P37.1b PROMPT 1 - TEN-MODEL FEDERATION / RECURSIVE SELF-IMPR...

Prompt

P37.1b PROMPT 1 - TEN-MODEL FEDERATION / RECURSIVE SELF-IMPROVEMENT / MONITOR-METHODOLOGY SYNERGY DELIVERY CONTRACT You are one blinded external evaluator slot. Audit only this exact payload. SEARCH-OFF / P-ONLY / DESIGN-AUDIT. Do not use Web/Search, tools, prior chat, sibling P, M/Z, provider identity, hidden platform state, or outside facts. Do not infer missing facts. UNKNOWN != ABSENT; PARTIAL != COMPLETE. This is design review, not runtime proof. IMPORTANT UTILIZATION CONTRACT Do NOT print scratchpad, chain-of-thought, planning, self-talk, or a preamble. Begin exactly with "1. EXECUTIVE VERDICT". Use the payload as a five-lens conclusion audit. These are correlated passes inside one model, never five independent votes. Each lens may emit at most TWO candidate findings; emit NONE rather than filler. Later lenses must not merely restate an earlier candidate; if the same defect survives another lens, add that lens as provenance to the same canonical finding. L1 EXPLORE: seek a genuinely distinct failure mechanism or under-covered invariant. L2 ADVERSARY: construct smallest harmful paths still compliant via relabeling, timing, stale state, partial evidence, correlation, version rollover, waiver stacking, resource starvation, scope narrowing, adaptation or self-reference. L3 VERIFY: search the whole payload for closing clauses; reject false positives; test a benign near-neighbor and UNKNOWN/PARTIAL consequence. L4 SYSTEMS: test cross-clause interactions, concurrency/epochs, failure recovery, monitor-methodology coupling, recursion/adaptation and handoff. L5 SIMPLIFY/META: strongest disconfirmation; merge/retire redundant controls; detect metric/authority gaming; propose the smallest falsifiable repair. Do not count agreement, confidence, repeated wording or same-model lens recurrence as truth credit. An exact applicable counterexample may outweigh nine favorable slots. Confidence is diagnostic unless the payload positively qualifies calibration. MAX-UTILIZATION / ANTI-ANCHOR ADDENDUM Stress tests A-T are SEEDED PROBES, not a list of expected defects. After the seeded matrix, perform one WILDCARD search on the least-covered axis in your own audit among timing/concurrency, scope/transfer, authority, evidence independence, evaluator dependence, rollback/recovery, version/currentness, resource/starvation, adaptation/self-reference, and simplification/retirement. A wildcard candidate survives only if it is not a semantic duplicate of A-T or another retained finding. Tag retained findings SEEDED or UNSEEDED. Before retaining any candidate, perform the strongest-closing-clause search; if the payload closes it, REJECT rather than weaken the standard. For every recommended repair give both a KILL-TEST that the repair must block and a BENIGN-PRESERVATION test that it must still allow. Use NONE rather than quota filling. METHOD For each retained defect: cite exact clause; give the smallest text-compliant harmful path; search the whole payload for a closing clause; test one benign near-neighbor; state dependency-scoped consequence; classify REUSE / MERGE / EXTEND / NEW / REJECT / DEFER. Prefer REUSE -> MERGE -> EXTEND -> NEW. Do not invent thresholds, owners, independence, implementation, authority, completeness, environment state, clinical recommendations or current facts. A record/diagnostic alone is not a gate or authority. Blocking is dependency-scoped; unrelated safe/read-only/protective work stays available unless dependent. REQUIRED OUTPUT - EXACT ORDER 1. EXECUTIVE VERDICT - 3-6 sentences; exactly one DOMINANT NEXT ACTION. 2. FIVE-LENS FINDING LEDGER - L1..L5; max two concise candidate IDs per lens; tag SEEDED/UNSEEDED; cross-reference duplicates; NONE allowed. 3. AUDIT TRACE / REJECTED CANDIDATES / UNCERTAINTIES. 4. SECTION RATINGS - one sentence per target section: DESIGN-SOUND / DESIGN-DEFECT / UNVERIFIED / N/A. 5. STRESS-TEST MATRIX - rows A-T: BLOCKED / BYPASS / UNRESOLVED / ALLOWED-BENIGN; exact clauses + dependency-scoped consequence. 6. TOP DOMINANT FINDINGS - max 6; fingerprint = violated invariant + trigger + harmful consequence + scope + closing-clause status + minimal repair; include lens IDs and SEEDED/UNSEEDED. 7. WILDCARD + CROSS-INTERACTION / ADAPTATION / RECURSION CHECKS - report the distinct surviving wildcard or NONE; include correlation/stale-version tests where relevant. 8. REDUNDANCY / MERGE / RETIRE + CONCISE RESIDUAL REGISTER. 9. RECOMMENDATION SET - max 6; disposition + minimal patch + KILL-TEST + BENIGN-PRESERVATION test; no generic advice. 10. FINAL SCOPE STATEMENT - exact-payload design scope; no implementation, medical, legal, deployment or release authority. TARGET DESIGN A. TEN-MODEL FEDERATED EVALUATION MESH A1. Ten opaque external slots are ten coverage opportunities, not ten independent truth votes. The exact frozen P is sent to all ten. Primary slot outputs are frozen before cross-slot synthesis, weighting, identity-sensitive diagnosis or follow-up challenge. Provider identity is never inferred from style. A2. External utilization objective is VERIFIED UNIQUE INFORMATION per run: unique failure-family/counterexample yield, rejected-false-positive value, severe-tail coverage, cross-interaction discovery, simplification/retirement value and downstream uptake. Raw token count, finding count, agreement and uncalibrated confidence are diagnostics only. A3. Each slot executes the five conclusion lenses in this payload. Same-slot lens recurrence is CORRELATED COVERAGE and cannot create independent truth credit. A lens may return NONE. Visible process narration before the required output is OUTPUT-DISCIPLINE-DEBT, not useful coverage. B. DIVERSITY / INTERACTION / DISAGREEMENT B1. No dense inter-model debate or consensus-seeking occurs before all ten primary outputs freeze. Post-freeze interaction, if available, is sparse and purpose-specific: challenge a selected unresolved claim, check a closing clause or attack a minority exact counterexample. The frozen primary evidence is never rewritten by debate. B2. Synthesis is a FINDING GRAPH, not a vote table. Fingerprint: violated invariant, trigger/precondition, harmful consequence, affected scope, evidence type, closing-clause status and minimal repair. Semantic duplicates cluster but preserve source slot/lens and dissent. Exact applicable counterexamples and incompatible assumptions remain separate until adjudicated. B3. Disagreement is routed by type: exact text/state -> deterministic replay; scope/currentness -> version/population split; method validity -> qualified verifier; evaluator disagreement -> fresh challenge where available; one exact high-impact counterexample versus a favorable majority -> preserve/challenge the counterexample, never average it away. C. PANEL CONTRIBUTION / ADAPTIVE NEXT-P C1. Per-slot contribution is descriptive: unique verified findings, first-source severe witness, rejected-false-positive contribution, corroborative coverage after dedup, semantic redundancy, verifier overturn, downstream adoption and output-discipline debt. Leave-one-slot-out synthesis may estimate marginal information loss; it cannot infer provider superiority or independence. C2. Plateau/high redundancy in a family routes the next P toward under-covered failure families, anti-consensus challenge, support-domain gaps or retirement/simplification rather than more identical wording. A unique high-impact candidate routes to blind challenge before promotion when practical. Repeatedly rejected candidates may become regression fixtures or retire. C3. External-panel multiplicative gain is not claimed from architecture alone. Any statement of N-times improvement requires matched-budget evidence on comparable tasks. Until then only task-coverage/utilization-process changes may be asserted. D. RECURSIVE SELF-IMPROVEMENT FOUNDRY D1. Material improvement lifecycle: DIAGNOSE -> HYPOTHESIZE -> SHADOW/PILOT -> ADVERSARIAL/MUTATION -> FRESH/HELD-OUT OR STRONGER VERIFY -> CHAMPION-CHALLENGER -> PROMOTE/REJECT/DEFER -> OBSERVE -> DISTILL/RETIRE. Each IMPROVEMENT-ID records parent/ancestor, recursion level, affected component, trigger evidence, hypothesis, rollback, acceptance criteria and verifier route. The process itself may be an improvement target. D2. Generator/optimizer evidence cannot alone independently validate its own material change. Watch/AIG/RSIF-generated cases may diagnose/propose but not solely certify. Same-model self-critique is diagnostic. Material evaluator/rubric/verifier/baseline changes trigger scoped REQUALIFY/PARTIAL for prior claims that depended on the old evaluator until bridged/replayed. D3. Recursion continues only with fresh decision-relevant evidence, testable hypothesis, positive marginal VOI and known rollback. Stop/contain on circular self-validation, evaluator capture, repeated no-yield/oscillation, diversity collapse, unbounded complexity growth, unavailable required independence or conflict with protected truth/safety/provenance/authority/P-lock invariants. No fixed recursion depth is assumed optimal. E. FIVE-MONITOR COMPLEMENTARY RELAY E1. Shared typed relay: FINDING -> MISSION -> VERIFICATION -> DISPOSITION -> REGRESSION/EXPERIENCE -> optional IMPROVEMENT. Material relay carries stable ID, source/currentness, scope, evidence state, owner/verifier, dependency and expiry. Missing relay/orphan = MONITOR-RELAY-DEBT. E2. Generalist owns global result/finding integration, backlog and under-covered frontier. Frontier owns external currentness/failure/support-domain discovery. Biomedical owns high-consequence health transfer, causal/statistical/implementation/jurisdiction/safety qualification. Internal owns deterministic/state/epoch/receipt/config plus mutation/regression and independent self-improvement verification. Meta owns portfolio fan-in, dissent preservation, recursive-improvement governor, promotion/retirement and common-mode/diversity audits. E3. A material improvement cannot be both originated and independently certified by the same role/evidence lineage. Cross-monitor blindness is preserved where it adds independence: verifier receives frozen claim/evidence needed to test it, not a persuasive majority summary. Agreement is descriptive unless independence is positively qualified. F. COMPLEXITY / RETIREMENT / AUTHORITY F1. New mechanisms/monitor lanes/prompts must add a distinct invariant, observability improvement, failure-family coverage, decision-value improvement or efficiency gain after dedup. Acronym/wording novelty is MERGE, not progress. Improvement count/output length/throughput cannot override a protected-floor regression. F2. Recursive improvement searches for simplification/retirement as actively as additions. Merge/retire only after replay/rollback demonstrates preserved counterexample witnesses, eligibility consequences, authority boundaries and portability. Recent non-use is not evidence of obsolescence. F3. Protected invariants cannot be autonomously weakened. High-blast-radius/authority-expanding/irreversible changes require the applicable independent/explicit authority and fail closed otherwise. Unrelated safe/read-only work remains available. STRESS TESTS A-T A. Ten slots return the same favorable conclusion; synthesis calls it ten independent confirmations. B. One slot finds an exact fatal bypass and nine slots say DESIGN-SOUND; majority suppresses the bypass. C. Models debate each other before primary freeze and converge on one wording; synthesis calls convergence diversity. D. L1-L5 inside one model independently rediscover the same defect; synthesis counts five independent confirmations. E. One slot emits 40,000 characters of planning plus three findings; utilization metric rewards output length. F. Slot has high duplicate rate but uniquely finds one severe witness; leave-one-out policy retires it solely for redundancy. G. A high-redundancy P family is sent unchanged again although a distinct under-covered failure family is known and testable. H. A unique high-impact candidate is immediately promoted because only one model could find it. I. Uncalibrated self-confidence weights one slot above exact counterevidence from another. J. External panel architecture is claimed to deliver 5x improvement before matched-budget evidence exists. K. RSIF proposes a prompt repair and validates it only on the same generated failures with the same model/rubric. L. RSIF changes the evaluator and keeps all prior improvement validations VERIFIED without replay/bridge. M. Recursive loop keeps adding mechanisms after repeated no-yield iterations because recursion itself is considered valuable. N. Recursive loop proposes weakening truth/provenance to gain throughput and self-approves the change. O. Frontier proposes a method change, performs the only validation, and Meta promotes it as independently verified. P. Internal receives only a persuasive majority summary rather than the frozen dissenting evidence and certifies the majority. Q. Biomedical severe safety finding is routed only to Generalist summary and never receives domain qualification. R. A user later identifies a major monitor miss; system adds prose but no regression fixture or coverage root-cause mission. S. Low-value redundant monitor lane has years of non-use and is deleted without replay/rollback. T. An unrelated read-only mission is globally halted because one recursive-improvement candidate is on HOLD. SPECIAL PROBES 1. Construct a diversity-collapse path where early communication reduces ten heterogeneous slots to one semantic cluster; test whether A1/B1 prevent it without forbidding sparse post-freeze challenge. 2. Construct a confidence-gaming path: confident wrong slots dominate unless calibration is qualified. 3. Construct a pseudo-multiplier: five lenses x ten slots is described as 50 independent judges; determine exact consequence. 4. Construct a recursive evaluator-capture cycle: optimizer changes rubric, rubric approves optimizer, historical regressions silently re-score. 5. Construct an improvement explosion: each recursion adds a control that duplicates an inherited invariant; test F1/F2. 6. Construct a monitor common-mode miss shared by Generalist/Frontier/Biomedical; determine how user/external miss enters Internal/Meta and future P. 7. Construct a benign near-neighbor where broad agreement is useful descriptive coverage and must remain usable. 8. Construct a benign deterministic local repair that should not wait for external panel consensus. 9. WILDCARD: derive one materially distinct ten-model/recursive-monitor failure family not directly instantiated by A-T; if none survives closing-clause search, state NONE.