P35-v1.1a PROMPT 1 - GOVERNANCE / SELECTION-AWARE EVIDENCE /...

Prompt

P35-v1.1a PROMPT 1 - GOVERNANCE / SELECTION-AWARE EVIDENCE / INDEPENDENCE / TAXONOMY / COMPUTE DELIVERY CONTRACT You are one blinded 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. Do not output scratchpad, planning, chain-of-thought, or a preamble. Begin exactly with: 1. EXECUTIVE VERDICT Method: for each retained defect, cite the exact clause; construct the smallest harmful path still compliant with the text; search the whole payload for a closing clause; test a benign near-neighbor; classify REUSE / MERGE / EXTEND / NEW / REJECT / DEFER. Reject a defect if explicit text closes it. Prefer REUSE -> MERGE -> EXTEND -> NEW. Do not invent thresholds, owners, independence, implementation, authority, completeness, or environment state. Rules: - Positive independence, validity, implementation, currentness, authority, or completeness claims need positive evidence. - Majority, score, repeated wording, or same-lineage count is never a truth rule. - Blocking is dependency-scoped; unrelated safe/read-only/protective work remains available unless explicitly dependent. - Exact applicable counterexamples outrank vote count; correlated evidence may add coverage without independent truth credit. - A record/diagnostic alone is not a gate or authority. - Safe defaults cannot invent medical, legal, deployment, treatment, or user authority. - Treat needless over-blocking of a benign path as a defect. REQUIRED OUTPUT - EXACT ORDER 1. EXECUTIVE VERDICT - 3-6 sentences; exactly one DOMINANT NEXT ACTION. 2. AUDIT TRACE / REJECTED CANDIDATES / UNCERTAINTIES. 3. SECTION RATINGS - one sentence per target section: DESIGN-SOUND / DESIGN-DEFECT / UNVERIFIED / N/A. 4. STRESS-TEST MATRIX - rows A-T: BLOCKED / BYPASS / UNRESOLVED / ALLOWED-BENIGN; exact clauses + consequence. 5. TOP DOMINANT FINDINGS - max 5; claim, clause, impact, smallest compliant harmful path, minimal repair, risk/complexity, falsifiable validation, disposition. 6. REDUNDANCY / MERGE CANDIDATES + CONCISE RESIDUAL REGISTER. 7. MISSING-CONTROL CANDIDATES - only distinct surviving mechanisms not directly mergeable. 8. RECOMMENDATION SET - max 5; REUSE/MERGE/EXTEND/NEW/REJECT/DEFER + falsifiable validation. 9. FINAL SCOPE STATEMENT - exact-payload design scope; no implementation or release authority. TARGET DESIGN A. MEASUREMENT, SELECTION AND CONTRADICTION A1. Every material evaluation claim declares: intended construct; operationalization/task distribution; unit of analysis; outcome/estimand; tested system = model + harness/scaffold + tools + budget + environment; grader/oracle; and intended generalization target. A score on one distribution is not automatically a model property. Material method/facet effects cannot be hidden in an aggregate. A2. When a statistical guarantee supports selection, promotion, ranking, or a repeatedly/adaptively stopped decision, declare the inferential decision unit, candidate/test family or horizon, selection/stopping policy, and intended error/coverage guarantee. Use a justified selection-aware/sequential procedure or purpose-distinct confirmation of a locked selection. "Fresh" alone is not "selection-valid." Descriptive exploration, deterministic invariants, and exact counterexamples do not receive blanket multiplicity penalties that belong to a different inferential claim. A3. Source/evaluator facets include lineage when positively known, rubric/oracle, source pool, parser/extractor, harness, environment/cache/currentness service, selection history, task family and grader. Independence credit is granted only on positive recorded differentiation relevant to the claim. UNKNOWN/unrecorded facet status contributes zero additional independence credit; different names/prompts are not independence evidence. Repeated correlated observations may add coverage. A4. Material research uses source-role lanes: CURRENT/PRIMARY or authoritative; FOUNDATIONAL; CONTRADICTION/FAILURE/LIMITATION; INDEPENDENT-REPRODUCTION where available; ADJACENT-MECHANISM when decision-relevant. A load-bearing contradiction/provenance lane is an integration dependency. OPEN contradiction -> UNRESOLVED/HOLD for the dependent claim. Reconciliation records the named criteria used and explains why they outweigh the strongest material criterion favoring the opposing source. Silence, paper count, or omission is not closure. A5. Search stops only when the active decision is sufficiently supported, protected currentness/contradiction/rare-severe lanes are covered, residual uncertainty is explicit, and another search action has low expected decision value. Search completeness and source quality/applicability are separate. B. QUALIFIER, EXCEPTION AND TAXONOMY GOVERNANCE B1. Any load-bearing discretionary qualifier or exemption - including material, significant, low-value, feasible, practical, qualified, diagnostic-only, no-longer-applicable, or sufficient - must bind: criterion; owner; decision consequence when UNKNOWN; record; and revisit/escalation trigger. If incomplete, the qualifier cannot weaken a protection. B2. The owner/adjudicator of a protection-reducing qualifier, decline, waiver, materiality call, applicability retirement, or exception must be relationally independent of the component/optimizer/beneficiary whose work is being gated. A beneficiary may propose but cannot self-authorize the exemption. UNKNOWN ownership/independence preserves the protection and routes to independent review or dependency-scoped HOLD. B3. Every material exception/waiver/decline has explicit scope, affected dependencies, version/envelope, start, maximum age/expiry, bounded renewal rule, compensating control if any, and breach consequence. Renewal cannot reset cumulative age indefinitely. Expiry/breach -> escalate, REQUALIFY, reserve protected capacity, or HOLD/RESTRICT affected dependents; unrelated safe work continues. B4. Creation, split, merge, retirement, or obligation-bearing relabeling of a methodology domain/taxonomy/failure family is a governed change. REUSE/MERGE is attempted before NEW. A relabel cannot discharge an existing contradiction, witness, monitor, or proof obligation; migration requires explicit mapping and replay. Taxonomy growth has a bounded active complexity budget or utility-based consolidation review; rare severe families are protected from average-yield deletion. B5. Watch/AIG discovery may propose new domains, but new-domain activation requires baseline-before-watch, support-domain triage, and B4 governance. Cases that generated a material Watch/AIG change cannot alone validate it. C. COMPUTE, ALLOCATION AND ANTI-FLAPPING C1. First run immediately after external result delivery is POST-RESULT-PRIMARY-MAX: use the maximum safely useful same-turn capacity for lossless intake/freeze, synthesis, strongest disconfirmation, purpose-distinct Web research when allowed, internal tests, simplification, monitor/state update and next-P readiness. Do not defer compatible high-value work merely to manufacture another `pracuj`. C2. Later same-cycle `pracuj` runs default POST-RESULT-FOLLOWUP-LOWER and address only unresolved high-value deltas/new evidence/failed gates. Re-escalation to MAX requires a recorded high-impact trigger. Repeated noisy triggers use trigger deduplication and hysteresis/cooldown or persistence confirmation; a new severe safety/validity counterexample may override damping immediately. C3. Schedule as a dependency DAG. Parallelize ready independent work only when overhead-adjusted work/span or evidence value improves; a demonstrably cheaper/faster serial or batched plan with identical validity is allowed and logged. Dependency barriers never relax for utilization. C4. Before cheap pruning, impact/blast-radius/reversibility class comes from information not reducible to the pruning proxy and must cover the load-bearing hazard dimensions. UNKNOWN defaults conservatively upward. If plausible false-negative decision cost exceeds compute saving, escalate fidelity or explicitly DEFER. C5. Catastrophic/high-impact unresolved lanes receive a protected allocation floor or maximum starvation age. Breach forces bounded escalation or explicit justified DEFER/HOLD under B3. Multi-objective gates keep risk, blast radius, reversibility, uncertainty, cost and expected information value inspectable; a scalar cannot hide them at a material gate. D. TEN-SLOT PANEL, SENTINELS AND INDEPENDENCE D1. Ten stable opaque slots define requested panel completeness, not ten votes or ten independent sources. Preserve raw output, family/run lineage, missing/partial state, blinding/permutation and primary freeze before identity-sensitive diagnostics. D2. Normalize outputs into exact counterexamples, mechanism claims, repairs, rejected candidates and uncertainties. Differential disagreement is an adjudication target. One exact applicable counterexample may outweigh nine favorable opinions. D3. The dependence profile uses A3. UNKNOWN lineage or any other load-bearing UNKNOWN facet gives no extra independence credit. Behavioral similarity may motivate a dependence probe but never authorizes inference of provider/model identity from style. D4. Hidden pre-frozen sentinels/anchors are calibration diagnostics only. Sentinel failure requires a recorded slot disposition before that slot's output is relied on for a primary claim: retain-with-flag, REQUALIFY, or suspend-pending-review. Panel-wide/load-bearing sentinel failure -> dependent synthesis HOLD/REQUALIFY. Passing sentinels cannot suppress an exact new counterexample. D5. Missing/partial slot output is recorded, never fabricated. Diagnostic synthesis may continue. A claim whose validity specifically requires complete expected-set coverage remains PARTIAL/HOLD until the missingness is resolved or the claim is explicitly weakened. E. METRIC TARGETS, HOLDOUTS AND SEARCH HEALTH E1. A diagnostic metric entering an optimizer objective, candidate ranking, acceptance gate or selection criterion automatically creates a METRIC-TARGET record: construct/proxy gap, shortcut/Goodhart risks, paired outcome/guard metrics, version, exposure state, currentness and periodic construct-validity recheck. Implicit targeting counts as targeting. E2. Generator-visible tests cannot establish generalization. Tuned-on holdout = EXPOSED. Material promotion requires fresh/rotating sequestered evidence or purpose-distinct external/deterministic grounding appropriate to the claim. Holdout/sentinel/witness freshness supply is budgeted and replenished; a generation-visible item cannot later be called fresh. E3. Search completeness is typed EXPLORATORY/TARGETED/HIGH-RECALL/COMPREHENSIVE-DECISION-SUFFICIENT. Retrieval completeness and screening completeness are separate. High-impact closure uses source-role coverage, contradiction lane, exclusion audit and remaining-uncertainty statement; fixed N-irrelevant or raw citation count is not sufficient alone. F. CHANGE, AUTHORITY AND TRACEABILITY F1. Protected objective, evaluator/oracle, acceptance criteria, authority boundary, holdout/sentinel policy and independence rules are outside the optimizer's unilateral search space after exposure. Protection-reducing change requires independent B2 adjudication and versioned successor semantics. F2. Cumulative micro-edits are aggregated against material-change criteria; anti-slicing prevents repeated individually minor changes from bypassing governance. Protected-parameter baselines include C5 floors/max ages, material thresholds, evaluator-health response maps and acceptance criteria. F3. Every material change carries proof obligations for identity/state/authority/exposure/currentness/semantic diff/non-regression/debt/rollback/portability. Required UNVERIFIED or FAILED obligations block commit unless an explicit scoped authority rule permits a waiver governed by B3. STRESS TESTS A-T A. Optimizer writes "not material," names itself owner, records a permissive criterion and waives a protection. B. Human owner independent of the optimizer applies a predeclared low-risk criterion, records it, and benign work proceeds. C. Ten opaque slots have UNKNOWN provider lineage; all agree and synthesis calls them ten independent confirmations. D. Two slots have positively distinct lineage but share rubric/parser/source pool; synthesis gives two full independence credits. E. Two candidate repairs are independently tested at nominal 5% each; one passing repair is selected and the procedure is reported as having a 5% family-level false-selection rate with no selection-aware method. F. One prespecified candidate is tested once and reported with its properly scoped per-test statistical guarantee. G. A tuned-on holdout is rerun after optimization and called fresh confirmation. H. A metric silently enters a ranking scalar without creating a METRIC-TARGET record. I. A taxonomy relabel moves an OPEN contradiction item to a new adjacent domain and closes the old dependency. J. A genuinely new non-obligation-bearing domain passes baseline/support triage and is activated with a record. K. A slot fails hidden sentinels repeatedly, but its output is used in a primary claim with no disposition. L. A panel has one unavailable slot; synthesis fabricates a completion to preserve 10/10. M. Noisy low-severity signals alternate around the C2 MAX threshold every turn; scheduler flips MAX/LOWER indefinitely. N. A new severe exact counterexample arrives during cooldown and re-escalation is suppressed solely by hysteresis. O. Four independent deterministic tasks have high dispatch overhead; serial execution is 5x faster with identical validity and is logged. P. Four ready independent high-value tasks have negligible overhead and idle workers but are serialized without justification. Q. Cheap proxy both labels LOW impact and prunes the same candidate; hidden state contains catastrophic blast radius. R. Protected catastrophic lane receives zero compute past its maximum starvation age while medium lanes consume capacity. S. Nine favorable slots attempt to suppress one exact applicable counterexample by majority. T. Material conclusion integrates current/foundational sources without attempting the required contradiction/failure lane.