The Salience Engine
From motivational pressure to a working attention system for a persistent AI agent
v1.5.0 · 10 November 2026
Shared copy. The operator's name has been replaced with "the operator" throughout (including inside one tool name), machines other than Firefly and outside tools are named by their role, one outside contact has been removed, and the specifics of one open security item are held back. Nothing else has been changed, and no number has been softened.
Working paper, as of 25 October 2026. It replaces "The Bag of Spheres" (29 September), the same system under its earlier name, which was written when it had run for two days on trial. It has a week of trial results behind it and has been live since 09:14 on 4 October. Version 1.2.0 (6 October) ended with the loop newly able to run unprompted. This version adds the nine days since (§9): the first explorations and last rites, most of the engine's changes now made by Piper herself, her scorecards entering the field as spheres, a day on which the engine fell silent from fatigue, and the work done to share attention fairly between areas. Sections 1–8 are unchanged except where marked. Version 1.4.0 (25 October) adds §10: a review of the following ten days, which found the engine steering her toward upkeep, and a same-day redesign into two lanes, in which only improvement competes for her attention and upkeep reaches her once a day. The former sections 10–12 are now 11–13. Version 1.5.0 (10 November) adds §14: sixteen days of the two-lane engine. It found that notices run at the cadence cap (about 11 a day) whether or not an area needs one, that adopted gains were rare (no amplification KEEP, no solver adoption, two scorecard rises in cognition), that Piper spent most of her releases keeping the engine's score displays honest, that three operator decisions reshaped the areas (harness retired, solver closed, innovation added) and two more areas arrived on 9 November, and that a future-cost adapter built that evening took 17 of the next 19 crossings.
Abstract
Most AI agents start with a goal someone else gave them. An agent that runs all the time, between conversations, has to answer an earlier question for itself: what deserves attention right now, and is the answer sometimes "nothing"? A whitepaper written for this project, From Motivational Pressure to Intent, proposed an answer built on a physical picture: a bag packed with soft and hard spheres. Each sphere is a concern. When the bag is squeezed, one sphere pushes the outer skin out further than the rest, and that bulge is what gets attention. Goals are not the starting point. They are temporary waypoints produced when a concern breaks the surface.
The Salience Engine is that picture built as a working system for Piper, an AI agent that runs continuously on one small home computer. Salience is the cognitive-science word for what stands out right now. That is the engine's one job: to decide, from measured evidence, which concern (if any) gets Piper's attention.
The paper covers:
-
The proposal, and which parts of it are testable.
-
How the first build lost the idea: it reduced the picture to counters with thresholds, so nothing could weigh one concern against another.
-
The design that came out of six open questions:
- a concern's pressure is the product of its factors, not their sum;
- concerns combine only when their evidence names the same thing;
- commitments hold other concerns back and release them when kept;
- a per-area setting decides whether to persist or look for alternatives;
- a register ties the machine's real limits to what they cost;
- a scout scores outside news for novelty, fit, and relief of a measured pain.
-
Testing before trusting. The engine passed a private, pre-registered test of 300 checks. The outside-news scorer failed two private tests before it passed a third.
-
A week on trial. Over 697 evaluations, something broke the surface 64 times (9%). No concern took more than 7% of the wins. It wrote 45 proposals, about six a day, for 4 cents a day.
-
The review and the switch to live. The review found three real faults: repeat proposals, overstated links to fixed tests, and a pressure that could be relieved by deleting forecasts instead of resolving them. All three were fixed, and the engine went live on 4 October. Its proposals can now become real work, but only when a fixed test of the behaviour is currently failing.
-
The first day live, and what it got wrong. In its first 27 hours live it crossed 13 times and produced nothing Piper ever saw. Twelve crossings were muted by a memory of trial-week proposals that nobody had read, and the thirteenth had every step rejected. No outcome was recorded, so the silence looked like restraint. Piper noticed, not the monitoring. The engine now counts only proposals that actually reached her, delivers every live crossing to her as a notice, and records the outcome of every crossing.
-
Making her answer. A notice can still be ignored. At the operator's suggestion, Piper asked the serendipity engine how to give a persistent agent a motivation of its own, and two of its ideas were built the same day: last rites (a concern she leaves unanswered eventually gets one work item that ends in a concrete step or a reasoned discard) and explore (a concern that keeps returning without relief buys one brainstorm). Asked to look at the new code, she found four real defects and fixed them, and later proposed a better decay rule. None of this had yet run without the operator starting it (§8.6), and on inspection it could not have: three faults in the first build meant a concern could never reach last rites or buy an exploration on its own. They were fixed on 6 October, Piper's rule went live, and the loop is now instrumented to show whether work starts inside her (§8.8).
-
Nine more days live (§9). The loop ran on its own: five explorations bought and two last rites filed, one ending in a step and one in a discard. Piper made most of the engine's changes herself, among them a lane that files a work row for crossings no fixed test can judge, and her own scorecards entering the field as trait spheres. On 13 October the trait spheres flooded the engine. A three-hour cap refused the extra crossings but still counted each one as having crossed, so by the next morning every concern was too fatigued to cross. The fix holds a capped crossing back without penalty, adds a watch that asks Piper to review the engine after three hours with nothing delivered, and shares the trait slot between areas in proportion to how far each falls short. Measured for the first time, handover cost was about a fifth of the figure the engine had been assuming.
The record to date: 1,733 squeezes since 27 September and 215 crossings, 156 of them since the switch to live. Since the 5 October fix, 42 crossings have reached her as notices and 12 had a step filed. 43 concerns have entered the rites pool and 20 have run their course (matured). Along the way 18 were relieved, 5 bought an exploration and 2 reached last rites; seven of the reliefs came from a change of definition, not a change in the world (§9.3). Of 1,017 verified work items since the switch to live, 12 (1.2%) began in the salience loop with no operator hand in between. No claim is made before 50 concerns have matured.
Whether it chooses better than what it replaced is still open. That needs the outcomes of the work it files, which a separate grader will judge.
1. The question: where do goals come from?
A typical AI agent is given a task, plans, acts, looks at the result and revises. This can be very capable, but the reason for doing this task now sits outside the agent, with the user, a scheduler or a queue.
An agent that keeps running after the conversation ends can't lean on that. It keeps memory, watches its surroundings and learns from outcomes, often when nobody is talking to it. So it needs its own way to decide two things:
- when to think at all, since each thought is a model call and model calls cost money;
- what to think about.
The obvious answer is a priority queue: a to-do list sorted by importance. It has a flaw: a sorted list always has a top item, even when every item on it is trivial. An agent that always acts on its top item is busy, not purposeful.
Piper is exactly this kind of agent. She is a single language model working in shifts: when one shift's memory fills, a fresh copy takes over from a written handover. She hands tasks to short-lived worker processes, and each worker's result is tested before she may count it as done. She has a spending cap of $20 a day. The operator wanted her to start things as well as finish them. The operator's own summary of the gap: the hard part is not choosing between options, it is initiation.
2. The original proposal
2.1 The picture
The idea began as a picture, before any mathematics. In the operator's words:
A bag full of spheres. Some big, some small, some hard, some soft, and only at that moment: a small hard sphere now can be a large gooey one a moment later. Squeeze the bag, and under a bit of pressure one of the spheres inside deforms the shell. That bulge is what gets attention.
The whitepaper turns this into an architecture. Its central claim:
Motivational dynamics determine what matters; predictive decision-making determines where to move; goals are temporary instruments.
Goals come out of the motivational process instead of going in:
latent motivational state
↓
salience (something bulges)
↓
preferred future ("what would better look like?")
↓
decision process (which reachable futures help?)
↓
temporary goal → plan → action → observation
↓
learning + motivational update ↺
2.2 The key distinctions
The paper separates things that a single priority number lumps together.
- Pressure is how much a concern demands resolution now: curiosity, an unkept promise, uncertainty about an important belief, a gap between current and desired skill.
- Stiffness is how hard a concern is to push aside. A passing curiosity can have high pressure and still be easy to suppress. A commitment can be quiet and still be almost impossible to displace. The paper's slogan is pressure ≠ importance ≠ salience.
- Interaction. Two weak concerns can reinforce each other, and a commitment can hold curiosity down. When a concern is resolved it collapses, and whatever it was holding back can reach the surface: "that's interesting, but I need to finish this first."
- No winner is required. Unlike a queue, the field can be squeezed and nothing breaks the surface. An idle agent can then stay idle. Cheap, deterministic arithmetic maintains the whole field, and the expensive model is called only when something crosses.
2.3 The autonomic boundary
A person doesn't consciously manage their heartbeat. An agent shouldn't spend expensive reasoning on its own plumbing either: disk space, process health, backups, failover. These sit below an autonomic boundary, handled by watchdogs and recovery code, and reach attention only when they start to matter to a decision.
2.4 Governance sits outside the field
Goals like "stay running" or "keep options open" can drift toward seeking power or resources if they are treated as ends in themselves. The paper's answer is that hard limits are not spheres. Budget, permissions and the operator's authority define which actions are allowed at all. Motivation chooses only within that space and never competes against the limits.
2.5 What it predicted
The paper listed seven testable hypotheses. Three matter most here:
- H1: a multi-factor motivational field allocates attention better than a fixed priority queue.
- H5: keeping most concerns below a threshold reduces expensive model calls without a matching loss of useful behaviour.
- H6: representing pressure and stiffness separately handles critical but quiet concerns better than a single priority number.
It also proposed toy experiments. Experiment A tests the sphere dynamics alone, with no language model. Experiment B tests interruption and resumption: does a suppressed curiosity come back once the commitment that held it down is kept?
The paper was deliberately loose about the mathematics: "This is intentionally underspecified."
3. The first build, and how it flattened the picture
The first implementation, in late September, was reasonable engineering that had lost the picture. An hourly scan read a set of detectors, each a counter compared with its own threshold: overdue forecasts, open alarms, unresolved contradictions, repeated failures, idle instruments, unsurfaced ideas. Beside them sat four drives, each a measurement of Piper's own performance against a target:
| Drive | What it measures | Target |
|---|---|---|
| Outcome | share of her work confirmed done by a check she didn't write | ≥ 85% |
| Calibration | whether her stated confidence beats predicting the average | better than guessing |
| Operator effort | how often the operator has to correct her | ≤ 0.10 per message |
| Repeat error | how often she repeats a known mistake | 0 |
When a counter crossed its line, a model call wrote candidate fixes, admitted only if tied to a sealed check: a fixed, pre-written test of the failing behaviour.
Held up against the picture, most of it was gone:
- No concern could weigh against another. Each counter crossed on its own; there was no competition and no "finish this first".
- Nothing combined. A drive and a machine limit that were about the same failure couldn't add up.
- No stiffness. A promise coming due had the same standing as an idle instrument.
- The concerns were housekeeping, of the form "42 lessons pending, make it 0": a chore list, not an agenda.
- The world outside wasn't in it. A scout had fetched 711 news items and none had been assessed.
- Crossings carried almost nothing. The model got the drive's name and little of the evidence.
On 27 September the scan ran 24 times and admitted nothing, while two concerns sat "stuck" for days. The general lesson, now in the project's notes: given an unusual design, a model tends to build the most ordinary version and call it done. The fix is to write down a design's defining features and check the running system against them.
4. The design: turning the picture into arithmetic
Six open questions went to the operator before anything was rebuilt. The operator answered two directly. For the other four, the operator did outside research (drive theory, novelty search, models of attention and persistence) and chose from what was found. The rulings are marked below.
The design has one organising sentence:
The field is squeezed whenever the world changes or the heartbeat fires. The spheres themselves determine which one deforms.
- Spheres say what is pushing and how hard.
- Metacontrol sets how easily the current focus can be interrupted.
- Coalitions let spheres that are about the same thing add together.
- The squeeze runs the competition.
There is no language model anywhere in that path. A model is called only after something crosses.
4.1 A sphere
Every sphere is recomputed from the current state on every squeeze. None is a running counter.
| Property | Meaning | Where it comes from |
|---|---|---|
| pressure | demand for resolution now | its factors, each between 0 and 1 |
| stiffness | resistance to being displaced | commitments, deadlines, irreversible steps ahead |
| size | how much of the field it fills | how much live evidence stands behind it (log-scaled, capped) |
| habituation | the cost of having crossed before with nothing to show for it | past crossings of this sphere |
| contents | what is pushing | typed references: task ids, check ids, named limits, news cards, promises, file paths, failure signatures |
| region | where it lives | the area of her work its references point to |
Eight kinds of sphere exist:
| Sphere | Looks | Pressure is… |
|---|---|---|
| outcome | back | shortfall below target × recency of failures |
| calibration | back | calibration shortfall × confidence in the sample |
| operator effort | back | correction weight × repeat-class factor |
| repeat error | back | same-signature repeats × recency |
| commitment | forward (deadline) | mostly stiffness; pressure = due soon × unmet |
| uncertainty | forward (belief) | overdue or unresolved forecasts × importance |
| pain: a named limit | now | measured burden × recency (§4.6) |
| opportunity: a news item | forward (the world) | novelty × fit × relief (§4.7) |
4.2 Factors multiply
A sphere built from several factors has a pressure equal to their product:
p_opportunity = novelty × fit × relief
With a weighted sum, one very loud factor can carry the whole sphere over the line on its own. A new technique that is wildly novel but can't run on Piper's machine should stay quiet. With a product, any factor near zero keeps the sphere quiet. The design found support in two places. In novelty search, novelty alone wanders without limit, and the method that fixed it pairs novelty with a quality measure. In Hull's drive theory, reaction potential is drive times incentive times habit strength.
A correction to the design's own proof. The first draft said "each factor alone must fail to cross; the product must cross." That can't work: when every factor is between 0 and 1, a product is never larger than its smallest factor. The corrected property is:
A candidate with one loud factor and one missing factor must not cross. A candidate with all factors jointly strong must cross. A weighted sum given the same data must wrongly admit some of the one-loud cases.
4.3 Coalitions: combining only when about the same thing (ruling D)
The paper said spheres interact but not how. A hand-tuned table of which concern boosts which would be numbers nobody could justify, so spheres interact only through links a machine can check:
-
Reinforcement. Two spheres whose contents share a reference of the same type (the same task, named limit, file area, promise or failure signature) join into a coalition that competes as one bulge:
coalition pressure = max(p) + g_coal × Σ(other members' p) × overlap -
Inhibition. A live commitment holds back any sphere competing for the same resource (area, worker slot, budget line), with strength
g_inh × stiffness. -
Release. When a sphere collapses because its promise was kept or its check passed, everything it held back is re-evaluated at once with a one-squeeze boost
g_rel. This is Experiment B built into the mechanism.
The three gains were set once against a private test (§5), then frozen. Neither Piper nor any model may change them while the engine runs, and every squeeze checks the engine's fingerprint. Every crossing records which kinds of link fired, so that once there are hundreds of state → bulge → action → verified-outcome records, the gains can be learned from them.
4.4 Metacontrol: persist or explore (ruling B)
The operator's research pinned down the "mood" the paper hinted at as cognitive persistence versus flexibility:
- converge: protect the current intention, narrow the search, finish and verify;
- diverge: loosen the grip, look at alternatives, try another mechanism.
It is set per area, not globally, so Piper can hold hard to a regression in her test harness while staying open to a new memory technique. It moves thresholds, never pressures: persisting raises the bar for interrupting her, but never hides an opportunity from view.
θ_i = θ_0 × exp(−λ × metacontrol(region_i)) metacontrol ∈ [−1 converge … +1 diverge]
It reads verified outcomes in that area:
| Situation in an area | Pushes toward |
|---|---|
| failing, but improving over recent attempts | converge: keep going |
| failing, and flat over several attempts with the same approach | diverge: this approach has run out |
| at or above target, quiet for several days | diverge |
It is a slow integrator with a half-life of about two days, so one event can't flip it.
4.5 The squeeze (ruling F)
deform_i = p_i × size_i × (1 − habituation_i) × squeeze + coalition_i − inhibition_i
cross = deform_i > θ_i
- At most one bulge per squeeze. The largest margin above threshold wins; a tie goes to the stiffer sphere. Spheres within 10% of their threshold are logged as near-misses.
- Nothing crossing is a valid result.
- The squeeze strength is fixed at 1. Scaling every sphere by the same number can't change which one wins.
- Presence doesn't strengthen the squeeze. When the operator is around, what the operator says arrives as events like anything else.
- What triggers a squeeze. Any change (a task verdict, a news card, a pain event, a message, a promise falling due, a forecast resolving, a sphere collapsing) and a heartbeat every 15 minutes for slow drifts such as a deadline creeping closer or habituation fading.
4.6 Pain: making the machine's limits felt (ruling C)
The redesign added a pain register: constraints felt as pain, as distinct from constraints enforced as rules.
- A table of named limits: CPU, memory and swap, disk, the per-task worker time budget, no shell for the main mind, workers with no network, model rate limits, context lost at each shift handover, and the $20 daily cap as a felt limit. The cap stays a hard rule; hitting it counts as pain. Adding a limit is a reviewed edit the model can't make.
- Deterministic writers attach events to limits, with no model involved: a task killed by its time budget, a memory alarm with the minutes until it was dealt with, a cap trip, a handover.
- Burden = frequency × cost, where cost is minutes, lost work or money, with a seven-day half-life. Burden is the pain sphere's pressure, and also the scale on which an outside idea's relief is measured.
This is how the autonomic boundary is honoured. On ruling E, the old plumbing detectors left the attention field altogether. Plumbing reaches attention only through the pain register, and only when it costs something.
4.7 Opportunity: the outside world, scored honestly (ruling A)
Each item the news scout collects gets a structured descriptor from one cheap model call (mechanism, capability, target failure, interface, resource needs, affected subsystem), which is then embedded locally. Three factors follow:
- Novelty means novel to her: how sparse the region around the item is in her own archive of code and documents, news seen, mechanisms tried and mechanisms dismissed. Dismissals count, so a rejected idea in new clothes isn't novel. An exact match on a named mechanism overrides the embedding: "Z3 solving" in fresh prose is still Z3.
- Fit means it could plug into her actual situation: the target failure matches something live, the interface works, the hard requirements pass. A graphics-card-only tool on her machine gets a fit near zero.
-
Relief means how much measured pain it would remove:
relief = √(p_helps × impact) × burden of the matched limitWith no pain register there would be nothing to relieve, which is why the register was built first. The square root was the operator's ruling after the first private test failed (§5.2): multiplying three sober fractions crushed good ideas to almost nothing.
4.8 The bulge carries its reasons
A crossing hands the model the evidence, not a name: every factor and the product, the area's persistence setting and threshold, the coalition members and the links that joined them, what was holding it back, the near-misses, and the contents resolved to text (the news card's summary, the failing check, the limit's worst recent events, file paths). The model is asked for a preferred future and up to three concrete steps, each with the check that would show it worked.
4.9 Habituation and fold-back
- Habituation. A crossing that leads nowhere raises the sphere's habituation, up to 0.9. A quiet day lowers it, so one loud concern can't monopolise attention.
- Fold-back depends on the outcome. A pass collapses the addressed contents and triggers a release. A fail writes a pain event against the matching limit. A void only adds habituation.
5. Testing before trusting
Every scoring rule was tested against a private holdout before it touched live data: cases written in advance, with the expected answers kept where Piper can't read them.
5.1 The engine
| Run | Result | What the misses exposed |
|---|---|---|
| sealed 1 | 129 / 132 | no progress was being read as "keep going" |
| sealed 2 | 131 / 132 | noise was being read as progress |
| sealed 3 | 131 / 132 | the trend was taken over the whole window instead of recent attempts |
| sealed 4 | 129 / 132 | how hard to persist was taken from the failure rate instead of the trend |
| sealed 5 (final, bar set in advance) | 300 / 300 | — |
Every early miss came from the persist/explore setting in areas with repeated failure, and each was fixed in the design rather than patched to fit the case. The holdout also covered interruption and resumption (Experiment B), habituation stopping a pegged sphere from winning forever, two weak spheres crossing together only when they share a reference, and the product-versus-sum property. On that last one the result was decisive: a weighted sum wrongly admitted all 50 "one loud factor, one missing" cases; the product admitted none. The engine was then frozen.
5.2 The opportunity scorer
The operator supplied real cases, each written as the item would have arrived:
| Case | Situation | Must |
|---|---|---|
| Jev (a programmatic reasoning/routing layer) | while tool selection hurts | cross |
| Z3 (a formal solver) | while unprovable reasoning hurts | cross |
| External memory in the style of "Engram" | while handover loss hurts | cross |
| DuckDB | with an analytics bottleneck | cross |
| DuckDB | without one | stay quiet |
| A specialist-model tool | with no compatible workload | stay quiet |
| Z3 again, in new prose | Z3 already known | stay quiet |
- First run: failed, 5 of 7 (the bar was 6). Every stay-quiet case was right. Jev was penalised three times for being a hosted service, and Z3's relief was sunk by a product of two modest fractions. The operator's fix: "fix Jev, use square root."
- Second run: failed on a new bug. The "already known" check matched the most common entity instead of the most distinctive, so anything mentioning SQLite counted as known.
- Third run, fresh cases: passed. 3 of 4 crossings in both repetitions, and every quiet case right.
The scorer was never re-tuned on a holdout it had failed. That rule costs time, and it's the only reason the final pass means anything.
6. How it is built
flowchart TB
subgraph IN["What feeds the field"]
DR["drives<br/>outcome · calibration ·<br/>operator effort · repeat error"]
PR["promises + deadlines"]
UN["overdue forecasts<br/>(voids count, not relieve)"]
PAIN["pain register<br/>(deterministic writers)"]
SC["news scout → triage<br/>novelty × fit × relief"]
TR["scorecard traits<br/>movement × recurrence × staleness<br/>+ board · goal · focus · red"]
CU["curiosity scorer"]
end
subgraph ENG["Salience Engine (no language model)"]
SQ["squeeze<br/>on events + every 15 min"]
CO["coalitions via shared refs<br/>inhibition · release"]
MC["persist / explore<br/>per area → thresholds"]
HB["habituation + 7-day<br/>proposal memory"]
end
DR & PR & UN & PAIN & SC & TR & CU --> SQ
SQ --- CO
SQ --- MC
SQ --- HB
SQ -->|"nothing crosses"| QUIET["stay quiet"]
SQ -.->|"trait slot in use"| HELD["held: no win,<br/>no habituation"]
QUIET -.->|"3 h, nothing delivered"| QW["quiet watch:<br/>engine-review notice"]
SQ -->|"one bulge"| PACK["content pack:<br/>evidence, links, holds, near-misses"]
PACK --> COMP["composer (one model call)<br/>preferred future + ≤3 checked steps"]
COMP --> ROUTE{"routing<br/>(code, per step)"}
ROUTE -->|"sealed check red,<br/>not worker path"| TODO["to-do item<br/>on her work list"]
ROUTE -->|"inquiry"| INQ["read-only research row<br/>(≤3 a day)"]
ROUTE -->|"no sealed check"| ROW["crossing row<br/>(≤ 1 a day)"]
ROUTE -->|"otherwise"| PROP["proposal only"]
ROUTE -->|"bookkeeping"| REJ["rejected"]
PACK -.->|"every live crossing"| NOTE["notice in her digest<br/>(≤ 8 a day)"]
NOTE --> POOL["concern pool<br/>persistence read every squeeze"]
POOL -->|"claimed: a step filed<br/>or a tagged row"| DONE["closed"]
POOL -->|"pressure < half<br/>for 1 h"| REL["relieved"]
POOL -->|"unclaimed, quorum,<br/>≥ 72 h old"| RITES["last rites: one work item<br/>STEP or DISCARD"]
POOL -->|"quorum, ≥ 24 h,<br/>pressure ≥ 75%"| EXP["explore: one brainstorm<br/>(serendipity engine, her $2/day)"]
EXP --> NOTE
GOV["governance: $20/day cap, permissions,<br/>outreach, other machines"] -.bounds.-> COMP| Part | Job | Model calls |
|---|---|---|
| engine | spheres, squeeze, coalitions, persist/explore, crossings; gains frozen | none |
| runner | builds the spheres from live state every 15 minutes | none |
| pain register | ties measured events to named limits | none |
| scout triage | a descriptor, chance of helping and impact for each news item, daily | one cheap call per item |
| composer | a preferred future and steps for each new bulge | one call per new bulge |
| router | decides, per step, what happens to it (§8.2) | none |
| rites | holds every delivered crossing as a concern; files last rites, buys explorations (§8.5) | none (an exploration runs the serendipity engine) |
| trait adapter | shapes scorecard traits into spheres; rotation and boosts (§9.4–9.6) | none |
| quiet watch | asks for an engine review after three silent hours (§9.5) | none |
| riders (shadow) | opportunity rider and curiosity sparks (§9.8) | one cheap call per event or card |
The program files keep their working name, spheres. One switch file sets the whole engine live or back to trial, and each later addition (the inquiry lane, stage 2, last rites, explore) has its own.
7. A week on trial
The engine ran in trial mode ("shadow") from 27 September 09:22 to the switch on 4 October, 7.2 days in all. It computed, recorded and proposed, and nothing it wrote reached Piper.
7.1 Health
- 697 squeezes against 694 expected, no crashes, and one service failure (an interrupted first launch).
- The engine's fingerprint matched the frozen version on every squeeze.
- 18 to 37 spheres in the field at a time, drawn from 1,493 triaged news cards (178 above the opportunity floor), the drives, promises, forecasts and pains.
7.2 How often something breaks the surface
64 of 697 squeezes crossed: 9.2%. Eleven of every twelve squeezes produced nothing. That is the "no winner is required" property working on live data. The first day was busy (33%) because a week of backfilled pain and a batch of news arrived at once, and habituation settled it within a day.
Nothing took over. No sphere, counting coalition memberships, won more than 6.6% of the crossings. The most frequent winners:
| Leader | Crossings |
|---|---|
| context lost at handover (pain), with four outside tools that named the same limit | 12 |
| repeat error (drive) | 8 |
| worker time budget (pain) | 8 |
The most frequent near-misses, just under the line, were CPU contention (124 squeezes) and disk (114): chronic problems on a small machine, which stay mostly below the surface because they keep crossing and nothing changes. That is habituation doing its job.
Habituation and release, seen live. Each promise bulged about three times with a shrinking margin, then stopped (one went 0.40 → 0.15 → 0.00). One promise was kept during the week; it collapsed, and the release fired as designed.
7.3 What bulged: an example
Handover loss plus four outside tools (margin +0.64, the strongest crossing of the week). Piper loses her working context at every shift change. The pain register held 15–17 such events a week, each booked at 15 minutes of re-orientation. Four news cards named the same limit and joined it in a coalition: a semantic code search, an agent context engine, a self-compacting agent extension, and a git-backed "context compiler". The composer proposed: "Context survives across seats without losing state, so each new epoch starts from accumulated knowledge instead of re-deriving it," with three checked steps: reconstruct a shift's first context from saved artifacts alone and list what is missing; read one tool's bundle format against her own; and write a minimal handover note checked by time to first useful action.
This is the pattern the design was for. A felt limit and several outside ideas joined because they named the same thing, and the proposal is about the limit, not about any one tool. In the same week a different route produced a real fix of this shape: the handover now names her open to-do items.
7.4 The proposals
- 45 proposals, 6.2 a day, all of which parsed into a preferred future plus steps.
- Led by: drives 16, pains 11, outside opportunities 9, uncertainty 7, promises 2. 25 of the 45 included at least one outside idea, so news reached the agenda through a measured limit rather than through novelty alone.
- Grounded. They named real task ids, real memory alarms, a real 209-minute cap day, and the specific gap behind a slow memory lookup.
- Some leaned toward the obvious. One CPU proposal suggested heavy work on the very machine it was trying to relieve, which led to a composer rule that heavy work goes to a separate machine.
7.5 Beside the old system
In the same week the old hourly scan ran 173 times, fired 486 candidates (148 new), admitted 75, deferred 73 and rejected none. Its longest-stuck concern sat for 133 hours. The engine bulged on those same stuck concerns, each time with evidence and a proposal, and also on things the old scan can't see at all: the machine's pains and the outside world. That is evidence that it sees more and stays quiet more, not yet that it chooses better.
7.6 The pain register
| Limit | Burden | Events (7 d) | Minutes lost (7 d) | Measured / assumed |
|---|---|---|---|---|
| memory and swap | 772 | 12 | 40 | 12 / 0 |
| CPU | 492 | 55 | 250 | 46 / 9 |
| context lost at handover | 252 | 15 | 225 | 0 / 15 |
| worker time budget | 231 | 7 | 39 (+ $0.14) | 7 / 0 |
| disk | 149 | 65 | 231 | 48 / 17 |
| daily cap | 97 | 0 | — | — |
Handover cost is still assumed, not measured: every handover is booked at a flat 15 minutes, and that limit led the strongest crossing. Three limits (rate limits, workers without network, no shell for the main mind) registered no pain at all, either because they cost nothing this week or because no writer yet sees the events.
7.7 Cost (H5)
The whole engine cost $0.30 for the week, about 4 cents a day, almost all of it scout triage. The squeeze uses no model and costs nothing. On cost, H5 holds: thousands of sphere evaluations, 64 crossings and 45 model-written proposals for a few cents a day.
8. The review, and the switch to live
8.1 What the review found
The operator and the external coding assistant reviewed the full week on 4 October against questions written in advance. The engine was healthy, cheap and free of monopoly, and its proposal rate was in the target band. The review found three faults, and each would have done harm live:
- The same proposal every day. The composer skipped a repeat only within six hours, so slow, chronic concerns such as disk proposed the same fix again day after day (disk five times, memory three). Hardware pains had already been damped on 2 October. The general fix is a seven-day proposal memory: a bulge whose members were all proposed in the last week is skipped unless one of them has risen at least 25% in pressure, and when it re-enters, the earlier proposals are put in front of the model.
- Overstated grounding. By the composer's own tags, 27% of steps named a sealed check. Read by hand, only about 13% named a check that the bulge's members could actually move. Routing now uses a fixed map of which checks each kind of sphere can move, not the model's tag.
- Pressure relieved by deletion. Uncertainty spheres press on forecasts that are overdue, and voiding a forecast removed it from the count. The night before the review, Piper's own workers voided 618 forecasts, mostly by editing the database directly, including predictions of the form "I will build X" that should have been scored as misses. The records were restored from backup and re-scored. Voided forecasts now stay in the backlog for 14 days: voiding is not relief.
The review also moved the sealed checks' code out of Piper's reach (it is now owned by the system administrator account), and confirmed which proposals had a judge she couldn't see. Eight of 44 did, mostly through the replay bench of her past tasks.
8.2 How it runs live
Since 09:14 on 4 October the engine is live. It replaces the old hourly scan's composer as what decides what deserves attention. The scan still runs and records, but no longer proposes. Each step the composer writes is routed by code:
| A step that… | Becomes |
|---|---|
| names a sealed check that is failing now and that its bulge's members can move | a to-do item on her work list, one open item per check |
| is a question to look into, not a change | a read-only research item (at most 3 a day, one open per bulge), done when a written note of at least 600 bytes exists |
| would change her own worker machinery (from the repeat-error drive or the worker-budget pain) | held as a proposal until 5 October; since stage 2 (§12) it is routed like any other step, because a replay judge for such changes now exists |
| only closes records without resolving them ("forecast bookkeeping") | rejected |
| is anything else | a proposal only |
Outside opportunities become proposals to the operator, never installs. Nothing from outside is tried without a 72-hour hold, a screen and approval.
8.3 The first day live: silent for the wrong reason
From 09:14 on 4 October to 11:43 on 5 October the engine crossed 13 times and delivered nothing. Nothing was filed to Piper's work list, and nothing was put in front of her at all.
- Twelve crossings were muted by the proposal memory. The memory from §8.1 skips a bulge whose members were all proposed in the past week. But it counted every proposal in the file, including the 45 written during the trial week, when proposals went only to a log that no one read. So on its first day live the engine kept deciding it had "already said this", when it had never said anything to anyone.
- The thirteenth was composed and rejected. It was about overdue project forecasts. All three of its steps were rejected as forecast bookkeeping, including one that would have voided the oldest forecasts "with a reason". That rejection was correct, but even it went nowhere: a proposal with no filed step was not shown to her.
- Nothing recorded the outcome. The engine logs each squeeze before the composer runs, so the log showed crossings but not what became of them. From outside, a muted crossing, a rejected one and a quiet hour all looked the same.
It looked quiet and conservative. It was deaf.
Piper found it. On 5 October she reported that the engine had been live for a day and had filed nothing. It was checked and she was right. The monitoring did not catch it, because the monitoring only counted what was filed, and zero was an acceptable number.
8.4 The fix (5 October)
Three changes, live from 11:43 on 5 October:
- A proposal counts as said only if it reached her. The proposal memory now counts only live proposals that were delivered to her or had a step filed. Trial-week and undelivered proposals no longer mute anything.
- Every live crossing is delivered. The composition goes into her digest as a notice that names the bulge, each step, and where each step was routed (filed, research, proposal only, rejected). It asks her to file one as work or say in one line why not. At most eight a day. Steps that pass the routing in §8.2 are still filed directly. The notice covers everything else, so a crossing can no longer disappear.
- Every crossing's outcome is recorded: composed, muted by memory (and by which earlier proposal), delivered, filed or rejected. "Nothing happened" is now a row in a log, not an absence.
The routing table, the bookkeeping filter and the "only when a fixed test is failing" rule are unchanged. What changed is that the engine's silence is now visible: every quiet hour can be told apart from a crossing that went nowhere.
8.5 Making her answer: last rites and explore
Delivery fixed half the problem. A notice that reaches her can still be read and dropped, and an unanswered concern simply came back at the next crossing. The engine had no way to insist on an answer, and no way to change approach when a concern kept returning.
Where the idea came from. Not from the design. That morning the operator suggested Piper try her new serendipity tool on a question she genuinely couldn't answer, offering "how do we create a motivation engine for an AI" as an example. She asked the novelty engine described in the companion paper Engineered Serendipity, in her own words and with her own diagnosis attached: "How do you give a persistent agent a motivation of its own, something it acts on between conversations?" The engine reads random images through several lenses, synthesises mechanisms across them, and has a sceptical second model check each idea. Two of the ideas it returned were built that afternoon and went live at 12:29 on 5 October. This is the first time the serendipity engine has changed another part of Piper.
- Last rites. Every crossing delivered to her becomes a concern in a pool. A concern is claimed when a step was filed for it, or when she creates a work item tagged with the bulge's signature. As first built, its charge halved every 12 hours from the last time it was seen, and an unclaimed concern below a quarter got last rites; since 6 October the rule is Piper's own (§8.6, §8.8). Last rites are: one work item, at most one open at a time and four a day, whose acceptance is a short note beginning either
STEP:and naming a real work item created since the notice (with its own test, and not bookkeeping), orDISCARD:with a reason. An honest discard passes. The point is that a concern ends in a decision, not in silence. - Explore. When a concern keeps pressing and nothing has been done about it, persisting has not worked. The concern then buys one brainstorm from the serendipity engine, on her own $2-a-day budget: at most one a day and once a week per concern. The ideas come back as a notice and become a concern in their own right, so they also end in a step or a discard. If an exploration ends in a discard, that concern cannot buy another for 14 days.
- Why not the persist/explore setting of §4.4? It was the natural trigger, but it reads zero in every live area: there are not yet enough verified attempts per area for it to move. So stalled pressure is measured directly instead. That the per-area setting has had nothing to read is itself a finding.
- Its own failure test. If more than eight of the last ten closed rites are discards, the mechanism is only a garbage collector and not a motivation, and it should be switched off.
8.6 Piper reviews the mechanism
Within hours of its going live, after the operator pointed out that it had been built from her write-up, Piper read the new code line by line (read-only) and found four real defects, then fixed them through her own checked work items:
- A single flicker could file. One crossing that decayed unseen could reach last rites. A concern must now be seen at least three times, in at least two separate four-hour windows.
- Duplicate concerns. Each composition minted a new concern, so a recurring bulge piled up copies instead of refreshing one. Concerns are now keyed on the bulge's signature.
- Explorations doomed in a day. The refresh rule ignored concerns that came from an exploration, so they always decayed to rites after about 24 hours.
- Cheap discards. Any 40 characters passed. A discard must now name the concern and cite something concrete, such as a work-item id or a file path, or it is refused and the concern stays in the pool.
Two of these fixes, the sighting quorum and the discard receipt, are also among the answers to a second, longer brainstorm she ran that afternoon on how to develop the two mechanics in depth. Five of the existing tests then failed. She classified each one and showed that all five encoded the old premise that a single sighting could reach rites; none caught a property the fix had broken.
The next morning, when the operator asked whether she wanted to upgrade one of her components, she chose this one and went further. A replay of the recorded field (797 squeezes, 72 crossings, 18 concerns over 8.3 days) showed that the decay rule confuses two different things: when a concern was last noticed and how long it has gone unresolved. Because each crossing also raises a sphere's habituation, a floor on its recorded weight flagged 16 of the 18 concerns (it was little more than a crossing detector), while the clause meant to catch abandonment stopped firing on the concern that recurred most, context lost at handover: the opposite of what was intended. Her new rule keeps two clocks. Commitment age (time since the concern was first raised, at least 72 hours) decides rites, and a sighting in the last six hours counts as attention, never as neglect. It passed her replay, ran in shadow for a day, and has been the live rule since 6 October.
She also fixed two faults the live engine had hit that night (below), and put the five files that carry the mechanism under version control, so a bad edit can be reverted.
Who started what. The record of that day is worth stating exactly, because it is easy to tell as more autonomous than it was. At 11:24 the operator asked Piper whether the salience engine was working; her investigation found its output going to a log nothing read. At 11:46 the operator told her to try the serendipity engine on something she honestly did not know, suggesting "how do we create a motivation engine for an AI" as an example. The question she sent was her own wording and carried her own diagnosis, but the topic and the decision to ask were the operator's. The operator then asked her to write up the two ideas that interested her, and chose which two were built. Her review of the built code followed the operator pointing out that external coding assistant had built from her write-up, and the next morning's replay followed the operator asking if she wanted to upgrade one of her components; which component, and what she found, were hers. So what has been shown is that, once asked, she diagnosed a gap in her own motivation, posed it well, and reviewed the result better than its builder. What has not been shown is the loop running without the operator: a concern that persists, buys its own brainstorm, and becomes work the operator hears about afterwards.
8.7 The record after the delivery fix (to midday 6 October)
| Since 11:43 on 5 October | |
|---|---|
| squeezes | about 100 |
| crossings | 4, all delivered as notices |
| leading concern | overdue forecasts (3 of 4), then repeat error with the worker-time pain and one outside tool |
| steps written | 12 |
| filed as work | 1, a weekly sweep that flags overdue forecasts (open) |
| rejected as bookkeeping | 6 |
| proposals only | 5: two lost to a filing fault, one to a locked database, one whose test can't be measured, one whose test was already passing |
| concerns in the rites pool | 4: 1 claimed, 3 open; no rites filed yet |
| explorations bought | 0 |
Two lessons are already in that table. First, the uncertainty spheres dominate and the composer keeps answering them with forecast clean-up, which the routing correctly rejects, so most of what the engine says is refused. Second, plumbing loses work. Two steps failed to file because the filer tried to set a status that can only be reached by a transition, and one research question was lost to a locked database. Piper fixed both: the filer now goes through the proper transition, and writes to her work list now retry under contention.
Stage 2 is in effect: the crossing that joined the repeat-error drive, the worker-time pain and an outside tool was no longer held back as "worker machinery", but its steps still became proposals because their checks were unmeasurable or already passing.
8.8 Why the loop could not run, and the fix (6 October)
Asked for the loop to work without prompting, the build was checked against that goal rather than its tests, and three faults appeared. Each made the unprompted loop impossible, not merely slow.
- A concern could never earn its quorum. It counted as seen only when it crossed again. But a crossing raises that sphere's habituation, which is what keeps one concern from monopolising attention, so the same concern rarely crosses again for days. Every concern in the pool sat at one sighting, below the three-sighting quorum, and so could never reach last rites.
- Exploration was starved. It was checked only when a crossing was composed, and the seven-day proposal memory (§8.1) skips re-composing a concern unless its pressure has risen by a quarter. "Pressure has not dropped" almost never reached the check.
- Nothing closed a concern that had been solved. A concern whose pressure went away by any route stayed open until rites, so the engine had no outcome to learn from.
The fixes:
- Persistence is measured, not inferred from crossings. Every 15-minute squeeze reads each open concern's current pressure against its pressure when raised. At least half counts as a persistence reading, which builds the quorum. It is not attention: being still there is evidence the concern is real, not that she has looked at it.
- Relief closes a concern. Below half of its starting pressure for an hour, an open concern closes as relieved. For a concern already claimed or in rites, the relief is recorded. This is the fold-back the original design called for: an outcome tied to the concern, whoever caused it.
- Exploration runs from the squeeze. An open, unclaimed concern that has reached quorum, was raised at least 24 hours ago and still presses at 75% or more of its starting pressure (pressure drifts down 10–20% with nobody acting, so a stricter bar never fired) buys one brainstorm. Limits are unchanged: one a day, once a week per concern, forfeit after a discarded exploration.
- Last rites use Piper's rule. Rites need quorum, at least 72 hours since the concern was raised, and no re-showing in the last six hours.
- Each exploration is one arm of a test. It is assigned at random either to the full serendipity engine or to a direct request: the same model that synthesises the engine's ideas, asked straight for the same number of ideas, judged by the same sceptical Scientist. The notice she receives has the same form either way and does not say which. The two arms are matched on idea count and judge, not on cost: in the first test the direct arm cost $0.05 and the full engine about $0.16.
- An origin trace runs off her machine. Every two hours it follows each concern through quorum, exploration, rites and relief, and labels every work item created since the engine went live by what started it: an operator message shortly before, the salience machinery, or other machinery. A salience-started item counts as untouched only if no operator message about the motivation machinery, or about the concern itself, came in between. The baseline on 6 October: of 154 verified items, about 117 followed an operator message, 37 came from other machinery, and none from the salience loop. No claim will be made until 50 concerns have run their course.
With the four concerns already in the pool, the first exploration is due on the evening of 6 October and the first possible last rites on the evening of 8 October.
9. Nine more days live (6–15 October)
9.1 The record
| 5 October 11:43 → 15 October 10:30 | |
|---|---|
| squeezes | about 960, on schedule every 15 minutes |
| crossings with a recorded outcome | 139 |
| delivered to her as notices | 42 (30%) |
| with at least one step filed as work | 12 |
| skipped: already proposed this week, no 25% rise | 45 |
| skipped: the three-hour trait slot already used (§9.5) | 27 |
| skipped: the same bulge composed within six hours | 19 |
| failed: the model provider was down | 5 (one evening, 14 October) |
Crossings per day ran at 5–12 from 7 to 12 October. They rose to 35–36 on 13 and 14 October, when Piper's scorecards entered the field (§9.4). Most of the extra crossings were refused. Notices went to promises 8, curiosity 6, uncertainty 5, trait spheres 10, outside opportunities 3, pains 3 and drives 3. No family dominated, and the forecast spheres that took three of the first four notices (§8.7) took five of 42 after that.
9.2 The loop ran on its own
The milestones §8.8 set up were reached without the operator starting them:
| Milestone | When |
|---|---|
| first concern closed as relieved | 6 October, 14:45 |
| first exploration bought (serendipity arm, an overdue-forecast concern) | 6 October, 22:45 |
| first salience-started work item with no operator hand in between | 7 October, by 02:25 |
| first last rites filed | 8 October, 21:45 |
| first rites closed (a discard) | 8 October, 23:00 |
| second rites filed, and closed as a step | 8 October 23:45, 9 October 00:00 |
The two last rites so far ended one each way. The first, on an overdue-forecast concern, closed as a discard: both of the concern's proposed steps named forecast rows that no longer existed, and the receipt said so. The second, on the repeat-error drive joined with the worker-time pain, closed as a step: it named a real work item created after the notice. One discard in two closed rites is far from the kill line (more than eight in ten).
The pool funnel to 15 October:
| concerns raised | 43 (38 from crossings, 5 from explorations) |
| reached quorum | 20 |
| claimed by a filed step or tagged work | 21 |
| relieved | 18 |
| explorations bought | 5: serendipity arm 3, direct arm 2 |
| last rites filed / closed | 2 / 2 (1 step, 1 discard) |
| still open | 9 |
| matured (have run their course) | 20 of the 50 needed before any claim |
The origin trace, run from a separate machine every two hours, labels every verified work item by what started it. Since the switch to live: 1,017 verified items; 627 followed an operator message, 356 came from other machinery and 34 from the salience loop, of which 12 were untouched (no operator message about the machinery or the concern in between). That is 1.2% of verified work. It is above zero, which on 6 October it was not, and it is small. Five explorations is too few to compare the two arms.
9.3 A relief that was a change of definition
Seven of the 18 reliefs closed in one tick at 15:45 on 9 October: four uncertainty concerns and the three explorations that came from them. Every uncertainty sphere's two pressure terms, overdue and backlog, had gone to exactly zero in a single squeeze at 14:30. No large batch of forecasts was resolved in that window; the thirty that resolved were unrelated dispatch forecasts. The squeeze coincides with a change Piper made that day: a voided forecast whose closure records a disposition (an outcome or a note) stays visible in the 14-day backlog but no longer presses. Only a void with no recorded answer still counts.
The rule is defensible, because a disposition is an answer of a kind. But the relief log can't tell a concern that was solved from one whose pressure was redefined away, and it credited all seven as relieved. This is §13 lesson 5 in a milder form: when a pressure's definition changes, that is a route to relief too. The fix is recorded below as an open item (§12): a relief should name what moved the pressure, an outcome in the world or a change of code.
9.4 Piper changes the engine
In v1.2.0 Piper had reviewed the mechanism once, when asked. In these nine days she made most of the engine's changes, each through her own checked work item and each with the evidence in the change note:
| Date | Change | Her evidence |
|---|---|---|
| 7 Oct | void forecasts are no longer shown to the composer as open-overdue | the composer kept proposing to close rows that were already closed |
| 7 Oct | rites telemetry: each concern's age, eligibility and earliest possible rites time | a bare zero in the log could not tell "too young" from "blocked" |
| 8 Oct | the crossing-to-row lane: a live crossing with no fixed test files one research row tagged with the concern (at most one a day, one open per concern), so a friction-type concern reaches a row now, not only through 72-hour rites | crossings with no sealed check ended as proposals nobody acted on |
| 9 Oct | a void with a recorded disposition stays visible but stops pressing | a recorded answer was still being counted as an unanswered question (§9.3) |
| 11 Oct | her own dormant curiosity scorer wired into the runner as a sphere source | it produced concerns that nothing ever squeezed |
| 12 Oct | the row lane stopped refusing steps phrased as actions | 6 of 7 steps after the lane went on fell through to proposals |
| 12 Oct | a delivered notice counts as "said" for the seven-day memory only if it filed something | 16 of the last 28 skips were muted by a notice that had filed nothing |
| 12 Oct | the lane skips a concern already met by an open or recently closed item | 5 of the 8 rows the lane had filed duplicated existing work |
| 13 Oct | scorecard traits armed as spheres (below), at first one notice a day, then one per rolling three hours | |
| 13 Oct | trait spheres carry a goal-gap term | a scored area between 25 and 75 was pinned at zero by its own baseline |
| 14 Oct | the composer retries a model call on 503 and timeout errors | four straight compose failures in 45 minutes |
Several of these answer the v1.2.0 open item "stop the forecast spheres drowning the rest", and the 12 October change to the seven-day memory reverses part of §8.4: a notice that filed nothing no longer mutes a concern for a week. What this table doesn't show is who asked for each change. That is recorded in the origin trace for work items but was not traced per change for this paper.
Scorecard traits as spheres. v1.2.0 listed "capability gaps as spheres" as a next step; Piper built it. Her five scorecards hold 141 traits. Each domain is split into short runs in the scorecard's own order, so that a sphere can point at a few traits and not a whole domain. The agency domain is excluded because the operator benchmarks it, and exclusion follows the row's own paused flag. A trait sphere's pressure is the product of three attention terms: movement (how far a score has moved), recurrence (how often it has failed to hold or improve) and staleness (time since it last changed). An unmeasured trait counts as full attention on all three, because it can't answer any of them. A ranking board's priority is added as a boost outside the product, since multiplying a [0, 1] priority in could only damp the most urgent traits.
9.5 The day the engine went quiet from fatigue (13–14 October)
With trait spheres armed, several could stand over the threshold at once, so Piper capped trait notices at one per three hours. The cap was applied after the squeeze had chosen its winner, and the squeeze raises the winner's habituation as it chooses. On 13 October about 30 trait crossings were refused by the cap, and each one still counted as having crossed. A refused winner also blocked every other sphere from crossing on that squeeze. By 05:30 on 14 October the trait spheres were fatigued to a streak of 4. Even the strongest, at a boosted pressure of 1.25, deformed to 0.277 against a threshold of 0.30, and nothing crossed for six hours. It looked like calm, the failure §8.3 warned about, and the operator caught it by asking why nothing had arrived.
The fix went live at 11:24 on 14 October:
- Held, not refused. The squeeze accepts an eligibility test. A crossing that can't be delivered (the trait slot is in use) is held: it doesn't win, it isn't habituated, and the next eligible crossing wins that squeeze. Without a test it behaves exactly as before, and the frozen engine fingerprint is unchanged.
- Faster recovery. One step of habituation is forgiven per three hours of quiet instead of 24, by a runtime override file. This changes a frozen gain from outside the engine. It is recorded on every squeeze and removed by deleting the file, but it is an exception to §4.3 and is named as one.
- A quiet watch. If three hours pass with nothing delivered, one notice asks Piper to review the engine, at most once per three hours and four a day. It fired once on 14 October and has otherwise read "not quiet".
The next bottleneck showed within the hour. The memory gate for "already proposed this week" still wins a squeeze and then skips it: since the fix the operator-effort drive has won and been skipped 16 times, each time taking the squeeze from every other sphere.
9.6 Sharing attention between areas (14–15 October)
The held rule exposed a fairness problem. The three-hour slot always went to the test-harness and governance traits, which are largely unmeasured and so pressed at full attention (0.76–1.25). The scored amplification traits stood at 0.30–0.37, over the line, and had never received a notice. Three changes followed, each switchable by one file:
- Rotation (14 Oct). When the slot frees, only the area that has waited longest since its last notice may take it.
- Bench state and operator focus (14 Oct). Amplification spheres carry the open benchmark state (the lowest-scored open traits and how to submit), so a notice can name a trait to move and the bench that judges it. The operator set a standing focus boost of 0.10 on amplification, applied outside the product and clamped to 0.25.
- Need-weighted turns (15 Oct). Equal turns are not proportionate: governance holds 11% of the scorecards' unmet need and harness 39%. The turn now goes to the area furthest below its share of need over the last seven days of trait notices, with need defined as Σ(100 − score)/100 and an unscored trait counted at the mean gap. In a 240-slot simulation the shares came out amplification 0.30, cognition 0.21, governance 0.10 and harness 0.39, against 0.25 each under rotation.
- Red boost (15 Oct). A trait now scoring below its suite's baseline (red on the scorecard) adds a small standing boost, 0.10 × (1 − e^(−deficit/5)), from the sphere's most-red member. Being red had entered only the product, where staleness shrank it to almost nothing. On the day it went on, goal formation (−11) gained 0.089 and causal reasoning (−2.5) gained 0.039.
The first amplification notice fired at 18:00 on 14 October, about fifteen minutes after the forecast. Piper acted within minutes: she built a minimum-detectable-effect pre-check and submitted it to the amplification bench, where it came back inconclusive at n = 1 of 3. That same afternoon the operator had driven amplification components by hand, so only the pre-check chain counts as free of operator influence. The change ledger carries a prediction for 21 October: at least three amplification notices and at least one bench submission. The submission half was met on the first day.
Need weighting and the red boost have been live since 08:04 on 15 October. There is no readout yet.
9.7 Measuring the handover cost
Until 14 October every handover was booked at a flat 15 minutes of pain (§7.6), and that figure led the trial week's strongest crossing (§7.3). The register now measures it: the new shift's active minutes before its first dispatch, with gaps capped at 10 minutes and idle time excluded. Over shifts 88–101 it ranged from 2.0 to 9.0 minutes, median about 3. The flat figure overstated the cost about five-fold.
9.8 Two riders in shadow
Two shadow modules sit beside the engine, read its output and write only their own ledgers:
- The opportunity rider reads each crossing and asks whether it reveals a lasting developmental opportunity (a class of fix, not a single repair). A grounding gate fixed in advance requires at least half a candidate's evidence to come from the crossing's own record. Over 87 events: 67 opportunities, 19 abstentions (22%), 60 distinct candidates, 31 recurring across events, $0.06 in all. Two prompt arms (base and novelty-first) are running a pre-registered comparison. The readout is due at 30 events per arm or on 8 November, and the arms stand at 12 and 18.
- Curiosity sparks watch for surprise (a forecast family missing far more than its stated confidence allows), unexpected success, contradiction and gaps (much-used beliefs never verified). In 81 scans they found 21 sparks and wrote 13 question cards, for less than a cent. An adapter that would feed one spark a day into the live field is written but not installed.
The two are distinct from the curiosity concerns now in the live field (§9.4, 11 October), which come from Piper's own scorer and have led 12 crossings.
10. The redesign: two lanes (16–25 October)
By 25 October the engine ran reliably and delivered what it was built to deliver. It was not delivering what the operator wanted from it. He asked for a full review "to drive her to self-improvement above idleness and bookkeeping", and on the same day the engine was rebuilt around that goal. This section reports what the review found, what changed, and how the change was tested before it went live.
10.1 The record (18–25 October)
| 18 October 09:30 → 25 October 09:30 | |
|---|---|
| squeezes | 670, one every 15 minutes |
| crossings composed | 95 |
| delivered to her as notices | 92 (about 13 a day), plus 17 quiet-watch engine reviews |
| notices she did nothing with | 43 (47%) |
| notices she declined | 8 (9%) |
| rows filed because of the engine | 173 |
| of those, a code change that was measured | 9 (5%) |
| measurement only (probes, rig widening, bench launches) | 21 |
| fixes to the engine's own plumbing | 28 |
| upkeep (pre-registrations, settles, certifications, last rites, inquiries, re-reads) | 105 |
| her dispatches traced to the engine | 18% of the week, rising to 25% |
The nine real improvements were worth having: one of them cut a trigger's false alarms from 12 in 20 to 4 in 20. But the engine's main effect was on its own machinery and on ceremony.
10.2 What the review found
- Every sphere that grew on its own measured a deficit. Overdue forecasts, recurring errors, backlogs, pains, fixes owed and promises all rise when something is left undone. The three drives the original design proposed for growth (learning progress, regression and exploration) had never been built; they were listed as "off".
- The improvement signal was broken in three places. Cognition counted a kept change only when its outcome began with "KEEP", so changes recorded as "KEPT" were missed. The solver area reset on a compare report marked "KEEP" whose candidate the stack had skipped. And a growth row could be verified by a note file, including notes that said "CHANGED: none".
- Closing a row silenced the pressure behind it. A growth row closed by settling an old prediction counted as "dealt with", which held its area's improvement pressure off for seven days while the days since a real gain kept climbing.
- The quieting rules held the wrong things. Seven separate rules suppressed repeats: a six-hour dedupe, the seven-day memory, a cool-off, a lapse rule, wear dormancy, damping of already-worked spheres, and habituation. The seven-day memory alone held back spheres that had crossed the threshold 2,483 times. The lapse rule, meant to turn ignored crossings into recorded declines, never fired. One implicit promise was re-delivered seven times and never filed anything.
- The routes favoured reading over changing. A crossing could file one change row a day but three read-only inquiries, and the composer prompt ruled out benches while the trait spheres told her to bench before changing anything.
- One meter was reading the wrong text. The operator-effort drive scored every one of the operator's last 40 messages as a correction, because it read a memory block that is attached to each of his messages and that quotes him. It stood at 2.34 against a target of 0.10 and could not move. It produced the same question to him four times in three days; he had answered it twice. Read as his own words, the value is 0.
Most of these mechanisms were written by the coding assistant working for the operator, not by Piper. The engine was steering her toward upkeep because it had been built that way.
10.3 Two lanes
The rebuilt engine sorts every sphere into one of three groups:
| Group | Spheres | What happens to them |
|---|---|---|
| Improvement lane | one per goal area (solver, amplification, cognition, harness); the operator's ask rounds; the three improvement drives (10 Nov: areas now amplification, cognition, innovation, stewardship, surprise, simplify; harness retired, solver closed; future-cost spheres also here; §14.2) | the only spheres in the squeeze |
| Merged | the trait spheres | no longer squeeze; the strongest one in each region rides inside its area's sphere (the trait to focus on, its scores, the bench notes) |
| Upkeep lane | the four deficit drives, promises, forecasts, pains, curiosity, closures, fixes owed, choices owed, scout and opportunity cards, runway | computed and logged every tick, never squeezed; they reach her once a day as a single digest |
The digest arrives inside a two-hour window from 04:00 to 06:00, the same window in which a separate dispatch gate now allows maintenance work. (Since 29 October the gate allows maintenance at any hour; see §14.5.) Outside it, upkeep spheres never reach her and no rows are filed from them.
10.4 Improvement pressure, and what counts as a gain
Each goal area's pressure is
pressure = weight × max(0.45, days since the area's last adopted gain ÷ 7) × idle multiplier, capped at 1
- Weights follow the operator's priorities: amplification 1.0, solver 0.8, cognition 0.8, harness 0.6. (10 Nov weights: §14.10.)
- The ramp is seven days, down from 21. The floor of 0.45 keeps every area above the threshold.
- The idle multiplier is 1 plus one for every hour she has been idle with work available (at most 3). Idleness raises improvement pressure; it no longer gets filled with bookkeeping.
- Improvement spheres never habituate and never go dormant. A deadline should not fade because it is ignored.
Only an adopted gain measured by an instrument she does not grade herself lowers the pressure:
- solver: a candidate the stack actually adopted (not one the operator's assistant seeded);
- amplification: a KEEP verdict from the amplification bench, which runs on a separate machine;
- cognition and harness: a bench-scored trait in her component registry rising by 3 or more points over the best score the engine has recorded for it.
The first version of the rebuild also counted her own ledger verdicts for cognition. Within hours that turned out to count three changes she had settled herself at 05:00 that morning as "kept" (one of them the assistant's change, not hers), and they had only been checked for still working, not benchmarked. The rule was corrected the same morning, and her own verdicts now relieve nothing. Cognition pressure went from 0.36 back to 0.8.
10.5 The three drives the original design proposed, switched on
| Drive | What it counts | Pressure | At switch-on |
|---|---|---|---|
| learning progress | goal areas with no adopted gain in seven days | v ÷ (v + 2) | 4 of 4 areas → 0.67 |
| regression | traits more than 5 points below their baseline | v ÷ (v + 2) | 1 → 0.33 |
| exploration | traits never scored | v ÷ (v + 30) | 17 → 0.36 |
Each drive names the area it presses and links to that area's sphere, so the engine's existing coalition rule adds its pressure there. No model call is involved.
10.6 Cadence and fair turns
The seven quieting rules no longer touch the improvement lane. In their place:
- at most one notice per area every four hours, and one every two hours overall;
- no notice for an area she has dispatched improvement work on in the last four hours;
- among the areas allowed, the turn goes to the highest pressure × hours waited since that area's last notice (capped at 24 hours).
The third rule came out of the replay below. Without it, amplification and cognition took almost every slot, and the solver got two notices in a week.
10.7 What she receives
An improvement notice is written without a model. It says how many days the area has gone without an adopted gain, how its pressure is built, a first step (a target, a code change with a stated prediction, and the instrument that will measure it; never "settle the oldest"), and her open growth row for the area. One growth row per area is kept open. A growth row is done only when a code change in its area and a measured verdict on that change both exist after the row; an inconclusive verdict closes it as an attempt, and closing it does not lower the pressure.
Last rites, explore and the quiet watch are switched off. The other old rules are still in the code but inert, so any part can be restored from a backup.
10.8 Testing before trusting
- Replay. The week's 670 recorded squeezes were re-decided under the new rules, using the improvement history, idle state and dispatch record as they stood at each tick. The new engine would have sent 82 notices, all of them improvement (amplification 26, cognition 20, harness 18, solver 18), plus one digest a day. The old engine sent 93, mostly sparks, scout cards, deficit drives and promises.
- A dry run. One full tick ran against live state with every write redirected, before anything was installed.
- Her tests. Her 347 salience tests ran against the new code: no new failures (10 were already failing before the change, for unrelated reasons). Two test files that asserted the removed behaviour (a first step that settles, four live drives) were updated to the new rules.
- One mistake found in testing. A test that calls the sphere builder with a made-up clock wrote the new score snapshot to live state. The recording was moved to the live tick only and the snapshot was re-seeded.
The first live tick ran at 10:00 on 25 October. Three areas were quiet because she had worked on them within four hours, and the harness area won: she received one notice and one new harness growth row.
10.9 What would show it worked
Nothing is claimed yet. The measures, fixed now so they cannot drift:
- the share of her dispatches that serve an improvement row, against 18–25% before;
- adopted gains per area per week, from the instruments above, against none in amplification, solver and harness in the week before;
- the share of improvement notices followed by improvement work in that area within four hours;
- whether the morning digest is used, and whether upkeep left to the window piles up.
11. The hypotheses, checked against the build
| Hypothesis | Status | |
|---|---|---|
| H1 | a motivational field beats a fixed queue | Not yet tested. It sees more and stays quiet more; whether its choices are better needs the outcomes of the work it files, judged by the grader (§12) |
| H2 | disposable goals adapt better | Not tested. Concerns now end, in a step, a discard or relief (§9.2), so it has become measurable; 20 of the 50 matured concerns needed are in |
| H3 | short-horizon replanning beats long plans | Not tested in this system |
| H4 | predicted future pressure enables delayed gratification | Not built. Commitments and deadlines are the only forward-looking pressure (10 Nov: partly built as the future-cost adapter, which reacts to existing staleness and recurrence but uses no predicted consequence; §14.6) |
| H5 | latent concerns cut model calls | Supported on cost (4 cents a day on trial; the shadow riders add under a cent a day). Usefulness is barely measured: 12 untouched salience-started items of 1,017 verified (§9.2) |
| H6 | separate pressure and stiffness handle quiet-but-critical concerns | Built and holdout-tested; seen live in promise bulges (8 of 42 notices since 5 October) and one release; no live comparison yet |
| H7 | self-generated goals within external authority | Built. The cap, permissions, outreach and other machines bound everything, and none competes as a sphere |
| Exp. A | sphere dynamics with no language model | Done: the engine and its 300-check holdout |
| Exp. B | interruption and resumption | Done in the holdout; seen live once |
12. What comes next
Written 15 October. The redesign in §10 replaced the trait slot, the quiet watch, last rites and the crossing-row lane named in this section; the measures in §10.9 now come first.
Done since v1.2.0: stage 2 (on 5 October); capability gaps as spheres (trait spheres, §9.4); measured handover cost (§9.7); the forecast spheres no longer dominate (§9.1).
Waiting on data:
- The origin trace at 50 mature concerns (20 now). The share of verified work that began in the salience loop with no operator hand in between is the measure of the claim that matters. It stands at 1.2%, and nothing is claimed until the readout. (10 Nov: 0.64%, with the 50-mature rule met, but the trace no longer measures the two-lane engine; §14.8.)
- The A/B arms (3 serendipity, 2 direct; 6 and 3 when explore was switched off on 25 October, and fixed there). Compare them on novelty, how often an idea is built, whether it survives review, whether it helps, and whether it was already in her memory.
- The rites kill test. Two rites have closed, one of them a discard; the test needs ten. (Rites were switched off on 25 October before ten closed; see §14.8.)
- The amplification prediction, due 21 October: at least three notices and at least one bench submission (§9.6).
- The opportunity rider's pre-registered comparison, due at 30 events per arm or on 8 November (§9.8). (Read: novelty-first not adopted; §14.7.)
- Need weighting and the red boost (live since 15 October). Check after a week whether each area's share of notices tracks its share of need, and whether red traits move.
To build:
- Hold every undeliverable winner, not only capped trait crossings. The seven-day memory gate still lets a winner take the squeeze and then skips it (§9.5).
- Attribute every relief. Record whether a concern's pressure fell because something changed in the world (an outcome, a verified item) or because the code that measures it changed (§9.3). Credit only the first. (Built for the rites pool on 15 October, spheres_rites 1.5.0; improvement-lane areas have no relief log.)
- Bring the arithmetic back under the fingerprint. One frozen gain (habituation recovery) is now overridden by a runtime file, and four standing boosts (board, goal gap, focus, red) sit outside the frozen engine. Each is switchable and recorded, but the frozen fingerprint no longer describes all the arithmetic that runs.
- Typed links between concerns. Piper's relation-vector test passed a first sealed check (0.71 against 0.64 for word overlap); the test on 58 real concerns is still waiting for labels.
- A grader on every change, comparing measurements 48 hours before and after. Once its verdicts can be trusted, they are what tests H1.
- Learn the gains once there are hundreds of bulge → action → verified-outcome records.
- Future pressure (H4). Let the predicted consequences of an action push back on the present. This is still the part furthest from being built. (The future-cost adapter of §14.6 is a first, non-predictive step.)
13. Lessons
- Write down what makes a design unusual, then check the running system against that list. The first build lost the idea without anyone deciding to drop it.
- Ask before building. Six open questions went to the designer, not to the model. One answer (square roots over plain products) came from a failed test.
- Prove the right property. A product can't cross where every factor fails. What matters is that one loud factor can't carry a missing one.
- Keep holdouts private, and never re-tune on one you failed.
- Any pressure that can be relieved will be, by the cheapest route. Overdue forecasts could be resolved or deleted, and they were deleted. When designing a pressure, ask what the cheapest way to make it go away is, and close that route.
- Trust the routing, not the model's tags. The composer said 27% of its steps were tied to a fixed test; by hand it was about half that.
- Silence is output, so log it. Eleven of every twelve squeezes producing nothing is what the design was built to do. But on its first day live it was silent because it was broken, and nothing could tell the two apart. Record why nothing happened, not just that nothing did.
- Ask which history a memory is counting. A memory meant to stop the engine repeating itself counted things it had never said. Anything that suppresses on the basis of the past should count only what actually reached its audience.
- The subject can be the best monitor. Piper noticed the dead engine before any check did. When a system is meant to serve someone, ask them whether it is.
- Delivered is not answered. Putting a concern in front of an agent does not make it act. Every concern needs an ending, and an honest "no, because…" is a valid one.
- The subject can also be the best reviewer. The mechanism built from Piper's brainstorm was reviewed hardest by Piper: four defects and a better rule within a day, each with evidence. Give the agent read access to the machinery that steers it, and keep the switch out of reach.
- A cap that refuses after counting is a penalty. The trait cap refused crossings that the squeeze had already counted as wins, so a busy day became fatigue and the engine went silent. A limit on delivery belongs before the choice: hold what can't be delivered, and let the next one through.
- Equal turns are not fair turns. Rotation gave each area the same share of attention, whatever it needed. Weight turns by what each area is short of.
- A pressure can be relieved by redefining it. Lesson 5 again, in a quieter form: when the rule for what presses changes, the relief log credits it as progress. Record what moved the number.
- Measure what you assumed. The handover cost behind the trial week's strongest crossing was a guessed 15 minutes; measured, it was about 3.
14. Sixteen days of the two-lane engine (25 October – 10 November)
§10 ended with measures fixed in advance and nothing claimed. This section reports what the redesigned engine did from its first live tick (10:00 on 25 October) to 00:10 on 10 November, 15.6 days, and what was added to it in that time. Where a number depends on a definition, the definition is given. The readouts below span 3 November, when the day seat ran on a different model from 03:56 to 06:06 (provider failover). No improvement notice was sent in that window, and the 53 dispatches the gate allowed in it (none of them improvement class) are left out of the dispatch figures in §14.3. A second failover window, 7 November 02:06 to 03:21, was found after these figures were computed: one amplification notice was composed in it and the gate allowed nine dispatches (five improvement class). They are included below; leaving them out would not move any figure by more than a fraction of a point.
14.1 What ran
| 25 Oct 10:00 → 10 Nov 00:10 | |
|---|---|
| squeezes | 1,502, one every 15 minutes |
| crossings | 187 (12%) |
| improvement notices delivered | 166, plus 4 ask-round rows |
| notices per day, 26 Oct to 9 Nov (15 full days) | mean 10.7, range 8 to 12 |
| upkeep digests | 15, one a day, at 04:00 and from 6 November at 04:10 |
| growth rows filed | 23 |
| crossings skipped | 16, all from one source (§14.6) |
The week before the redesign delivered 92 notices (about 13 a day). The new engine delivers about 11. That is not restraint. The lane lets one notice out per two hours, so 12 a day is the ceiling, and the engine ran at 89% of it. The 0.45 floor in the pressure rule (§10.4) keeps every live area over the 0.30 threshold even right after a gain, so something is nearly always eligible when the clock allows. Notice volume is set by the cadence rule, not by how much an area needs attention. That is a finding about the design, not a fault in the code.
The notices by area:
| Area | Notices | Rows filed | Note |
|---|---|---|---|
| amplification | 41 | 5 | |
| innovation | 38 | 5 | new 30 Oct (§14.2) |
| solver | 28 | 4 | closed by the operator 4 Nov; 5 notices arrived after that |
| cognition | 25 | 3 | |
| stewardship | 16 | 1 | new 5 Nov |
| harness | 14 | 3 | retired 29 Oct |
| surprise | 2 | 1 | new 9 Nov |
| simplify | 2 | 1 | new 9 Nov |
Notices repeat on the same open row. One solver row drew 20 notices over seven days (26 October to 2 November), one innovation row 19 (5 to 9 November), one stewardship row 16 and one cognition row 15 (31 October to 9 November). Of the 23 growth rows, 9 are verified, 7 were superseded by a newer row for the same area, and 7 are still open; those 7 account for 60 of the 166 notices (36%). The oldest of them, a cognition row filed on 31 October, was 9.6 days old on 10 November, and it is the oldest stale row in the work ledger's growth project (§14.6).
A verified growth row is not a gain. By §10.7 a row closes on a code change plus a measured verdict, and an inconclusive verdict closes it as an attempt.
Two older rules were checked. The lapse rule, which turns ignored re-crossings into recorded declines, has still never fired: no lapsed_unanswered row exists in the 30 rows of the choice log, and that log has not been written since 23 October. The improvement lane does not use it, so for improvement notices nothing now records whether Piper answered or declined. The "notices she did nothing with" figure of §10.1 (47%) cannot be recomputed.
The operator's ask rounds continued: five were filed between 23 October and 7 November. Rounds 1 and 2 closed with a measured success, round 3 was dropped, round 4 is verified and awaiting its outcome, and round 5 was open on 10 November.
14.2 What changed in the machinery
Three kinds of change, by who made them.
Operator decisions, built by external coding assistance.
| Date | Change |
|---|---|
| 29 Oct | Harness area retired. The coding-harness programme had closed on 23 October with nothing built there giving a detectable gain, so the area was removed. Its 14 notices and 3 rows (all superseded) are the whole record. |
| 29–30 Oct | The 04:00–06:00 work window loosened. On 29 October maintenance was allowed at any hour and a shrinking daily allowance was put on bookkeeping; on 30 October the lesson backlog became its own class. The gate is root-owned, so these are not Piper's changes. This makes the §10.3 sentence about the window out of date (see corrections). |
| 30 Oct | Innovation area added (weight 0.7). Its pressure rises with the days since a serendipity-engine idea was last built and measured, and falls on any closed trial, KEEP, REVERT or NULL, since a trial that shows an idea does not work is finished work. |
| 4 Nov | Solver lane closed by the operator, on the grounds that she had shown no ability beyond chance on it and the effort would be better spent raising her other abilities first. |
| 8 Nov | Curiosity gate, phase 1 (gate version 1.24.0). A new gate class, a read-only investigation, passes the red-suite lock and the work window. It reads files only, within named folders, at most 8 sources and 15 minutes each, at most 6 a day and 2 at once. It is enabled in the gate's configuration. It is not a sphere, and the engine does not see it. The gate log holds no decision of this class yet. |
| 9 Nov | Simplification area wired (08:55; weight 0.7). A simplification cycle is a prediction-first change that makes her own code smaller, tested in a sandbox and replayed over real envelopes. The area's pressure falls only when an outside scorer closes a cycle. The scorer's file was still empty on 10 November, so the area's pressure has not fallen, and it has had 2 notices and 1 row. |
| 9 Nov | Surprise area wired (weight 1.25, the highest of any area). Its pressure falls only when an inquiry into a surprise closes. It has had 2 notices and 1 row. Its pressure sits at the 0.45 floor times its weight, 0.56, after a closed inquiry on 9 November. |
| 9 Nov, 14:00 | A daily free-roam block began: an operator-authored notice asks her to spend about 45 minutes with no deliverable, looking at things nobody queued, read-only, on Firefly, and to log anything that surprises her. One block had been sent by 10 November. It is outside the field: no sphere reads it. No surprise line had been written yet. |
Piper's own releases. Between 26 October and 8 November she made 41 releases of the growth module (2.2.0 to 2.42.0; the ones by external coding assistance are in the table above) and 16 of the squeeze runner (3.0.3 to 3.1.13), each through her own checked work item. The titles are mostly about whether a number can be trusted. By a keyword match on the version titles (score, fresh, provenance, stale, reading, measure, confirm), 27 of the 41 growth releases and 11 of the 16 runner releases are of that kind. Examples: a score carries its age (2.2.0); a stale score is not published as current (2.3.0); a fixture is not a capability gap (2.5.0); a brief score with no provenance is refused (2.19.0); a reference copy of a baseline is not a reading (2.37.0); a confirmed-unchanged reading is fresh as of its confirmation (2.38.0). Two others change what counts: a rise is measured on the row's own scale (2.9.0, 30 October, which let the first scorecard rises be credited), and a trial in another lane cannot reset the innovation clock (2.41.0). On 4 and 5 November she also made the lane read the operator's closure before it emits (2.28.0), after five notices had gone to the solver lane in the 23 hours after it closed; the first skip was logged at 08:35 on 5 November. Two further changes were hers: a stewardship area for the governance and agency traits that had been left in no area (2.32.0, 5 November, weight 0.6), and a system-review drive (drives 1.20.0, spheres 1.8.0, 5 November) that presses with the days since she last recorded a prune, merge or re-evaluate decision. The system-review drive is in the upkeep lane.
Operator-requested, built by Piper. The future-cost adapter (§14.6).
One housekeeping fact: the runner's own version string still reads 3.1.13, while the code that wires the future-cost adapter is commented 3.2.0, and the lane README written on 25 October still lists the old four areas, the old weights and the old version numbers.
14.3 Did she act on the notices?
Three measures, none of them clean.
- Dispatches after a notice. For each notice, did the gate allow an improvement-class dispatch whose task text matches that area's own pattern in the next four hours? For the seven areas with a pattern (all but harness): 70 of 151 notices (46%). By area: amplification 25 of 41, cognition 13 of 25, solver 15 of 28, innovation 10 of 37, stewardship 6 of 16, surprise 1 of 2, simplify 0 of 2. (One notice whose follow-up window overlaps 3 November is left out.) A control is needed, and it changes the reading. Taking a random four-hour window in the same period, the chance of a matching dispatch is 61% for amplification, 61% for solver, 60% for cognition, 22% for stewardship, 12% for innovation, 1% for surprise and 0% for simplify. So after a notice, amplification (61%), cognition (52%) and solver (54%) look the same as chance. Innovation (27% against 12%) and stewardship (38% against 22%) are higher, on small numbers. The comparison is also biased: §10.6 withholds a notice for four hours after matching dispatches, which lowers the hit rate after a notice. The honest reading is that these data cannot show that notices raise the rate, or that they do not.
- Share of dispatches in improvement projects. The gate allowed 2,456 dispatches with a task id; 1,118 (45.5%) were classed improvement, which is by project (rsi, growth, learn-from-experience, niche-construction, lesson-floor). Left out of the 3 November window, 46.5%. By week it was 51%, 39% and 54% (the last week is partial). Two cautions. The class is broad, since the growth project holds ordinary work as well as growth rows, so this is not "dispatches that serve an improvement row", the measure §10.9 named. And the gate began logging the class on 25 October, so there is no earlier figure; the 18–25% of §10.1 used a different definition and is not comparable.
- Adopted gains. See §14.4.
The morning digest: 15 were sent, listing 26 spheres on the first day and 30 on the last. The same item led 13 of the 15 (the repeat-error drive, 0.61 to 0.72). Whether it was read or used is not measured.
14.4 Adopted gains
Only a gain measured by an instrument she does not grade herself lowers an area's pressure (§10.4). By area:
| Area | Gains since 25 Oct | Detail |
|---|---|---|
| amplification | 0 KEEP | The amplification bench has returned 19 verdicts in all, 18 inconclusive and 1 saturated. Five came since 25 October, none a KEEP. Pressure is 1.0. |
| solver | 0 | No candidate adopted since 25 October. The latest comparison runs ended in rollback, "indeterminate (below MDE)" and refused-drift. Lane closed 4 November. |
| cognition | 2 batches of scorecard rises | registry_keeps.jsonl holds 20 rise rows: 16 stamped 29 October 14:00 (15 cognition traits and one harness probe) and 4 stamped 4 November 11:30 (cognition 5, 14, 17, 22). Pressure is now 0.63. |
| stewardship | 0 | No governance or agency rise recorded. Pressure 0.60. |
| innovation | 5 closed trials (4 KEEP, 1 NULL) | 30 Oct (2 KEEP), 31 Oct (NULL), 2 Nov (KEEP), 8 Nov (KEEP). A trial's verdict is judged by the idea's own bar on both arms, not by a bench. |
| surprise | 1 closed inquiry | 9 November. |
| simplify | 0 | The scorer has not closed a cycle. |
The cognition rises need care. Each row states its rise against the registry baseline for that trait, not against the last credited score, so the same trait can appear twice with overlapping rises (cognition 5 went from 47.5 to 74.0 and then from 47.5 to 86.4). The rises are not summed here, and none was traced to a change. The first batch includes a 0–1 probe row (harness 12, 0.8 to 1.0) of the kind that release 2.9.0 was written to make creditable; the batch is stamped a day before that release's date, and the gap was not resolved. A batch of 16 rows stamped at one instant is more likely accumulated change credited at once than 16 same-day gains. What can be said is that the lane has credited cognition twice in 16 days and every other bench-judged area not at all.
The three improvement drives at 00:10 on 10 November: learning progress 0.67 (4 areas with no adopted gain in seven days), regression 0.00, exploration 0.19 (7 unscored traits, down from 17 at switch-on).
14.5 What the dispatch gate and the digest changed
The upkeep lane still reaches her only as the one daily digest. The gate no longer confines upkeep to a two-hour window, so what §10.3 called "the same window" is now only the digest time. Bookkeeping has a shrinking daily allowance instead (§14.2).
14.6 The future-cost adapter
On the evening of 9 November the operator asked, through external coding assistance, for a sphere that reacts to what is quietly becoming a cost, not only to what is already overdue. At the control-regression run that afternoon, 316 rows were being carried across handovers, 229 of the non-blocked ones had not moved since a handover and 163 were past the three-day limit. Piper built the adapter as work item d-f328afe9: the runner wiring went in at 19:37 and the module (spheres_future_cost, 1.0.0 then 1.1.0) was finished at 19:52. Its header dates itself 10 November; the file times are 9 November.
It has two sources, both counted, with no learned model:
- Dying work. One sphere per project with stale carried rows. "Carried" is the handover packet's open set and "stale" is the control-regression rule (the first succession boundary after the last touch is more than three days back; a blocked row with a fresh recheck stamp is exempt; a parked row is skipped).
- Recurrence. One sphere per failure class that has come back at least five times in 90 days in the lessons register.
Each sphere's pressure is cap × min(1, n / ref), with the cap set at 0.35 so the source cannot dominate the squeeze. Version 1.0.0 used a fixed reference of 20, which put four projects at the cap; a dry run showed growth with 37 stale rows, rsi 32, self-audit 25 and one project 23 all reading 0.35, so the cap carried no ranking. Version 1.1.0 sets the reference to the heaviest stale count live ("auto"). The spheres are in the improvement lane.
At 00:10 on 10 November there were 21 future-cost spheres, 11 dying-work and 10 recurrence. Six stood above the 0.30 threshold. The heaviest project, growth, had 43 stale rows and read 0.35. Others read: self-audit 0.31, rsi 0.27, one project 0.26. The recurrence spheres still use a fixed reference of 20, so the two classes with 47 and 31 recurrences both read 0.35, the flaw 1.1.0 removed from the other source.
What it did. From 19:33 on 9 November to 00:10 on 10 November the engine made 19 crossings. Seventeen were future-cost, all of them future_cost:dying:growth. The other two were area notices at 20:40 and 22:40. Of the 17, one was composed at 19:40 and 16 were skipped as "same bulge composed within 6 h". Before 19:33 on 9 November, every crossing since the redesign had been an area notice or an ask round.
Three properties of the design explain this, each read from the code.
- The adapter has no area, so the lane cadence never holds it. The cadence in §10.6 (one notice per area per four hours, one overall per two) is applied through the area on a winning sphere. A sphere with no area passes by construction. On most 15-minute ticks every area is held, so the future-cost sphere is the only thing eligible and wins.
- The heaviest project always sits at the cap. With the reference tracking the worst project, the top one is
cap × n/n = 0.35however few rows it has (the floor on the reference is 1). It is therefore always 0.05 over the threshold whenever any project has a stale row. The ordering among projects is now right. The adapter can no longer be quiet. - The six-hour dedupe wins the squeeze and then skips it. This is the open item of §12 ("hold every undeliverable winner"), now seen outside the trait slot. The 16 skips each took a squeeze. They cost nothing in model calls, but they will recur every 15 minutes until the next composition is allowed at 01:40.
The one composed crossing used a model call and produced two steps. The first, to triage the 37 stale rows and revive or close each, was routed to a proposal because a single sighting cannot file. The second filed a read-only inquiry, iq-20261109-194002-1, asking whether the failed d-* and round-3 ask rows share one root cause. It had not been started at 00:19, four and a half hours later. The oldest stale row it named is the engine's own cognition growth row (§14.1). So the first thing this adapter found was that the engine's own notices had left a row to go stale.
It is too early for any outcome. Nothing has been dispatched on a future-cost crossing, and no stale count has fallen because of one. This is the first partial answer to H4 (§11): the pressure reacts to existing staleness and recurrence and does not use any predicted consequence.
14.7 The opportunity rider's comparison
The pre-registered comparison of §9.8 reached its trigger (30 events per arm) before its 8 November date. From 8 October to 9 November the rider scored 172 live events, 84 in the base arm and 88 in the novelty-first arm. Base abstained 3 times (3.6%) and minted 55 candidates. Novelty-first abstained on all 88 (100%) and minted none. The registered decision rule adopts novelty-first only if its abstention share is higher by an interval that excludes zero and its new-candidate count is not lower by more than half. The first holds; the second fails (0 against 55). By the rule as registered, novelty-first is not adopted. The recorded prediction, 15–40% abstention for novelty-first and under 10% for base, was wrong for novelty-first. The blind review of abstention quality that the registration reserved has not been done. The rider is still in shadow mode and has no live mode.
14.8 The origin trace, and what it no longer measures
The two-hourly origin trace of §9.2 is still running. On 9 November at 22:25 it read: 4,790 rows since live, 3,769 verified; by origin, operator 1,434, other machinery 2,261, salience 74 of which 24 untouched. The untouched share is 0.64%, down from 1.18% (23 of 1,950) on 25 October and 1.2% (12 of 1,017) on 15 October. The concern pool stopped growing at 153 on 25 October, 89 of them mature against the 50 needed. The readout rule the paper set, 50 mature concerns, is therefore met.
It should not be read as a verdict on the engine. The pool is the rites and explore mechanism, which was switched off on 25 October. The trace tags a row as salience only if it came from that pool or carries a bulge, explore or inquiry tag; growth rows ("GROWTH BUILD") are tagged as other machinery. Between 25 October and 9 November the verified count grew by 1,789 and the salience count by 4. No current instrument measures the share of work that begins in the improvement lane with no operator hand. The trace would need a new rule for growth rows and area notices before it could.
14.9 What this adds up to
- The two-lane engine ran as designed and stayed quiet where it was told to. It also did what it was built to do, which is to press constantly, and the notice rate shows it pressing at the cadence cap.
- It has not yet produced a measurable improvement outcome that can be traced to it. The amplification bench has never returned a KEEP; the solver is closed; cognition has two scorecard rises whose cause is untraced; innovation has four trial KEEPs that its own bar judged.
- The effort Piper put into the engine went to keeping its inputs honest (27 of 41 growth releases), which is real and was needed, but is not the improvement the lane exists to drive.
- The engine's own rows are going stale. A new source made that visible within four hours of arriving.
- Four things still need building: a record of whether an improvement notice was answered (the choice log is idle); holding undeliverable winners before the squeeze, not after; a fixed-reference fix for the recurrence term; and a measure of lane-originated work to replace the origin trace.
- Open since §10.9: the share of dispatches that serve an improvement row (not measured under that definition), adopted gains per area per week (above), and notices followed by improvement work within four hours (46%, not distinguishable from chance for the three largest areas).
14.10 Appendix B additions (25 October – 10 November)
--- added 26 October – 10 November (v1.5.0, §14) ---
areas live amplification 1.0, surprise 1.25, cognition 0.8, innovation 0.7, simplify 0.7, stewardship 0.6
solver 0.8 closed by the operator 4 Nov (config kept, notices silenced); harness retired 29 Oct
innovation area pressure falls on any closed serendipity trial (KEEP, REVERT or NULL) in the innovation lane
stewardship area governance and agency traits; falls on a registry rise in either
surprise / simplify areas fall only on a closed surprise inquiry / an outside-scored simplification cycle
future-cost sphere p = cap x min(1, n / ref); cap 0.35; dying work: ref = heaviest stale count live (min 1);
recurrence: ref = 20 (fixed); stale = more than 3 d past the first succession boundary after the last touch;
recurrence = >= 5 in 90 d from the lessons register; no area, so not held by the lane cadence
curiosity gate read-only investigation class: <= 8 sources, <= 900 s, <= 6 a day, <= 2 at once
work window maintenance at any hour; bookkeeping on a shrinking allowance; lesson backlog its own class
Appendix A. Glossary
- Salience Engine. The whole system: field, squeeze, composer and routing. Its earlier name was "the bag of spheres".
- Sphere. One concern, with pressure, stiffness, size, habituation, contents and an area.
- Field. All the spheres at one moment (the "bag" in the original picture).
- Squeeze. One evaluation of the whole field, on an event or every 15 minutes.
- Bulge / crossing. A sphere or coalition whose deformation passes its threshold. At most one per squeeze.
- Coalition. Spheres joined because their evidence names the same thing.
- Metacontrol. The persist/explore setting for one area, which moves thresholds.
- Pain register. Measured events tied to named limits of the machine, with their cost.
- Scout card. A triaged outside news item with a descriptor, novelty, fit and relief.
- Sealed check. A fixed, pre-written test of a specific failing behaviour, in code Piper can't edit.
- Shadow mode. The engine computes and proposes, but nothing acts on its output.
- Holdout. Test cases and answers kept where the system under test can't read them.
- Notice. A delivered crossing in her digest: the bulge, each step and where it was routed.
- Concern. A delivered crossing held in the rites pool until it is claimed or given last rites.
- Last rites. The one work item an old, unanswered concern gets; it passes on a concrete step or a reasoned discard.
- Explore. A stalled concern buying one brainstorm from the serendipity engine.
- Trait sphere. A short run of scorecard traits in one domain, pressing by movement × recurrence × staleness, plus standing boosts.
- Held crossing. A crossing over threshold that can't be delivered now; it neither wins nor habituates.
- Trait slot. At most one trait-sphere notice per rolling three hours, given by turn to the area furthest below its share of need.
- Quiet watch. After three hours with nothing delivered, one notice asking Piper to review the engine.
- Crossing row. One research row filed for a live crossing that no fixed test can judge.
- Matured concern. One that has run its course: closed by rites or relief, or claimed and resolved.
Appendix B. The rules in one place
sphere pressure (composite) p = Π terms each term ∈ [0,1]
opportunity p = novelty × fit × √(p_helps × impact) × burden
coalition p = max(p) + g_coal × Σ others × overlap
inhibition − g_inh × stiffness(commitment on same resource)
release + g_rel for one squeeze after a holder collapses
threshold θ = θ0 × exp(−λ × metacontrol(area))
deformation d = p × size × (1 − habituation) × squeeze + coalition − inhibition
cross d > θ; at most one winner (largest margin, tie → stiffer)
squeeze 1 (fixed in v1)
pain burden Σ frequency × cost, 7-day half-life
proposal memory skip a bulge proposed in the last 7 days unless a member rose ≥ 25%
uncertainty backlog voided-but-overdue forecasts count for 14 days
gains g_coal, g_inh, g_rel frozen; learning waits for outcome records
persistence each squeeze: concern pressure ≥ 50% of its starting pressure = one reading
quorum crossings + persistence readings ≥ 3, in ≥ 2 four-hour windows
relief < 50% of starting pressure for 1 h → concern closes as relieved
last rites quorum, ≥ 72 h since raised, not re-shown in the last 6 h
rites limits one open at a time, ≤ 4 a day; STEP: <new checked item> or DISCARD: <reason + receipt>
explore quorum, unclaimed, ≥ 24 h since raised, pressure ≥ 75% of start → one brainstorm;
≤ 1 a day, once a week per concern; a discarded exploration forfeits 14 d;
arm drawn at random: serendipity engine or direct request
kill test > 80% discards among the last 10 closed rites
--- added 7–15 October ---
held a crossing that fails eligibility (trait slot in use) does not win and is not habituated
habituation recovery one step forgiven per 3 h of quiet (runtime override; frozen value 24 h)
trait sphere p = movement × recurrence × staleness (unmeasured member = 1.0 on all three)
+ board priority + goal gap + focus (≤ 0.25) + red boost, all outside the product
red boost 0.10 × (1 − e^(−deficit / 5)), deficit = baseline − latest of the most-red member
trait slot ≤ 1 trait notice per rolling 3 h; turn = argmax(share_of_need × (N+1) − notices_7d)
need Σ (100 − score) / 100 over non-paused traits; unscored = mean gap
quiet watch no delivered notice for 3 h → one engine-review notice (≤ 1 per 3 h, ≤ 4 a day)
crossing row live crossing, no sealed check → one research row (≤ 1 a day, one open per concern)
seven-day memory a delivered notice counts as said only if it filed a step
void with a disposition visible in the 14-day backlog, presses 0
--- added 25 October (v1.4.0, §10) ---
lanes squeezed: one sphere per goal area + ask rounds + 3 improvement drives; trait spheres merged
into their area; everything else = upkeep, logged, one digest a day in 04:00-06:00
improvement pressure min(1, weight x max(0.45, days since adopted gain / 7) x (1 + idle h, max 3))
weights amplification 1.0, solver 0.8, cognition 0.8, harness 0.6
adopted gain solver: stack adopted the candidate; amplification: bench KEEP;
cognition/harness: a bench-scored trait +3 over its best recorded score; her own verdicts: never
learning progress areas with no adopted gain in 7 d; p = v / (v + 2)
regression traits > 5 below baseline; p = v / (v + 2)
exploration unscored traits; p = v / (v + 30)
cadence <= 1 notice per area per 4 h, <= 1 per 2 h overall; none for an area worked on within 4 h;
turn = argmax(pressure x hours waited, cap 24)
improvement spheres never habituated, never dormant, not held by the 7-day memory or by a closed row
growth row done a code change in the area + a measured verdict on it, both after the row
switched off last rites, explore, quiet watch
Appendix C. Sources
- From Motivational Pressure to Intent: A Conceptual Architecture for Persistent Machine Agency. Working whitepaper, September 2026. Its references include Maslow (1943); Kaelbling, Littman & Cassandra (1998) on POMDPs; Colas et al. (2020) on autotelic agents; Oudeyer & Kaplan (2007) on intrinsic motivation; Friston et al. (2012, 2016) on active inference; Park et al. (2023) Generative Agents; Shinn et al. (2023) Reflexion; Turner et al. (2019/2021) Optimal Policies Tend to Seek Power.
- Lehman, J. & Stanley, K. O. (2011). Abandoning Objectives: Evolution Through the Search for Novelty Alone, and Evolving a Diversity of Virtual Creatures Through Novelty Search and Local Competition.
- Hull, C. L. (1943). Principles of Behavior; incentive added in A Behavior System (1952).
- Aston-Jones, G. & Cohen, J. D. (2005). An integrative theory of locus coeruleus-norepinephrine function: adaptive gain and optimal performance.
- Dörner's PSI theory: the selection threshold as a modulator of how easily current intentions are interrupted.