calibration: week of 2026-07-06

[github-actions]

Weekly review-prompt calibration for the week of 2026-07-06, synthesized from the raw verdict-labeled issues in HarperFast/ai-review-log (pre-fetched directly). Curated supplements Calibration log #846 and False-negative log #870 exist but carry only scaffolding headers this week (their populating routines have stalled), so all synthesis is from raw labels.

Verdict mix

102 issues triaged: useful 98 · noise 3 · partial 1 (96% useful). 13 verdict:pending issues remain open, so the week is largely but not fully settled. Findings were sparse — 10 of the 98 useful entries carried an actual blocker/finding; the other 88 were genuine "no blockers" passes.

Per model:

Model	useful	noise	partial	total
claude-sonnet-4-6	94	1	1	96
gemini-3-flash-preview	4	2	0	6

No claude-sonnet-5 canary entries this week — the harper canary produced no triaged verdicts, so there is no sonnet-5-vs-sonnet-4-6 read.

Per prompt ref (short SHA):

Prompt ref	useful	noise	partial	total
9cf49d23cb04	98	1	1	100
d1d9a655e969 (gemini)	0	1	0	1
accea2706fae (gemini)	0	1	0	1

9cf49d23cb04 has been the live claude ref across the whole window (no before/after prompt-change read to draw). The two other refs are what the gemini reviewer ran against on this repo's own prompt-layer PRs (noise).

Per repo: harper 69 (68 useful + 1 noise) · harper-pro 23 (all useful) · oauth 8 (7 useful + 1 partial) · ai-review-prompts 2 (both noise, gemini).

Recurring NOISE patterns (none to add)

All three noise calls are already-covered shapes; none is the review over-flagging — each is a vacuous "no blockers" on a diff with no reviewable surface:

1. Lint-only mechanical diff — #847 (harper#1612): a +1/−1 node:assert/strict → node:assert import swap. Covered by the existing "Mechanical, no-logic diffs" bullet + <!-- review:no-log --> marker.
2. Provider runs on prompt-prose PRs — #842 (ai-review-prompts#73) and #820 (ai-review-prompts#72): gemini reviews of markdown-only prompt-layer PRs; vacuous by construction, same convention already logged as #670/#747. Bot-only (gemini).

Recurring PARTIAL / FALSE-NEGATIVE patterns (none new; reinforces open PR #73)

One partial:

1. Severity deflation / under-calling on a clean review — #903 (oauth#167): Claude returned "no blockers" while gemini's leg on the same PR (#902) caught a real finding the author then fixed (commit 2c1b2a9; gemini re-review confirmed "Prior 1 finding(s) resolved"). Human-corroborated (triaged by Nathan), author-fixed, and peer-corroborated — a strong single point, same class as the #699/#700 precedent.

Signal weighting: highest-weight point of the week, but a single point whose specific technical sub-class isn't recorded in the triage note (the missed finding lives in the gemini #902 leg). It is the same "under-called clean review / severity deflation" class that the still-open #73 (week of 2026-06-29, unmerged) already addresses by adding severity-deflation guidance to universal.md. Per the sweep rule on open prior calibration PRs, this week does not add competing/duplicate text — #903 corroborates #73's direction and should be weighed when reviewing it.

Positive corroboration: on the oauth signing-key GC bug, Claude (#852) and gemini (#851) independently flagged the same blocker — the #903 under-call is not a systemic collapse of Claude's oauth calibration.

Changelist

log-only — no prompt changes this week. The noise calls are all already-covered non-event shapes (nothing over-flagged), and the single false-negative (#903) is one data point in a class already handled by the open, unmerged PR #73 — below the "≥2 independent, ideally human-corroborated points, and don't duplicate an open PR" bar.

• CALIBRATION.md — prepended the ## Week of 2026-07-06 section (the synthesis above).

🤖 Generated with Claude Code