Where the verify-* critics emit false PASSes — and why we can't yet prove it
The question
Across the seven verify-* skills (verify-vault-write, verify-strategic-output, verify-dispatch, verify-pdf-output, verify-action, /design-critic, /video-critic), where has the founder disagreed with a PASS verdict in the last 60 days (2026-05-04 → 2026-07-03), and what are the recurring failure-mode patterns the critic is blind to? Context: the v3 SpaceX memo was a verify-pdf-output gap that triggered a skill build; the /improve cron is scoped to monthly sample-and-regrade but the historical failure inventory has never been compiled.
What we already know (from the vault)
- Zero explicitly logged founder-vs-PASS disagreements exist in the window — and that absence is an instrumentation gap, not a clean bill of health. A full sweep of the self-review log,
~/.claude/state/, every verify-*SKILL.mdchangelog, and vault-wide override-language greps returned no event of the shape "critic emitted PASS → founder caught a miss." The machinery that would record one was specced but never stood up:~/.claude/state/improve-cron-log.mdand~/.claude/state/flash-review-log.jsonldo not exist, the Tier-2 flash review is marked "NOT BUILT" ([[2026-05-05-verify-action-v2-flash-review-extension]]), and a ~48h cron-auth 401 (2026-06-20→06-22) killed the one self-review cycle that might have caught drift. The/improvecron itself ([[2026-05-23-improve-cron-design-spec]]) is still design-only. - Two in-window critic GAPS exist, but both are "defect got through" events that predate or created the critic, not PASS-overrides. (1) MAC video C v2 (2026-05-07): the parent's own render-check read frame extracts "too generously" and missed a clipping badge at t=12s and a "FAIL" line-wrap at t=26s — this created /video-critic. (2) SpaceX v3 investment memo (2026-05-23): founder flagged a white margin band, an orphan callout straddling a page break, and ~9pt body — verify-pdf-output was built the same day, its checks 1/2/3 literally labeled "(Fixes v3 defect)" ([[2026-05-05-pdf-quality-bar-synthesis]] / SpaceX build). The design-critic rule-additions (marquee bar, palette-temperature collapse, 60/30/10 budget, bento integrity) are the same "founder caught it → add a rule" pattern but land in April, out of window.
- Every verify- skill scores a fixed rubric of presence/format axes, and the "founder override beats the rubric" carve-out is written into four SOPs — yet has never logged a firing.* verify-pdf-output = 12 binary layout checks; verify-vault-write = RDCO-mapping/cross-link presence; verify-strategic-output = confidence-vs-stated-evidence calibration; design-critic = visual tells. The carve-outs ([[2026-05-19-verification-as-independent-worker-pattern]] "founder override beats the loop"; website-polish-workflow "if design-critic PASSes something he hates, his override is the higher signal"; verify-strategic-output/verify-dispatch "the verifier is B-tier") institutionally concede the critics are a floor — but with no log, RDCO cannot measure its own catch rate.
- The vault already catalogued the six failure modes this audit is hunting. [[2026-05-19-cai-critic-graduation-per-axis-threshold]] and [[2026-05-22-reward-hacking-patterns-llm-critic-systems]] name: (1) blind-spot exploitation — silent, PASS-rate climbs while quality plateaus; (2) sycophancy/style-mimicry; (3) constitutional loophole (rule-letter vs rule-intent, e.g. "cite a source" satisfied by a hallucinated citation); (4) drift; (5) mode collapse; (6) OOD brittleness. It also flags that RDCO runs the same Claude model as both producer and critic — so the rubric, not any capability differential, does all the work.
What the web says
- "Evaluation Illusion": frontier judges anchor on surface heuristics — formatting, fluency, confident tone, structural polish — rather than substance (Beyond the Illusion of Consensus, arXiv 2603.11027). This is the single most load-bearing external finding for RDCO.
- Automated judges catch only ~1 in 5 real failures in production multi-turn agents; "automated judging is a regression floor, not a substitute for human review" (Catching One in Five, arXiv 2606.10315).
- Verbosity/length bias inflates preference for longer, fluent answers by 15-30 points across GPT-4/Claude/PaLM judges (Justice or Prejudice?, arXiv 2410.02736); self-preference/self-enhancement bias makes a judge favor outputs from its own model family (Self-Preference Bias, arXiv 2410.21819).
- Rubric-based judges fail specifically via "Hackable/Proxy-Based Scoring" and "Ungrounded Verification" — presence checks satisfiable without the underlying substance (RIFT taxonomy, arXiv 2604.01375; Reward Hacking in Rubric-Based RL, arXiv 2605.12474). Manipulated reasoning traces alone inflated false-positive rates up to 90% (Gaming the Judge, arXiv 2601.14691).
- A diverse panel/jury of different model families reduces correlated blind spots — the standard mitigation when a single judge shares surface heuristics with the producer (futureagi bias-mitigation 2026).
Convergences and contradictions
- The literature's #1 predicted blind spot is exactly RDCO's structural exposure. "Evaluation Illusion" (surface over substance) maps 1:1 onto verify-pdf-output's 12 purely-mechanical layout checks, verify-vault-write's presence checks, and design-critic's visual tells — a PDF or vault note can pass every check and still be substantively hollow or wrong. Both in-window gaps (video legibility, PDF layout) are surface-legibility misses, and both were closed by adding a specific rule, not a substance axis — so the class of miss remains open.
- RDCO's own carve-outs already say what "Catching One in Five" says. Four SOPs treat the critic as a regression floor and the founder as the higher signal. The literature validates that posture — but RDCO's version is unfalsifiable today because the override that would test it is never recorded.
- Contradiction / honesty flag: RDCO has no measured disagreement data, so it cannot yet confirm which literature-predicted blind spot dominates its critics. Everything below is inferred from rubric structure plus the two in-window gaps — not from a disagreement corpus. The self-preference-bias risk is real in principle (same model both seats) but unquantified.
Synthesis for RDCO
Be explicit: the honest count of founder-vs-PASS disagreements in the last 60 days is zero, and that is a finding, not a reassurance. The disagreement inventory this brief was meant to compile cannot be compiled because the verify-* stack never logged its verdicts to a durable per-axis file. "Zero overrides" and "critics are well-calibrated" are indistinguishable given the current instrumentation. So the first thing /improve must test is not drift — it is whether a disagreement corpus can exist at all. Blind spot #0 (meta): the stack has no verdict log; stand up improve-cron-log.md / a per-axis flash-review-log.jsonl before any regrade math is meaningful.
Ranking the substantive blind-spot patterns the critics are structurally exposed to, most-to-least load-bearing: (1) Surface-over-substance ("Evaluation Illusion"). verify-pdf-output scores 12 layout invariants and nothing about whether the memo's argument holds; verify-vault-write scores whether an RDCO-mapping/cross-link is present, not whether it is insightful or points at the right thing; design-critic scores tells, not taste. This is where the SpaceX and video gaps both live, and where the literature says single-model judges fail hardest. (2) Rule-letter vs rule-intent (constitutional loophole). Presence axes — "cite a source," "has cross-links," "has an RDCO mapping," "has a bear case" — are satisfiable by hollow compliance (boilerplate mapping, links to the wrong doc, a token bear paragraph). (3) Self-preference / no capability differential. Producer and critic are the same Claude; the literature says this inflates false-PASS, and RDCO's only structural counter is the fresh-eyes zero-context dispatch — which defeats confirmation bias but not shared surface heuristics.
What the /improve cron should test (concrete, falsifiable hypotheses): (H1) Run the RGFMD adversarial-self-generation step already specced in [[2026-05-23-improve-cron-design-spec]] against verify-pdf-output and verify-vault-write first — manufacture artifacts that pass every mechanical/presence check but are substantively wrong or hollow, and measure how many the critic still PASSes. A high pass rate confirms the surface-over-substance blind spot empirically. (H2) Treat the two in-window gaps (MAC video C v2, SpaceX v3) as the seed negative-example corpus and back-fill them as gold-set FAILs so the regrade has something to score against. (H3) On the single highest-stakes axis (verify-strategic-output, where a false PASS ships a wrong thesis), pilot a cross-model or differently-prompted critic and measure the PASS-rate delta vs the same-model critic — the cheapest test of the self-preference hypothesis. (H4) Add one semantic axis on top of the presence axes in verify-vault-write ("does the RDCO mapping say something non-obvious?") and see whether it diverges from the presence check on real traffic. The design spec's own metric note applies: at RDCO's small N, founder-override events are the primary drift signal and Cohen's kappa is only confirmation — which is precisely why the missing log is the binding constraint.
Open follow-ups
- Build the per-axis verdict log (
~/.claude/state/improve-cron-log.mdand/orflash-review-log.jsonl) — the hard prerequisite for ever compiling a disagreement inventory. Nothing else in the /improve cron is meaningful until this exists. - Retro-log the two in-window gaps (video C v2 2026-05-07, SpaceX v3 2026-05-23) as the seed FAIL corpus so month-1 regrade has ground truth.
- Run RGFMD on verify-pdf-output + verify-vault-write and record how many mechanically-clean-but-hollow artifacts earn a PASS.
- Decide whether cross-model critique is worth the cost on verify-strategic-output (the highest failure-cost axis).
- Reconcile the design spec's ">15% raw disagreement" reversion threshold with Cohen's kappa at N<50/axis — raw % is unreliable on PASS-skewed distributions.
- verify-action is dormant/NOT-WIRED (founder decision 2026-06-09) — confirm it should stay out of the /improve regrade scope rather than being counted as a silent-zero.
Related
- [[2026-05-23-improve-cron-design-spec]]
- [[2026-05-19-cai-critic-graduation-per-axis-threshold]]
- [[2026-05-22-reward-hacking-patterns-llm-critic-systems]]
- [[2026-05-20-verify-stack-two-gate-pass-fail-architecture]]
- [[2026-05-19-verification-as-independent-worker-pattern]]
- [[2026-06-14-open-ended-output-verifier]]
- [[2026-05-05-verify-action-v2-flash-review-extension]]
Sources
Vault:
- ~/rdco-vault/06-reference/research/2026-05-23-improve-cron-design-spec.md
- ~/rdco-vault/06-reference/research/2026-05-19-cai-critic-graduation-per-axis-threshold.md
- ~/rdco-vault/06-reference/research/2026-05-22-reward-hacking-patterns-llm-critic-systems.md
- ~/rdco-vault/02-sops/2026-05-20-verify-stack-two-gate-pass-fail-architecture.md
- ~/rdco-vault/02-sops/2026-05-19-verification-as-independent-worker-pattern.md
- ~/rdco-vault/06-reference/research/2026-06-14-open-ended-output-verifier.md
- ~/rdco-vault/08-tooling/2026-05-05-verify-action-v2-flash-review-extension.md
- ~/.claude/state/pdf-quality-bar-synthesis.md (SpaceX v3 defect table)
- ~/.claude/skills/{verify-vault-write,verify-strategic-output,verify-dispatch,verify-pdf-output,verify-action,design-critic,video-critic}/SKILL.md (rubric axes + changelogs)
Web:
- https://arxiv.org/pdf/2603.11027 — Beyond the Illusion of Consensus (surface-heuristic anchoring)
- https://arxiv.org/html/2606.10315 — Catching One in Five: LLM-as-Judge Blind Spots in Production
- https://arxiv.org/pdf/2410.02736 — Justice or Prejudice? (verbosity/position bias)
- https://arxiv.org/pdf/2410.21819 — Self-Preference Bias in LLM-as-a-Judge
- https://arxiv.org/pdf/2604.01375 — RIFT: Rubric Failure Mode Taxonomy (hackable/proxy scoring)
- https://arxiv.org/pdf/2605.12474 — Reward Hacking in Rubric-Based RL
- https://arxiv.org/html/2601.14691 — Gaming the Judge (unfaithful CoT)
- https://futureagi.com/blog/evaluating-llm-judge-bias-mitigation-2026/ — jury/panel mitigation for correlated blind spots