calibration: decision-level confidence is a constant 1.0 — derive it from finding-level signal instead

[JSONbored]

Problem

Empirical finding from the #8170 retro analysis, and the single most important measurement the corpus has produced: every one of the 460 historical close-class decisions recorded decision_json.confidence = exactly 1.0 — zero variance — while 38% of them were reversed (same-PR reopen-and-merge or same-author rework that merged). The decision layer claims certainty on every close and is wrong on a third of them. Consequences:

• Any confidence floor applied at the decision level has never held anything back — the signal it gates on does not vary.
• The backfilled corpus cannot discriminate close-confidence thresholds at all (every case sits above every candidate), so the knob correctly proposes nothing from it — the discipline held, but only because the loop refuses to act on flat evidence.
• The FINDING-level confidences the live capture writers record DO vary — the go-forward corpus is sound; the broken layer is specifically the decision aggregate.

Requirements

⚠️ Required pattern. Find where the decision-level confidence is produced (the review_targets decision writer) and replace the constant with a derivation from the contributing findings' own confidences — the natural candidate is the MINIMUM across the findings that drove the verdict (a close is only as certain as its least-certain load-bearing finding), but audit the writer first and record the chosen aggregation with rationale beside it. Behavior-pinning: the decision VERDICT must be byte-identical — only the recorded confidence changes.

• Backfill provenance note: rows synthesized by calibration: backfill the corpus from the 10k+ historical review ledger (labels + firings; raw context phase 2) #8157 carry the degenerate 1.0 — never re-derive them retroactively (the corpus era-tagging already distinguishes them); the fix is go-forward.
• Add an invariant test asserting the recorded decision confidence varies with finding confidence inputs (the exact failure mode this issue documents).

Links & Resources

#8211 (epic), #8170 (the analysis), #8226/#8227 (reliability curves — this fix is what gives them a real axis at the decision level)

Boundaries

Measurement only: no gate-authority change, no threshold movement.

maintainer-only — decision-writer instrumentation.

[linear-code]

JSO-5716

[JSONbored]

Scope reconciliation for the record (the gate's linked-issue concern on the closing PR was fair): this issue's premise — a LIVE decision writer stamping constant 1.0 — was disproven during implementation. The constant-1.0 rows all came from the retired legacy content-gate pipeline (review_targets decisions stopped 2026-06-22); the live capture path records finding-level confidences that genuinely vary, so there is no live derivation to build. What shipped instead was the piece that still had value for the flat era: reason-code segmentation (both stores, 460/460), which produced the per-reason reversal table above.

The one requirement that survives the reframe is the invariant test — pointed at the LIVE path rather than a retired writer: an explicit pin that the capture writers' recorded confidence VARIES with finding inputs, so a future regression to constant-confidence capture cannot land silently. Shipping that as a test-only follow-up PR now.