Changelog

2026-08-23: Settings get a page. /dashboard/settings lists every connected repository with its flag line, its PR comment toggle and a new deep read toggle, and the site header links to the dashboard. Deep read off means Doug scores that repository on structural signals alone — no diff leaves it. On a repository with no flag line of its own that also moves the line Doug bands against, from the deep-read default to the fallback one, so it asks for a human less often rather than merely differently.
2026-08-23: The reader reads harder and spends less doing it. Its reasoning effort goes to high, and the two mechanical passes behind it — the one that checks a finding against the file it cites, and the one that places a finding on a hunk — move to a cheaper model than the frozen reader (ADR-0016, ADR-0018). Grounding switches on for one named installation rather than for everyone at once (ADR-0017).
2026-08-21: Doug stops forgetting what he already said. When the code a finding cited is byte-unchanged in the next push, he carries that finding forward by construction and reports it under ### Since <sha>, with a count of how many of his own earlier findings on untouched code he did not mention again. Nothing is ever marked resolved: evidence that code was edited is not evidence it was fixed, and a sampled hand-check said so before this shipped.
2026-08-20: The sticky PR comment finishes rolling out. It is on for every repository Doug reviews, and the per-repository toggle beside the flag line is the only thing that turns it off.
2026-08-19: Doug leaves one sticky comment on each reviewed PR that repeats its check run word for word, edited in place on every push — on by default, opt-out per repository beside the flag line.
2026-08-13: The outcome instrument is visible. Every check run ends with adjudicated N · pending M · as of <date> and a deep-read meter; the public /scoreboard page shows the same counters from the same query, empty and dated. Bot-authored PRs no longer buy a deep read. Hosted /docs actually serves — Cloud Build had been stripping it. Finding file is carried on the Reason rather than rematched by description.
2026-08-12: Dashboard rebuilt on the console’s design grammar: a bounded table, a threshold view lens, a larger type scale, a space picker that navigates on selection. Floating header with separated sign-in. Hosted docs section.
2026-08: Doug’s own findings now get a durable disposition — real / disproved / adjacent, plus whether anything changed and which file settled it (docs/findings-log.jsonl). Every row at the time was backfill and excluded from every rate; the prospective denominator opened later. A design for an agent review crew with a falsifier-adjudicated ledger was put through six independent reviewers and cut: four of six proposed “agent lenses” turned out to be checks that already run or free platform features, and the ledger’s own motivating case — a finding that draws a conclusion from an absence in the diff — is one its mechanism could not have fixed, because the obvious falsifier restates the error. What survived is the log and one reviewing rule.
2026-07: Phase-1 entry experiments, pre-registered end to end. RandomForest on Kamei’s 14 confirms the metadata ceiling (a plain size sort still wins). An LLM diff-read probe out-ranks every deterministic baseline on both repos — AUC 0.69 sentry / 0.67 grafana, the first method here that held up on repo #2 — and survives a polarity-inversion counterfactual (findings track change semantics, not diff surface). Finding-pattern distillation: marginal pass on sentry, coverage fail on grafana. Scripts: api/scripts/rf_kamei.py, api/scripts/llm_probe.py.
2026-07: Backtest hardening: checkpoint/resume harvests, cache seeding, listing dedupe, --backfill-details (per-file stats + patch text). Cleared-band metric shipped. Rolling-window hotspot learning built; verdict pending deeper history. This docs site.
2026-07: Renamed to Doug (after a Saint Bernard). Landing page live. Replays published for sentry, grafana, ruff.

Changelog

2026-08-23: Settings get a page. /dashboard/settings lists every connected repository with its flag line, its PR comment toggle and a new deep read toggle, and the site header links to the dashboard. Deep read off means Doug scores that repository on structural signals alone — no diff leaves it. On a repository with no flag line of its own that also moves the line Doug bands against, from the deep-read default to the fallback one, so it asks for a human less often rather than merely differently.
2026-08-23: The reader reads harder and spends less doing it. Its reasoning effort goes to high, and the two mechanical passes behind it — the one that checks a finding against the file it cites, and the one that places a finding on a hunk — move to a cheaper model than the frozen reader (ADR-0016, ADR-0018). Grounding switches on for one named installation rather than for everyone at once (ADR-0017).
2026-08-21: Doug stops forgetting what he already said. When the code a finding cited is byte-unchanged in the next push, he carries that finding forward by construction and reports it under ### Since <sha>, with a count of how many of his own earlier findings on untouched code he did not mention again. Nothing is ever marked resolved: evidence that code was edited is not evidence it was fixed, and a sampled hand-check said so before this shipped.
2026-08-20: The sticky PR comment finishes rolling out. It is on for every repository Doug reviews, and the per-repository toggle beside the flag line is the only thing that turns it off.
2026-08-19: Doug leaves one sticky comment on each reviewed PR that repeats its check run word for word, edited in place on every push — on by default, opt-out per repository beside the flag line.
2026-08-13: The outcome instrument is visible. Every check run ends with adjudicated N · pending M · as of <date> and a deep-read meter; the public /scoreboard page shows the same counters from the same query, empty and dated. Bot-authored PRs no longer buy a deep read. Hosted /docs actually serves — Cloud Build had been stripping it. Finding file is carried on the Reason rather than rematched by description.
2026-08-12: Dashboard rebuilt on the console’s design grammar: a bounded table, a threshold view lens, a larger type scale, a space picker that navigates on selection. Floating header with separated sign-in. Hosted docs section.
2026-08: Doug’s own findings now get a durable disposition — real / disproved / adjacent, plus whether anything changed and which file settled it (docs/findings-log.jsonl). Every row at the time was backfill and excluded from every rate; the prospective denominator opened later. A design for an agent review crew with a falsifier-adjudicated ledger was put through six independent reviewers and cut: four of six proposed “agent lenses” turned out to be checks that already run or free platform features, and the ledger’s own motivating case — a finding that draws a conclusion from an absence in the diff — is one its mechanism could not have fixed, because the obvious falsifier restates the error. What survived is the log and one reviewing rule.
2026-07: Phase-1 entry experiments, pre-registered end to end. RandomForest on Kamei’s 14 confirms the metadata ceiling (a plain size sort still wins). An LLM diff-read probe out-ranks every deterministic baseline on both repos — AUC 0.69 sentry / 0.67 grafana, the first method here that held up on repo #2 — and survives a polarity-inversion counterfactual (findings track change semantics, not diff surface). Finding-pattern distillation: marginal pass on sentry, coverage fail on grafana. Scripts: api/scripts/rf_kamei.py, api/scripts/llm_probe.py.
2026-07: Backtest hardening: checkpoint/resume harvests, cache seeding, listing dedupe, --backfill-details (per-file stats + patch text). Cleared-band metric shipped. Rolling-window hotspot learning built; verdict pending deeper history. This docs site.
2026-07: Renamed to Doug (after a Saint Bernard). Landing page live. Replays published for sentry, grafana, ruff.