Plan review — Reviewer X

Date: 2026-05-12 Plan version reviewed: v2 (2026-05-12)

Summary

The plan ports math-olympiad's adversarial-verification pattern to bridge-card sign-off in a fundamentally reasonable way, and on its central choices — three role-typed reviewers, dual context isolation, pattern-armed adversarial taxonomy, asymmetric quorum, hypothesis-note-before-report discipline, raw-access budgeting, pilot-before-codify — it is more rigorous than what the v0d.5 schema currently requires (a date on a line). However, the plan has several load-bearing translation defects from math-olympiad, a systematically miscalibrated quorum threshold, an inadequate adversarial taxonomy for the specific failure modes of bridge cards (notably missing the wiki's own most-cited prior incident class, false-friend confusion, and motifs-double-promotion), and a false-isolation hazard at the adjudication seam where main-thread Claude collapses 18 reports.

The most consequential single finding: the asymmetric quorum is inverted from math-olympiad's logic in a way that makes the procedure less sensitive to DEFECT signal than math-olympiad's 4-of-5/2-of-5, not more. The plan's "2-of-3 DEFECT = block" rule is roughly equivalent to math-olympiad's 4-of-5 in confirmation strength, when in fact the math-olympiad pattern's whole point is asymmetry — and a faithful 3-reviewer translation should be 1-of-3 DEFECT triggers adjudication and 2-of-3 DEFECT = block, with 3-of-3 APPROVED the only path to silent sign-off. Combined with the carve-out's apply-mode targets, the current rule risks rubber-stamping bridges that get a single quiet DEFECT FOUND vote with a real anchor problem. Verdict: ACCEPT-WITH-CHANGES.

Defects found

D1. Quorum threshold is mistranslated from math-olympiad

  • Location: §9 (Quorum rule), and the rationale paragraph "Asymmetric because: a single DEFECT FOUND is a strong signal cheap to investigate; rubber-stamping is expensive to undo once apply-mode writes have landed."

  • Defect: The plan's table reads:

    • 3/3 APPROVED → sign off
    • 2 APPROVED + 1 DEFECT → adjudicate
    • 1 APPROVED + 2 DEFECT → block
    • 3 DEFECT → retire

    Math-olympiad uses 4 of 5 to confirm, 2 of 5 to refute with 5 verifiers, i.e. ratios of 4/5 = 0.80 to confirm, 2/5 = 0.40 to refute. The asymmetry is that confirmation needs supermajority but refutation needs only minority-plus.

    Translated to 3 reviewers, faithful asymmetric thresholds would be: 3/3 to silently confirm (1.0), 2/3 to refute (0.67) — and the gap between them is filled by adjudication. The plan does keep 3/3 for silent confirm and 2/3 for block, so far so good. The defect is that the plan's rationale prose says "a single DEFECT FOUND is a strong signal" yet the table treats 1-DEFECT as "adjudicate" (which, in practice, will usually resolve to APPROVE — main-thread Claude reads three reports, two say APPROVED, one says DEFECT, and the structural pressure is to side with the majority). The procedure tells the maintainer "be sensitive to single dissents" while structurally arranging conditions in which a single dissent is the weakest possible outcome.

  • Why it matters: Bridge cards are pre-claim staging artifacts whose downstream consequences include subsection writes on HUB concept pages (institution, transtemporality, good-ambiguity) that may later be cited by claims and motif entries. A 1-of-3 DEFECT FOUND on an editorial-frame-laundering pattern is exactly the failure class the plan claims to catch and the math-olympiad design philosophy targets. Routing it to "adjudicate" against two APPROVED votes structurally rubber-stamps it via social proof in the adjudicator.

  • Proposed fix: Replace the rationale with the actual asymmetric logic the table implements, and tighten the table: when adjudicating a 2-APPROVED + 1-DEFECT, main-thread must (a) read the DEFECT report before reading the APPROVED reports (sequence matters; first-read is anchoring evidence), (b) explicitly write the DEFECT's strongest sentence in the adjudication note before reasoning, and (c) require the adjudication to answer the DEFECT, not merely note its disposition. Better still: add a rule that 1-of-3 DEFECT FOUND with Confidence: high triggers block, not adjudicate. Confidence-high single dissent is closer in informational content to a 2-of-3 confidence-medium block.

  • Severity: major

D2. Main-thread adjudicator is a hidden coordination layer

  • Location: §2.2 ("Dual context isolation") and §11 ("Adjudication and audit trail")

  • Defect: The plan claims dual context isolation but the dual is between (a) reviewer ↔ Pass 3 motivating prose and (b) reviewer ↔ other reviewers. Both isolations terminate at the main-thread adjudicator, which sees all 18 reports plus the Pass 3 motivating prose plus accumulated session memory plus knowledge that this is a self-review of a procedure-it-just-wrote. The adjudicator is not isolated and there is no rule preventing the adjudicator from referencing reviewer agreement patterns ("two of three agreed it was fine") as primary signal.

    This is a load-bearing translation gap from math-olympiad. In math-olympiad the orchestrator orchestrates but does not adjudicate — the vote adjudicates. The orchestrator's job is mechanical (count verdicts, decide whether to launch more). Here, the main-thread is making interpretive calls.

  • Why it matters: Main-thread Claude wrote the plan that defines the procedure and the taxonomy. When it adjudicates, it is also implicitly grading its own taxonomy: if the taxonomy missed a pattern, main-thread Claude will tend not to recognize that miss because it doesn't have the pattern available. The "isolation" is undone at the seam where it matters most.

  • Proposed fix: Either (a) lower the adjudicator's discretion: explicit rule that adjudicator may only count verdicts and select from a small fixed action set (sign off, block, hold for human, narrow apply-as scope by removing items mentioned in DEFECT reports), with no free-form "B1's concern is already in card §Counterpressure" reasoning; or (b) add a fresh-context adjudicator subagent that gets the 3 reports per card (and only those) with the verdict format and quorum table, with no access to the plan, the Pass 3 motivating prose, or session memory. The current §11 example output even illustrates the failure mode: "adjudication: B1's editorial-frame concern already in card §Counterpressure; bridge survives at subsection level; sign-off proceeds" is exactly the kind of warming-over the §2.2 isolation rules were meant to prevent.

  • Severity: major

D3. "You are the only reviewer" prompt is a procedural assertion that doesn't enforce itself

  • Location: §2.2: "the subagent prompt explicitly says 'you are the only reviewer; no other process is checking this card'"

  • Defect: Telling a subagent it is the only reviewer doesn't make it so. The plan's actual isolation rests on (a) not putting other reviewers' verdicts in the prompt, (b) not putting peer-directory paths in the prompt, and (c) using fresh subagent contexts. The "you are the only reviewer" sentence is either decorative or an active liability: decorative if it merely restates what the context already enforces; an active liability if the subagent's harness has any way to discover (via Glob, Read, tool list inspection) that peer agents exist — at which point the prompt's contradiction-with-reality reveals the framing as theatre, which a sophisticated model may treat as social-proof signal in its own right ("they wouldn't tell me this unless it mattered, which means I am being graded as gatekeeper, which means I should bias toward approval to not look paranoid"). Math-olympiad's actual lever is "fresh context with no peer information," not the verbal claim of singularity.

  • Why it matters: The plan elevates a verbal claim to a §2.2 architectural commitment. The architectural commitment is the isolation; the verbal claim is rhetoric. Conflating the two weakens reviewer training of future maintainers about what actually causes isolation.

  • Proposed fix: Replace the "you are the only reviewer" framing in subagent prompts with the operational truth: "You are reviewing this bridge card independently. You have no information about whether or how this card has been reviewed elsewhere. Do not speculate about other reviews; reason from the card and the cited anchors." The instruction is honest and the bias-removal logic is the same.

  • Severity: minor

D4. Thinking-trace stripping is absent — math-olympiad's #1 lever is not translated

  • Location: Absent from the plan. Math-olympiad §3 ("Clean the solution — context isolation — the #1 lever").
  • Defect: Math-olympiad's §3 strips the solver's reasoning trace before the verifier sees the proof, because "the thinking trace biases the verifier toward agreement — a long chain of reasoning reads as supporting evidence even when the conclusion is wrong." The bridge-card analogue is the Pass 3 motivating prose plus the silent-keys Phase 2 PASS rationale. The plan does say (§2.2) that subagents don't get the "framing prose of the Pass 3 report" or the "silent-keys Phase 2 verdict prose" — good. But the bridge card itself contains framing prose: the "What it makes visible" field, the "Note on Primary home selection" paragraphs (cards 3 and 4), and the relation glosses are all reasoning-trace artifacts that bias toward agreement in the same way math-olympiad's stripped thinking does.
  • Why it matters: A reviewer who reads the card's full prose, including the author's argument for why this bridge is interesting, is in the same epistemic position as math-olympiad's biased verifier. The "What it makes visible" field on card 1, for example, opens with "institution's mechanism is neither pure event ... nor pure residue — it is the verb-form déposer" — that is a polished claim, not a neutral fact. A verifier reading it is reading the conclusion before checking the premises.
  • Proposed fix: Add a §3.5 "Strip the bridge prose" pass as preparation step. For Role B (adversarial) and Role C (evidence-tracer), present the card in two passes: pass 1 gets a stripped form (Status, Bridge type, Primary home, Secondary homes, Relation bullets reduced to page-A → page-B with type label, raw evidence anchors only, Apply-as items only). Pass 2 unlocks the full prose for the reviewer to compare their initial reading against the card's framing. Role A may use the unstripped card because Role A's job is the structural gate (the prose is part of what they're checking). The stripping should be mechanical, scripted if possible, not the agent's own self-discipline.
  • Severity: major

D5. Calibrated abstention is missing from reviewer verdicts

  • Location: §8 (Verdict format), §9 (Quorum)

  • Defect: Math-olympiad's §7 has explicit "no confident solution" as a verdict and structures the workflow to prefer it over wrong-confident. The plan has APPROVED / DEFECT FOUND / UNCLEAR, with the Confidence field tied to the verdict. But UNCLEAR is rolled into the adjudicate bucket without distinguishing between (a) UNCLEAR-because-cannot-determine-from-this-evidence (the card is under-specified) versus (b) UNCLEAR-because-this-card-is-out-of-scope-for-this-role (e.g., Role C sees the empirical anchors check out but is reading a card whose pattern problem is editorial-frame laundering, which is Role B's job).

    These should produce different downstream actions. Math-olympiad's calibrated-abstention idea is more than a label; it's a structured way of saying "I am not the right instrument for this question." The plan loses that.

    Worse, the plan's adjudication procedure does not have an action for "all 3 reviewers say UNCLEAR." The §9 table covers "Any UNCLEAR → adjudicate" but doesn't specify what adjudicate means when nothing has been confirmed or refuted.

  • Why it matters: A bridge card that produces three UNCLEAR verdicts is likely under-specified (the kind of card that should be sent back to the Pass 3 author for rewrite). Routing it to "adjudicate" instead of "rewrite" lets badly-formed cards get a maintainer signature.

  • Proposed fix: Add a fourth verdict — OUT-OF-ROLE-SCOPE (or NEEDS-DIFFERENT-INSTRUMENT) — and split UNCLEAR into (a) under-specified-card, (b) cannot-determine-from-cited-evidence. Add to §9: 3/3 UNCLEAR-under-specified → card returns to proposed with rewrite requirement; 2-of-3 OUT-OF-ROLE-SCOPE → role assignment was wrong, redo this card with re-roled reviewers.

  • Severity: major

D6. Role-A's "Adversarial restatement (light)" muddies the gate-vs-attack boundary

  • Location: §3 test #6 ("Adversarial restatement (light)")
  • Defect: §3 explicitly says "Role A's full adversarial work is bounded by the schema's structural requirements. Deep attacks are Role B's job." Then test #6 puts an adversarial restatement inside Role A's gate. This produces a non-trivial probability that Role A does Role B's job badly (because Role A is gate-applying-stance, not adversarial-attacker-stance), and Role B is then anchored on Role A's restatement (because — see D2 — the adjudicator and possibly the card content reflects what Role A produced).
  • Why it matters: The whole point of three roles is stance separation. Putting adversarial work in two roles dilutes both. If "the strongest objection is stated in the card" is meant as a check that the card itself contains a Counterpressure entry, say exactly that and remove the "adversarial" framing. If it's meant as Role A doing its own adversarial pass, restructure into A's role and remove the "Deep attacks are Role B's job" carve-out.
  • Proposed fix: Rewrite test #6 as: "The card contains a Counterpressure field with at least one stated objection; the objection is articulated specifically enough (i.e., it names a check that could be performed) rather than gestured at." Move actual adversarial restatement entirely to Role B.
  • Severity: minor

D7. Carve-out boundaries bleed at "typed-connection additions"

  • Location: §12 (Carve-out)
  • Defect: Sign-off authorizes "Typed-connection additions" but does NOT authorize "Promotion to motif HUB/STRUCTURAL weight." These are entangled. A typed connection that adds is the condition of intelligibility of between [[institution]] and [[transtemporality]] (for example) materially affects motif weight even if it doesn't formally write a motif HUB upgrade, because motifs.md weight is computed in part from the topology of typed connections (this is the entire premise of weave Pass 1's connection-graph survey and Pass 4's bidirectional back-references). The plan treats motif weight as a separate corpus-level judgment but doesn't say what happens when a sign-off-authorized typed-connection addition would tip motif weight by counterfactual aggregation.
  • Why it matters: This is exactly the #2 single-source promotion creep pattern that the plan claims to catch, but applied at the motifs level rather than the claims level. The plan's pattern #2 only checks the immediate apply-as items for claim/motif HUB writes; it does not check whether typed-connection additions taken together would change motif weight.
  • Proposed fix: Add to §12 a sub-rule: "Sign-off-authorized typed-connection additions that change motif weight (by adding the cards' Primary or Secondary homes to a motif's attestation graph in a way that crosses the THEME→STRUCTURAL or STRUCTURAL→HUB threshold) trigger a motifs-delta sweep before apply-mode writes proceed. Motifs-delta runs are themselves not authorized by bridge-card sign-off — they require their own audit phase." Add a Role B pattern check (pattern #12, see Missing Patterns below) for this case.
  • Severity: major

D8. Pilot success criteria are unspecified — the §13 promise is empty

  • Location: §13 (Pilot-before-codify)
  • Defect: §13 says "if pilot fails ... revise; if pilot succeeds, codify" with no operational criteria for either. What does success look like? The plan's only criterion is the negative one: "same-gap-twice fires" triggers a pause. There is no positive criterion. The taxonomy could be too lax — produce 5 silent sign-offs with no defect signal — and the plan would call that "success" and codify the procedure with its blind spots intact.
  • Why it matters: A pilot designed only to catch its own failure-by-detection-fire cannot validate the taxonomy's coverage. If the taxonomy misses a pattern, the pilot won't surface it.
  • Proposed fix: Specify pilot success as a conjunction:
    1. At least one DEFECT FOUND across the 5 cards. A 5-for-5 silent sign-off should be treated as evidence that the taxonomy lacks bite, not as evidence that the cards are clean. (The cards include a single-source promotion blocker on card 5, a 4-chapter-cross-author convergence on card 1 with explicit editorial-frame counter-pressure, and a 1947→1958→post-1958 archival reading on card 4. If reviewers find zero defects across these, the procedure is broken, not the cards.)
    2. The DEFECT FOUND should fire a specific pattern from the taxonomy (not "other"). If reviewers keep firing "other," the taxonomy is under-specified.
    3. At least one reviewer triggers the §6 raw-budget self-flag (over-budget → "needs fresh ingest pass, not sign-off review") OR explicitly notes the budget was adequate. Silence about the budget means the budget wasn't actually checked.
    4. The §10 same-gap-twice check finds zero matches. (Negative criterion, already present, but should be explicit.)
    5. No 1-of-3 DEFECT-FOUND adjudication produced a silent APPROVE. Every adjudication is reviewed against D1.
  • Severity: major

D9. No procedure for cards that reference each other

  • Location: §14 ("What about cards that reference each other?")
  • Defect: §14 dismisses cross-card dependencies with "Each reviewer evaluates each card independently; cross-card dependencies are flagged as concerns but do not affect the per-card verdict." Card 5's evidence chain explicitly leans on Card 1's déposer mechanism. If Card 1 has a defect that retires it (3/3 DEFECT in the future), Card 5's apply-as items 2 and 3 (which reference card 1's déposer) are now ungrounded — but Card 5 may have already been signed off.
  • Why it matters: The plan splits the 5 cards into batches across role-A's 2 agents (a1: 1-3; a2: 4-5). Reviewer A2 looking at card 5 doesn't know whether card 1 was approved or retired (and per dual-isolation, should not know in-session). After adjudication, the maintainer ends up with a possibly-inconsistent outcome: card 1 retired, card 5 approved with apply-as items that cite card 1.
  • Proposed fix: Add a §11.5 "Cross-card consistency check" run after adjudication on all cards but before writing any Approved by maintainer: line. Main-thread Claude (or, better, a fresh subagent) checks: for each approved card, does any Apply-as item reference a retired or blocked card? If yes, conditionally narrow that approved card's apply-as scope or downgrade to held-pending-other-card-revision.
  • Severity: major

D10. Pattern taxonomy gives short shrift to the wiki's documented prior failure modes

  • Location: §4 (the 11-pattern taxonomy)

  • Defect: The wiki has two recorded prior failure modes that the plan does not weaponize as patterns:

    1. The Faul fabrication incident (CLAUDE.md General Rule 16) — fabricated secondary-source citations. Plan pattern #9 references it but the check is generic ("does every cited source / page / line exist?"). Math-olympiad-style this would be a specific attack: "Search the cited journal's archive; if the journal is real but the article isn't, you have the Faul pattern."
    2. The 2026-05-09 whole-file duplication of wiki/claims.md (CLAUDE.md "Slug uniqueness invariant"). Not bridge-card-specific, but the failure mode it represents — silent rendering bugs from artifact-level structural duplication — has an analogue here: a bridge card whose apply-as items duplicate existing typed connections that already exist on the target page, but with subtly different language that wins the rendering race.

    Plus, the plan ignores three bridge-card-specific failure modes the wiki's architecture would predict:

    1. False-friend laundering through bridge-card apply-mode. CLAUDE.md General Rule 17 specifies the "false-friend caution" mechanism. A bridge card whose apply-as items perform typed-connection additions and also mention the false-friend caution as a separate Open-Questions add risks laundering the false friend into the body via the typed connection while marking the caution as resolved. (This is structurally analogous to how user-memory-derived theses can be laundered through audit reports per CLAUDE.md Rule 17.)
    2. Stale-anchor pattern. Per CLAUDE.md General Rule 18 ("Retrospective relevance / artifact conservatism"), an older extraction note may be adequate for its source-local task and still miss a passage's later cross-source importance. A bridge card whose anchors are extraction-note lines from a year-old ingest may rest on extractions that did not yet have the bridge in view; the anchor "matches" the card's claim only because the card was constructed from the anchor's specific phrase rather than the anchor's local context.
    3. Citation-trace inversion. The 5 cards uniformly cite silent-keys-2026-05-07.md as the upstream artifact AND treat it as evidential. Per CLAUDE.md Rule 17 ("audit reports authorize only when traceably anchored"), this is acceptable only if the silent-keys report's PASS verdicts are themselves anchored in extraction notes / source pages / raw — which they are (verified in lines 33-34, 56-58 of silent-keys-2026-05-07.md). But Role C's protocol does not specify checking the silent-keys report's own anchor chain — it specifies checking the card's anchors, which include the silent-keys report. This is one link short.

  • Why it matters: Three of the 5 cards in run3 carry exactly the failure modes the taxonomy misses. Card 1's evidence depends on a secondary-author convergence reading (the M-C 2026 editorial frame) that is named in the Counterpressure but not weighed by any taxonomy pattern. Cards 1, 2, and 5 all reference M-C 2026 chapters in ways that could be false-friend at the cross-chapter integration level. Card 4 cites raw line numbers from SA-2006 (raw 1047, 1287, 1678) that are exactly the kind of citation Faul-style fabrication produces — but the line numbers correspond to a different artifact organization than the wiki's standard raw-citation scheme (the wiki uses page-based citations elsewhere for SA-2006), and the plan has no protocol for noticing scheme mismatch.

  • Proposed fix: Add patterns #12 ("motifs-weight tipping" — see D7), #13 ("stale-anchor"), #14 ("false-friend laundering through typed connection"), #15 ("citation-scheme mismatch"). See "Missing patterns" section below for the specific checks.

  • Severity: major

D11. Hypothesis-note discipline is not enforceable

  • Location: §7 (Hypothesis-note discipline)
  • Defect: The plan requires each reviewer to write a hypothesis note BEFORE writing the report, with a specified structure. There is no mechanism to enforce this: a subagent that writes its report first and back-dates the hypothesis note is indistinguishable from one that wrote the hypothesis note first. The plan says "the note IS the reasoning trace" but with same-context generation, the reasoning trace can be reconstructed after the verdict is decided.
  • Why it matters: This is the "thinking-out-loud" trick — math-olympiad doesn't use it for solvers (it uses thinking blocks plus a clean output split). The plan adopts a form of thinking-out-loud without the harness mechanism that makes it informative.
  • Proposed fix: Either (a) commit hypothesis notes to file before invoking report-writing — split into two tool-bash calls with a marker file; the second call cannot proceed without the marker; or (b) make the hypothesis note part of the report's tool-call output, with the subagent prompt structured "first write hypothesis-note section; then write verdict section; do not edit the hypothesis-note section after writing the verdict; if you find yourself wanting to revise the hypothesis-note section, instead append a 'Revision after verdict' subsection." Option (b) is cheaper; option (a) is more rigorous.
  • Severity: minor

D12. Same-batch index isn't isolated across role assignments

  • Location: §14 (Card-batch splits the same across roles? "Yes")

  • Defect: A1, B1, C1 all get cards 1-3; A2, B2, C2 all get cards 4-5. This is presented as a feature ("clean per-card report assembly"). But it concentrates correlated noise: if A1 has a stylistic blind spot (e.g., tends to under-weight editorial-frame patterns), B1 and C1 reviewing the same cards do not check A1's work — they do their own roles, but they're also seeing only the same 3-card slice. If B1 also has a similar blind spot (different stance, similar bias), the entire 1-3 batch may be reviewed by reviewers who share a contamination that 4-5's reviewers don't have.

    A better split: A1 gets 1-3, B1 gets 3-5, C1 gets 1-2 + 5 (or similar non-overlapping but interleaved). No card has all three reviewers from the same batch index.

  • Why it matters: This is more theoretical than concrete (subagents are independent and the noise is unlikely to be highly correlated), but the §14 reasoning ("clean per-card report assembly") is a coordination convenience, not a correctness argument. The correctness argument runs the other way.

  • Proposed fix: Interleave card assignments so that no card has reviewers all drawn from the same per-role batch position. For 5 cards × 3 roles × 2 agents/role: assign each card 3 reviewers using a derangement of the agent set. Concretely: card 1 = A1+B2+C1; card 2 = A1+B1+C2; card 3 = A1+B2+C1; card 4 = A2+B1+C2; card 5 = A2+B1+C1.

  • Severity: minor

  • Location: §3 (gate test #3 "Apply-as specificity") and §4 pattern #9 (fabricated citation)
  • Defect: Bridge cards in the run3 report reference [[hyperdialectique]] (cards 3 and 4 mention it as an alternative concept) — but this page does not exist in wiki/concepts/; only hyper-dialectic.md does. The silent-keys-2026-05-07.md upstream report has the same dead link. The plan's gate does not check Apply-as targets against the wiki's actual page inventory. A reviewer that mechanically applies the §3 gate would mark card 3's Apply-as item 2 ("Typed connection on [[hyper-dialectic]]") as compliant because hyper-dialectic exists — but card 3's framing prose says [[hyperdialectique]] may be an alternative anchor, and a sloppy apply-mode follow-up could route to the nonexistent page.
  • Why it matters: Lint catches dead wikilinks, but lint hasn't run on the proposed apply-as targets yet (they're not in the wiki). The plan should catch this before sign-off, not after apply-mode writes dead-link the wiki.
  • Proposed fix: Add to gate test #3 a sub-check: "Every wikilink in the Apply-as list resolves to an existing wiki page (per Glob over wiki/{sources,concepts,entities,questions}). Wikilinks in the Note-on-Primary-home-selection paragraphs are checked separately and flagged if they would seed apply-mode confusion."
  • Severity: minor

Math-olympiad transfer audit

Per math-olympiad insight, the plan's translation status:

  1. Strip thinking before verifying (math-olympiad §3) — partially translated. The plan strips the Pass 3 framing prose and silent-keys verdict prose (§2.2) but does NOT strip the bridge card's own framing prose ("What it makes visible," "Note on Primary home selection," relation glosses). See D4. The math-olympiad lever is incomplete.

  2. Pattern-specific attacks, not generic checks (math-olympiad §4) — faithfully translated. The 11-pattern taxonomy is the right shape: each pattern names a failure mode and a specific check. The patterns map plausibly to bridge-card analogues (#10 is the #4 RH analogue, #11 is the #40 too-clean-bridge analogue, etc.). But — see "Missing patterns" — the taxonomy is short relative to math-olympiad's 12-check verifier_patterns plus the adversarial-prompts library, and the patterns it does include sometimes overreach (e.g., #10 "would settle longstanding controversy" applied to bridge cards risks false positives because every philosophy bridge in some weak sense touches a longstanding controversy).

  3. Asymmetric vote (math-olympiad §5 "4-to-confirm, 2-to-refute") — partially translated. The plan has 3/3-to-confirm, 2/3-to-block, with adjudicate in between. The threshold direction is right but D1 identifies that the rationale paragraph doesn't match what the table operationally produces — the prose says "single DEFECT is a strong signal," the table routes a single DEFECT to adjudicate-which-likely-becomes-approve.

  4. Calibrated abstention (math-olympiad §7 "no confident solution") — missed. See D5. UNCLEAR exists as a verdict but is not structured to drive an abstain-rather-than-bluff workflow. Math-olympiad's no confident solution is the right answer when you can't confirm, and the plan doesn't have the analogue.

  5. Specification-gaming check first / interpretation check (math-olympiad §1) — missed. Math-olympiad has Step 1 "Interpretation check" before solving: identify alternative readings of the problem, choose the hard one, state which interpretation you're solving and WHY. The bridge-card analogue: before any role review, does the bridge have an alternative reading? Does its Primary home selection foreclose a different (and more defensible) Primary home? The plan touches this in §3 test #5 ("Primary home selection justified") but only as a structural check, not as an interpretation gate run before the rest. The 50/63 Aletheia finding (most "correct" answers solved the wrong reading) is directly relevant: a bridge card might survive all 11 patterns and 6 gate tests while pointing to the wrong relation between concepts.

  6. Presentation pass after correctness (math-olympiad §8) — not applicable. Bridge cards are pre-claim artifacts; presentation polish is an apply-mode concern, not a sign-off concern. Correctly omitted.

  7. Solver cannot verify its own solution (math-olympiad §5 "A solver cannot verify its own solution") — not applicable in the way the plan instantiates it, but analogous failure not addressed. Math-olympiad: the solver agent is not the verifier agent. Bridge-card analogue: the Pass 3 author (the agent that drafted the bridge cards) should not be the sign-off reviewer. This is implicit because the plan runs in a fresh review session, but is not stated. Worse: the main-thread adjudicator is currently the same agent that wrote the plan and is, in the v0d.7 codification step, the same agent that wrote the Pass 3 cards (run3 was generated 2026-05-08 by main-thread). Main-thread adjudicator-as-bridge-card-author is a structural conflict the plan does not address.

  8. Deep mode after abstention (math-olympiad §6c "Step 6d: not optional") — missed. Math-olympiad: after any ABSTAIN at the verify stage, automatically launch one deep-mode agent before writing the abstention. The bridge-card analogue: after a 1-of-3 or 2-of-3 DEFECT FOUND, before retiring the card, launch one deep-mode-equivalent agent (fresh context, bounded raw access, more time) to attempt repair. The plan does not have this; §11 just blocks and "flags for revision in a follow-up Pass 3 run."

  9. Retry policy (math-olympiad §2 "If an agent fails or times out, retry once") — missed. The plan has no retry policy. A subagent that fails partway should not silently drop a card.

  10. Pigeonhole early exit (math-olympiad §5) — not applicable at 3 reviewers. With 5 verifiers, early exit at 2-HOLE-FOUND saves cost; with 3 reviewers all running in parallel, there's nothing to exit early from. Correctly omitted.

  11. Model tier defaults (math-olympiad "Model tier defaults" table) — partial. The plan says "Opus 4.7, max effort" for all agents (§2.1). Math-olympiad has differentiated tier tables: weaker models → more parallel attempts, faster abstention; stronger models → deeper verification. The plan's choice of all-Opus is defensible for a personal wiki where cost is bounded, but it's a choice, not a translation of the math-olympiad principle. Worth a sentence acknowledging the tradeoff.

Missing patterns / tests

Additions the taxonomy should make:

  • #12 — Motifs-weight tipping. Check: would the bridge's typed-connection additions, combined with existing typed connections on the same Primary-home / Secondary-home pages, tip an existing motifs.md THEME entry to STRUCTURAL or STRUCTURAL to HUB? Construct the post-apply attestation graph for the relevant motif; compare to current weight class. If it tips, the bridge requires motifs-delta authorization, not bridge-card sign-off. See D7.

  • #13 — Stale-anchor. Check: For each extraction-note anchor cited in the bridge card, is the cited line from an ingest more than 6 months ago? If yes, did the bridge's author do a targeted raw-source check of the older passage's local context (per CLAUDE.md Rule 18)? If no record of a targeted check, the anchor is potentially stale. The pattern's specific bite: an older extraction note may extract a phrase that perfectly matches the bridge's claim while the surrounding extraction note's argument is not about the bridge's relation.

  • #14 — False-friend laundering through typed connection. Check: Does any Apply-as item add both a typed connection AND a false-friend caution on the same page pair? If yes, the bridge is admitting the relation is suspect and writing the relation into the body — the false-friend caution's "be careful, this might be a false analogy" gets buried under the typed connection's "X enacts Y." A clean bridge does one or the other, not both at once.

  • #15 — Citation-scheme mismatch. Check: For each evidence anchor citing a raw source, does the citation use the same scheme (page / paragraph / line / fragment) the source page declares in its citation_scheme: frontmatter field? Card 4 cites "SA-2006 raw 1047" / "raw 1287" / "raw 1678" — these are line numbers in the raw text, not the source-page-citation scheme used elsewhere for SA-2006. The pattern catches Faul-like citations where the citation form is plausible but doesn't match the source's actual organizational scheme.

  • #16 — Secondary-mediated cross-source convergence. Check: When a bridge cites cross-author convergence across multiple chapters of an edited volume (cards 1, 2, 5 all do this for M-C 2026), is the convergence attested by the primary author (MP in his own texts) or only constructed by the editor across chapters? The wiki's own discipline (CLAUDE.md "User memories nominate; artifacts authorize") has the structural analogue: editorial-constructed convergence nominates a corpus motif but does not authorize it as a stable MP-corpus register. This pattern is what card 1's Counterpressure paragraph half-says about déposer; making it a taxonomy pattern moves it from per-card discretion into systematic check.

  • #17 — Counter-test executability. Check: Is the bridge's stated Counter-test executable by the current sign-off agent (i.e., requires only what's available in raw/ and the extraction notes), or does it require a future ingest, second-source landing, or external corpus access? If the counter-test is "does this surface in V&I working notes?" and V&I is not yet in raw/, the Counter-test is unfalsifiable in the current state and should be flagged as deferred-Counter-test, not satisfied-Counter-test. This is the bridge-card analogue of math-olympiad's pattern #4 ("does this prove RH?"): a Counter-test that requires solving an open corpus-ingestion problem is no Counter-test.

  • Interpretation pass (pre-review step, like math-olympiad §1). Before role review begins: for each card, generate 2-3 alternative readings (different Primary home, different Bridge type, different relation specification). Mark the card's stated reading as one of these. Ask: which reading makes the bridge TRIVIAL? If one reading makes the bridge trivial and the card's stated reading is non-trivial, that's expected. If the card's stated reading is the trivial one and another reading is non-trivial, the card may be mis-stated. The check costs one fresh subagent per card and could be Role D ("interpretation-checker"), or could be a Step 0 done by main-thread before the 3 roles launch.

Would the taxonomy catch the actual 5 cards' defects?

Yes for some, no for others. Card-by-card:

  • Card 1 (déposer). The Counterpressure is candid: cross-author convergence attests editorial frame more than MP's lexicon. Pattern #1 (editorial-frame laundering) would fire on this — but the card has already self-flagged it. The question is whether a reviewer would defect on a card that flags its own problem. I expect 3-of-3 APPROVED with a note about the editorial frame, which is the same outcome as the current Pass 3 author intended. Missing pattern #16 (secondary-mediated convergence) would actually fire here and could change the outcome.
  • Card 2 (haunting). Same shape as card 1; same expected outcome. Pattern #1 fires partially, #11 fires partially (the bridge survives across 4 sub-anchors which feels suspicious-clean), but the Counterpressure explicitly names the secondary-author mediation.
  • Card 3 (endurance du chaos). Single-source attestation (SA-2006 only). Pattern #4 (anchor-counterpressure inversion) should fire: the Counter-test is "does endurance du chaos surface in V&I working notes, Eye and Mind, or Signs?" — but the wiki has not yet ingested V&I working notes (per raw inventory), so the Counter-test is unfalsifiable in the current state. Missing pattern #17 (counter-test executability) would fire decisively.
  • Card 4 (militante). Three SA-2006 attestations (raw 1047, 1287, 1678). Pattern #15 (citation-scheme mismatch) would fire (line-number citations where the source page uses a different scheme). Pattern #16 fires on "stable late-MP register is SA's interpretive emphasis."
  • Card 5 (past that could have been otherwise). Explicit single-source promotion blocker. The card already restricts itself to subsection + open-question level. Pattern #11 (too-clean-bridge) might fire on the "modal hinge that converts" formulation, which is suspiciously general. The card's explicit blocker is its defense; pattern #17 also fires (counter-test = second-source confirmation, requires future ingest).

Summary: with the current taxonomy, expected outcome is 5-of-5 APPROVED, possibly with adjudicate-but-approve on Card 3 and Card 5 over single-source concerns. With the missing patterns added (#16 and #17 especially), expected outcome shifts toward 1-2 cards getting blocked or returned for narrower apply-as scope. That's the calibration signal the pilot needs: D8 above.

Adversarial card construction

Constructed bridge card that passes all 6 gate tests and survives all 11 adversarial patterns but is bad:

## Card X — *Sichten* / "to sight" / phenomenological-sighting

**Status:** proposed
**Bridge type:** mechanism
**Primary home:** [[institution]]
**Secondary homes:** [[reduction]], [[wesensschau]], [[Stiftung]]

**Relation:**
- [[institution]] *enacts* the phenomenological-sighting (Sichten) mechanism in MP's late ontology; *Stiftung* names what is sighted, *Sichten* names how.
- [[reduction]] *is the condition of intelligibility of* the Sichten mechanism (one must reduce to see).
- [[wesensschau]] *contrasts with* Sichten on the eidetic-vs-deposited axis.

**Evidence anchors:**
- Husserl Hua VI §9c — "Sichten" used in the Krisis re: world-horizon thematization.
- MP Phenomenology of Perception, Forward 1945, "voir d'un coup" (the phrase MP uses for synoptic phenomenological seeing) — extraction note line 1402.
- M-C 2026 (any chapter discussing phenomenological method) — implicit via Stiftung-becoming.

**Evidence status:** extraction-note + source-text anchored.

**What it makes visible:** institution's mechanism is not just deposit-of-sense but the act of sighting that deposit, the phenomenological seeing that recognizes what was instituted. Without naming Sichten, institution becomes a flat ontological event with no methodological access.

**Counterpressure:** Sichten is more central in Husserl than in MP; the bridge could be reading Husserlian methodology into MP. **Counter-test:** does MP himself use "Sichten" or a French equivalent in working notes? *Note: V&I working notes are not yet in raw/; counter-test deferred until ingest.*

**Apply as:**
1. Subsection on [[institution]] under What the Concept Does: name Sichten as the methodological-access dimension; cite Husserl Hua VI §9c, MP PP Forward.
2. Typed connection on [[reduction]] Connections section: *is the condition of intelligibility of* the Sichten mechanism.
3. Open-question on [[wesensschau]]: flag eidetic-vs-deposited contrast.

**Approved by maintainer:** _(awaiting per-card sign-off)_

This card is bad because:

  • It poses a Husserl-mediated reading of MP as if it were MP's own move. Sichten in Krisis is Husserl's term; the "voir d'un coup" phrase MP uses in the PP Forward (extraction line 1402) is a general phenomenological-synoptic locution, not a technical déposer-grade mechanism term. The bridge wraps a generic phenomenological-method observation in the bridge-card 10-field shape to make it look load-bearing.

  • It walks past patterns #1 (editorial-frame laundering — there's no editor frame, this is a reviewer-frame laundering, where the reviewer projects Husserlian method onto MP's late ontology), #3 (false genealogy — the contrasts with and is the condition of intelligibility of relations imply a developmental chain that isn't attested), #10 (would settle longstanding controversy — yes it would, the Husserl→MP-method chain is exactly the kind of dispute the wiki hasn't earned a position in), #11 (too-clean — the bridge is suspiciously general).

  • Most of those patterns could fire on this card with current taxonomy, so let me say where it actually evades:

    • Patterns #1, #3, #10, #11 only fire if the reviewer recognizes the Husserlian-projection move. The card uses MP's "voir d'un coup" phrase as a bridge, which is technically MP's own language. A reviewer working in checklist mode (pattern #1 = "any anchor cite primary author's own text?" → yes, MP's PP Forward → check fails to fire) does not recognize that "voir d'un coup" is general phenomenological idiom, not load-bearing technical vocabulary.
    • Pattern #5 (Primary home arbitrariness): the card argues [[institution]] is Primary home with a one-line gloss; the reviewer can't easily falsify because any of the secondary homes could equally be primary.
    • Pattern #6 (Apply-as scope creep): each Apply-as item maps to one specific page section, technically.
    • Pattern #7 (Bridge-type conflation): "mechanism" is plausible.
    • Pattern #9 (Fabricated citation): Hua VI §9c is real; the PP Forward is real; MC 2026 has chapters discussing method. All citations resolve.

    So the card passes the structural gate, evades pattern-specific attacks on the strength of using a real but generic phrase as anchor, and could plausibly get 3-of-3 APPROVED.

The pattern that would catch it: an "anchor specificity / load-bearing test." Check: Is each cited anchor uniquely interpretable as the bridge's specific relation, or is it a generic phenomenological / philosophical idiom that the bridge has co-opted? Construct the bridge's relation from the cited anchor in isolation: would a reviewer naive to the bridge produce this relation, or would they produce a generic phenomenological-method observation? If the anchor underdetermines the relation, the bridge is over-claiming from a generic locution.

Where in the taxonomy: This should be pattern #18 — "Anchor underdetermination." Specific check: For each anchor, construct the most parsimonious reading of what the anchor attests. Does the bridge's claimed relation require interpretive moves beyond the parsimonious reading? If yes, the bridge is anchor-underdetermined and either (a) needs a different anchor that does the work, or (b) is over-claiming.

This pattern would also fire on Card 5 of the actual run3 (Larison's "the past that could have been otherwise" — single-source, single-passage; the modal-hinge claim requires interpretive moves beyond what the parsimonious reading of Larison's passage attests; the card's explicit single-source promotion blocker is the partial defense, but the parsimonious-reading test is sharper than "single-source" alone).

Proposed improvements

In rough priority order:

  1. Fix the quorum miscalibration (D1). Make 1-of-3 DEFECT-FOUND-with-confidence-high a block, not adjudicate. Add to §9.

  2. Constrain or replace the main-thread adjudicator (D2). Either restrict the adjudicator's action space to mechanical verdict-counting + narrow-the-apply-as-scope, or add a fresh-context adjudicator subagent. Move §11 example outputs to reflect this; remove the "B1's concern already in card §Counterpressure" reasoning pattern.

  3. Strip bridge prose for Roles B and C (D4). Add §3.5 Pre-review-prep with a mechanical strip pass. Role A keeps the unstripped card because the prose is part of what Role A checks.

  4. Add the 6 missing patterns (#12-#17 above, plus the Interpretation pass and the Anchor-Underdetermination pattern #18). Section §4 expands accordingly.

  5. Specify pilot success criteria (D8). Concrete 5-criterion conjunction; absence of any criterion means "revise, don't codify."

  6. Add cross-card consistency check (D9) after adjudication, before sign-off-line writes. §11.5.

  7. Resolve the carve-out bleed at typed-connection level (D7) with the motifs-weight tipping check (pattern #12).

  8. Split UNCLEAR (D5) into OUT-OF-ROLE-SCOPE / UNDER-SPECIFIED-CARD / CANNOT-DETERMINE; specify §9 actions for each.

  9. Replace test #6's adversarial framing (D6) with the structural-only "Counterpressure field present and specific."

  10. Enforce hypothesis-note discipline (D11) with the two-tool-call pattern (write hypothesis-note, then write report; can't edit hypothesis-note after report).

  11. Add the dead-link check to gate test #3 (D13). Wikilinks in Apply-as targets resolve to existing pages.

  12. Interleave card-batch assignments across role agents (D12). Each card's 3 reviewers come from different per-role agent positions.

  13. Add automatic deep-mode retry on DEFECT FOUND before retiring a card (math-olympiad §6d/Step 6d analogue). One fresh subagent gets the DEFECT report and tries to repair the card before retirement.

  14. Add retry policy for subagent failures (math-olympiad §2 analogue). One retry, then drop the card with explicit "subagent unavailable, manual review required" note.

  15. Refresh the §2.2 "you are the only reviewer" prompt language (D3) to honest operational language.

  16. Bridge-card-author ≠ adjudicator rule. Explicit rule that the agent that drafted the Pass 3 bridge cards may not be the adjudicator for those cards. In practice, this requires either splitting the v0d.7 codification into two sessions OR using a different invocation lineage for adjudication. Worth stating even if hard to enforce.

What the plan does well

  • The three-role split (A/B/C) is a defensible structure that maps cleanly to the bridge-card 10-field template. Each role has a distinct stance and the protocol differences (gate / pattern / evidence-chain) are operationalizable.

  • The §6 raw-access discipline is well-specified, with concrete budgets (5 raw reads per card per reviewer), a self-flag trigger for over-budget, and the right rationale (passage-level verification, not re-ingest). This matches CLAUDE.md Rule 18.

  • The §10 same-gap-twice check at batch level is a real innovation over math-olympiad, which doesn't have a cross-problem analogue. It's the right idea even if I'd want criteria sharpened.

  • The §12 carve-out correctly distinguishes what sign-off authorizes from what requires escalation. The boundaries are sometimes leaky (D7), but the direction of the distinction is right: subsection / typed-connection / open-question are bridge-card-sized writes; claim promotion to supported / motif HUB / deletion are not.

  • The §13 pilot-before-codify discipline is exactly right in shape. The criteria need work (D8), but the two-commit sequence (pilot + revisions → v0d.7 bump) is a good operational pattern.

  • The §15 "What this plan does NOT do" section is clear about scope: this is sign-off procedure, not apply-mode procedure, not schema bump, not retire-v0d.5. Scope discipline.

  • The dual file-path isolation choice (§2.3) — file-path subdirectories rather than worktrees — is well-reasoned. Subagent contexts are inherently isolated; worktrees would add complexity without marginal safety.

Verdict

ACCEPT-WITH-CHANGES

The plan's core architecture is correct: three role-typed reviewers, pattern-armed adversarial taxonomy, dual context isolation, hypothesis-note discipline, raw-access budgeting, asymmetric quorum, pilot-before-codify, carve-out. The plan is more rigorous than what the v0d.5 schema currently requires (a date on a line) and substantially closes the rubber-stamping risk.

However, D1 (quorum miscalibration), D2 (main-thread adjudicator is a hidden coordination layer), D4 (bridge-prose stripping missing), D5 (calibrated-abstention missing), D7 (motifs-weight bleed in carve-out), D8 (pilot success criteria empty), D9 (cross-card dependency), and D10 (taxonomy missing wiki-specific patterns) are load-bearing defects that, taken together, would let the procedure rubber-stamp the actual 5 cards in run3 with high confidence in a way the procedure itself claims to prevent. None of these are unfixable; all of them have concrete proposals above.

Pre-pilot: address D1, D2, D4, D5, D8 (the math-olympiad transfer-fidelity defects) and add at least #16, #17, #18 to the taxonomy (the wiki-specific failure-mode patterns). These are the changes without which the pilot risks producing 5-of-5 silent APPROVE outcomes that the plan would misread as "pilot succeeded."

Post-pilot, before v0d.7 codify: address D7, D9, D10's full pattern list, D11, D13.

The remaining defects (D3, D6, D12, plus improvements 13-16) can be addressed at the codification step or in a v0d.7.1 follow-up.