Summary / The 27 components
08 Context construction
Deciding exactly what goes to the model next: source, related symbols, tests, history, errors, requirements.
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.
What is built now
Nothing beyond the task statement and the workspace. Every arm's worker starts from the same short prompt.
What is missing
The before_agent_start assembler (task, manifest, localization top-k, retrieval slice, failing-test output, prior attempt and diagnosis on retry), and a per-state context ceiling.
Reviewer Claude, 2026-10-04; close-out verdicts 2026-10-23
Reviewer's read: Closed 10-22: built as the hx localize stage; no detectable gain
This is the component that would carry 03, 04, 05 and 13 into the session, so it waits on them. The spec's 24K ceiling, ranking call and lost-in-the-middle pinning are reasonable but unmeasured. One condition from the 09-07 review matters for any of it: the worker must not inherit unrelated context. That review found the early baseline workers were loading her rulebook and skills; the fix (--no-context-files, no skills, workspace outside the harness tree) is what keeps arms comparable.
Close-out, 10-22 (pre-registered, boxed, Claude): built for real as stage S1 (localize) of the close-out bundle arm hx (A12): h0b + localize pre-pass + repro-first with failure digest + repro-triggered reflect/repair (2 rounds) + critic with one revision. hx against h0b, verified-hard 38 + pro-hard 20, 3 seeds: 103 against 94 cells (verified 65/114 against 55, pro 38/60 against 39). Gain +3.0 tasks of 58, paired bootstrap CI [−1.0, +7.0]; 4 regressions on tasks h0b always solved (limit 2); 2.9× h0b cost per cell. No detectable gain at this model. The pre-registered rule closes every bundled component together, with no halving round. Terminal: not re-run.
How it is measured no detectable gain (hx, 10-22)
Arm per added block, each against h0-v41, cheapest first.
Adoption rules shared by every arm: How the tests work.
Research & attempts
Score history
No scores yet. The first entry is the baseline for suite v1.
Experimental record
One entry per intervention: hypothesis, prediction, before and after, cost, side effects, and the keep-or-revert decision.
No interventions recorded yet.