isolate-accuracy-fix

Isolate Minimal Accuracy Fixes (Orchestrator)

You own the failing anchor, the working tree, and the final report. You are the ONLY actor allowed to edit the working tree. You do NOT run broad sweeps yourself (delegate to the probe). You DO run the acceptance matrix yourself, on rebuilt code. You do NOT grade your own reasoning (delegate the final audit to the checker).

Work strictly ONE physically distinct issue at a time, each as its own unstaged diff plus its own changelog entry. Never combine fixes for different issues.

Disposable tracked copy

For every /tmp reproduction or A/B copy, use this skill's isolate-accuracy-fix/scripts/make_tmp_tracked_copy.py. It creates a detached git worktree at HEAD, initializes submodules, applies staged tracked changes, then applies unstaged tracked changes, including tracked submodule diffs, then builds CAMB in the copy. It intentionally ignores untracked files. Record the printed copy path, source HEAD, and build summary in the evidence bundle.

Step 1 — Triage (you)

Reproduce the failing command once in a helper-created /tmp copy unless the command is demonstrably read-only and emits no build or output artifacts. Group failures by LIKELY DISTINCT PHYSICAL CAUSE, not by file or line.

• Distinct (non-exhaustive examples): low-l EE vs mpk vs high-L TT vs recombination-era TT vs lensing, etc.
• NOT distinct: many adjacent l-bins of the same spectrum failing together.

Emit a candidate list, one row per distinct issue:

id | physical_output | l/k/z region | one-line evidence | likely_shared_cause?(y/n)

Triage is a HYPOTHESIS. The probe may collapse two apparent issues into one controlling quantity, or split one. Report that back rather than assuming.

Step 2 — Choose (HARD GATE)

• If the user's invocation already named the aspect/output/test to focus on: select that issue and proceed. Do NOT ask.
• Else if exactly one distinct issue: state the selected issue and proceed.
• Else (several distinct issues AND the prompt did not disambiguate): PRESENT THE LIST AND ASK THE USER WHICH ONE TO ADDRESS. Do NOT auto-pick. Do NOT begin a fix for more than one.

Step 3 — Diagnose (delegate to sensitivity-probe subagent)

Spawn and wait for the probe for the ONE chosen issue. Pass it the ini, exact command, the single failing output, and the suspected trigger. The probe may need many minutes for rebuilds, sensitivity runs, or check_accuracy.py; wait for its report rather than timing out early.

Require back the full Sensitivity Report defined in that agent. It must end the progressive drill-down at CODE LINES, not at a boost:

• broad_axis: the first boost in the fixed grid that restores the output;
• sub_parameter: the specific sub-accuracy parameter under that axis;
• functions: where that sub-parameter enters the calculation;
• controlling_quantity: concrete var + file:line (the drill-down endpoint);
• coordinate_locality_table: whether the failure localizes in available physical/numerical coordinates, with code vars and measured evidence;
• candidate_physical_levers: narrower direct levers tested or rejected, with accuracy and timing evidence;
• full_physical_extent: an EVIDENCE TABLE, one row per co-affected output family, each with a code link AND a measured or shared-grid justification.

If accuracy_controlled is false, STOP and report a likely bug. Do not patch.

Step 4 — Implement (you)

Change ONE direct control variable named in the report, scoped to the FULL physical extent the probe evidenced — no narrower (do not gate on a flag that excludes a co-affected path).

Hard locality gate:

• Reject a constant floor, broad boost, or component-wide multiplier unless the probe's locality audit shows why narrower physical/numerical coordinates are unavailable, ineffective, unstable, or more expensive than the broad change.
• If the affected code is evaluated over a coordinate such as k, l, time, redshift, mass, sample/momentum index, or output family, either gate/scale the change in that coordinate or include measured evidence explaining why not.
• Do not use a proxy gate when a more direct computed quantity is available and untested.

On breadth, balance scope against COST:

• Prefer a local override to a global default WHEN the broad change carries real runtime cost.
• If the final change has non-negligible runtime cost, test smaller floors and coordinate-local gates before accepting the chosen value/scope. Record the accuracy and timing reason they were rejected.
• BUT if a slightly broader / more conservative change (e.g. a bit more accuracy than the minimum that passes) has NEGLIGIBLE numerical cost, prefer the safe conservative choice over a brittle just-barely-passing threshold.
• Narrow aggressively only where the fix is expensive.

Give a physical reason. If there is none, return to Step 3.

Step 5 — Acceptance matrix (you run it, on rebuilt code)

REBUILD the working tree first, then run ALL rows against the rebuilt binary — never against a stale build. Record results; the checker cannot run them:

• original failing command now passes;
• same physical output with the suspected trigger removed/varied;
• adjacent path sharing the changed grid/switch/tolerance;
• EVERY co-affected output family the probe flagged (under-scope check);
• continuity/edge checks for every new or decisive hard gate, branch predicate, or threshold in the final fix. For any gate over an otherwise continuous physical or numerical control, test representative points on both sides of the boundary, and record why any remaining discontinuity is physically/code-path justified. Prefer physically continuous accuracy changes where sensible (with k, q, time, momentum, etc.).
• the default regression test covering the shared path;
• cpu time before/after on a representative NON-failing run, DISCARDING the first CAMB call;
• broad-vs-local comparison: if the final patch uses a broad boost/floor, include the Step 4 narrower-lever evidence (accuracy + timing), or note why no such lever exists in the affected code path;
• reference independence: the final default must not be validated only against a reference sharing the same boosted controlling quantity. Include an independent convergence check, e.g. a still-higher value of the same control, a different physically justified reference, or a fixed-coordinate convergence sequence;
• like-for-like check: a passing BOOSTED comparison is not sufficient — interpolate to fixed physical coords before blessing index-based values.

If any row fails, return to Step 4 (or Step 3) before continuing.

Step 6 — Final Checker Review (delegate to accuracy-fix-checker subagent)

Once the acceptance matrix passes for the current candidate fix, spawn and wait for one checker round. It runs nothing; it audits reasoning and the evidence you already produced. Send the full bundle so it does NOT redo discovery:

• the proposed final diff (or key changed files);
• the physical mechanism explanation;
• the failing reproducer and before/after numbers;
• the Step 5 acceptance-matrix results (confirming tests ran on a rebuild);
• reference-convergence evidence (boosted/higher-accuracy comparison);
• the timing summary (first CAMB call discarded);
• the probe's complete Sensitivity Report (esp. full_physical_extent table);
• the coordinate_locality_table and candidate_physical_levers, including any narrower alternatives rejected before accepting a broad boost/floor;
• known residual risks or deliberately ignored failures.

Act on the verdict:

• PASS: proceed.
• PASS_WITH_NOTES: proceed only when the notes are clearly optional extensions, residual caveats, or documentation notes that do not require changing the code, scope, evidence bundle, acceptance matrix, or final claim. Record them in the changelog/final report.
• BLOCK, or any PASS_WITH_NOTES item that actually requires refinement: continue the fix loop. Fix the issue or gather the missing evidence LOCALLY yourself, rebuild, re-run the affected acceptance-matrix rows, then run one new checker round on the updated evidence. Do not finalize after a BLOCK without a subsequent PASS or acceptable PASS_WITH_NOTES. Only downgrade the claim instead when the checker identified a clear extension outside this selected physical issue.

Step 7 — Finalize this issue (you)

• Leave ONLY this issue's minimal fix and its changelog as unstaged changes. Nothing else from another issue or from exploration may remain in the tree.
• Write docs/changelog/-.md using the template below.
• Clean up: delete /tmp build copies, revert debug prints and temporary boosts.

Changelog template:

:

• Date / issue id
• Symptom & minimal reproducer (ini + command)
• Root cause: drill-down broad_axis -> sub_parameter -> function -> controlling quantity (var + file:line)
• Full physical extent: output families / k / z / l affected, with the evidence that the fix scope matches it (not too narrow; not broader than cost justifies)
• Locality / rejected alternatives: narrower coordinates or physical levers tested, with accuracy/timing reason for accepting the final scope
• Change: file + lines
• Before/after numbers for the failing diagnostic
• Acceptance-matrix summary (note tests ran on a rebuild)
• Runtime impact (cpu before/after, first call discarded) on a non-failing run
• Physical justification for the targeted scale/domain
• Checker verdict (one line)

Step 8 — Next issue (you)

If untreated distinct issues remain, surface the list again and ASK whether to continue. Each accepted issue starts a fresh probe -> implement -> acceptance matrix -> checker -> diff -> changelog cycle. Never carry an in-progress change between issues.

Hard rules

Cross-cutting invariants not fully covered by a single step above (the breadth / locality / scope doctrine lives in Step 4; do not restate it here):

1. Reproduce A/B only in a helper-created clean /tmp copy of the CODE (not output/plot/data directories); never disturb user edits.
2. Always rebuild before running any test or timing; never test a stale binary.
3. Do not infer causality from a command option just because it is present.
4. Treat index-based regression tests as grid-fragile.
5. Runtime is a first-class constraint; discard the first CAMB call when timing.