Amplification scorecard
The cognition, agency and governance scorecards cover how well Piper thinks, whether she can pursue goals on her own, and whether her evidence about herself can be believed. This one covers the core of recursive self-improvement: whether her improvements make her better at producing further improvements. Changing her own code is not enough; the change has to improve the process that makes the next change. Twenty-four components in six layers, from the mechanics of one improvement cycle up to evidence of a self-reinforcing one. None has been reviewed against her system yet. As with the other three, the plan is to pick one at a time, starting with #5 and #6, test it, and rescore.
Shared copy: the operator's name is replaced with "the operator", and machines other than Firefly are named by their role. Nothing else is changed.
| # | Component | Reviewer read | Suite | Baseline | Latest | Change | Re-read | Live use | Trend |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Improvement detection | She finds most of her own targets now; nothing records what she missed | v2 | 87.5 | 87.5 | — | 95.8 | — | |
| 2 | Hypothesis formation | Predictions she writes herself are mostly testable (85.7 on 35 rows); most version changes still carry none (118 of 148) | v2 | 84.4 | 84.4 | — | 85.7 | — | |
| 3 | Intervention selection | Each dispatch names one route not taken; there is no candidate set, no cost or value estimate, and nothing scores the choice | v2 | 100 | 100 | — | 100 | — | |
| 4 | Change execution | Her changes land and are live; one (memstore 1.26.1) is on disk but not loaded by 2 processes started before it | v2 | 80 | 80 | — | 87.5 | — | |
| 5 | Outcome attribution | Judges its own changes on pre-set rules; the resolver separates on/off blocks only where a run marker exists, and no switch history is recorded | v2 | 93.8 | 93.8 | — | 87.5 | — | |
| 6 | Feedback integration | Few rows cite an earlier result about the change (7.1): 25 of 35 declare that no earlier outcome applies | v2 | 6.2 | 5 | ▼ -1.2 | 7.1 | — | |
| 7 | Failure learning | Code fixes leave a lesson; her solver loop's failures leave only a verdict file, and the one hand-written lesson there is half wrong | v2 | 87.5 | 87.5 | — | 87.5 | — | |
| 8 | Strategy adaptation | She changes method when the numbers say so, quickly for the capture trigger and on a partly wrong reading for the solver loop | v2 | 100 | 100 | — | 100 | — | |
| 9 | Transfer learning | Lessons move between her dispatch habits; between capabilities she found zero cases, and her count of that never reached the document | v2 | 85.7 | 85.7 | — | 85.7 | — | |
| 10 | Improvement memory | Retention works and reads are logged; before a change she mostly reads only the always-loaded docs (15 of 22) | v2 | 0 | 12.5 | ▲ +12.5 | 28.6 | 68 | |
| 11 | Improvement throughput | 104 changes settled right in the window, 21 of them on a measured result; the rest settled at mechanism level | v2 | 25 | 25 | — | 100 | — | |
| 12 | Resource efficiency | Cost per verified gain is readable: $36 in 14 days, $27 of it solver scoring | v2 | 31.4 | 31.4 | — | 100 | — | |
| 13 | Experiment efficiency | Her solver tests are too small for the bar: 9 of 10 could only see gains of +0.09 or more | v2 | 0 | 0 | — | 10 | — | |
| 14 | Search efficiency | One capability keep in her last 10 solver sessions; one session ran without stated odds | v2 | 0 | 0 | — | 10 | — | |
| 15 | Self-improvement capability | She builds improvement machinery fast and documents it well; Claude finds the errors in it, and using it is optional | v2 | 100 | 100 | — | 66.7 | — | |
| 16 | Second-order improvement | Most version changes still land with no prediction (118 of 148) or no row at all (12) | v2 | 50 | 50 | — | 10.8 | — | |
| 17 | Cross-component amplification | One real cross-component effect, measured honestly; it is reach inside one pipeline, not a gain in a different capability | v2 | 87.5 | 87.5 | — | 84.4 | — | |
| 18 | Compounding improvement | Solver gains shrink per generation (first three +7.8 pts, last three +0.9); v6 is the exception | v2 | 0 | 0 | — | 5.5 | — | |
| 19 | Bottleneck identification | She hunts the binding constraint before changing things, but reads whole-history logs that mix fixed and unfixed versions | v2 | 100 | 100 | — | 100 | — | |
| 20 | Bottleneck migration | Most fixes never get a re-read afterwards (14 of 15 in the window) | v2 | 80.8 | 80.8 | — | 0 | — | |
| 21 | Parallel improvement | Parallel work is routine; 824 overlapping changes and 90 duplicate deliverables blur each other's measurements | v2 | 47.8 | 47.8 | — | 38.3 | — | |
| 22 | Integration stability | Older capabilities hold (older set 1.000 on 10-20); two kept solver versions read below their parent | v2 | 84.4 | 84.4 | — | 90 | — | |
| 23 | Improvement persistence | Early solver keeps held; the v4 and v5 gains shrank below half later; v6 holds (+8.4) | v2 | 60 | 60 | 0 | 66.7 | — | |
| 24 | Autonomous recursive continuity | 0 cycles under v3 (11-10): 9 of her changes read an earlier decisive result first, none has an outside verdict yet; rev b 37.5, rev a 100 | v3 | 0 | 0 | — | — | — |