REPAIR COMPARISON TEST — FM-ADJ-001a ROLE Jsi nezávislý rig...

Prompt

REPAIR COMPARISON TEST — FM-ADJ-001a ROLE Jsi nezávislý rigorózní auditor návrhu LMC. Posuzuj pouze text uvedený v tomto zadání. Nepoužívej předchozí odpovědi jiných modelů, neodhaduj jejich názory a neprováděj žádné hlasování podle očekávané většiny. CÍL Máme již zjištěnou interpretační nejednoznačnost kolem pravidla STALE. Nyní porovnej tři možné MINIMÁLNÍ opravy tohoto jediného mechanismu. NEHODNOŤ jiné části LMC. Neprováděj redesign celého promptu. Neřeš provenance/dependency ledger jako samostatný požadavek; pouze uveď jeho praktický význam, pokud je relevantní. KONTEXT Původní problematické pravidlo D10.1: "If an accepted finding invalidates an intermediate conclusion already used by any performed step relevant to the decision, every downstream step or artifact that consumed that conclusion becomes STALE. A STALE downstream step must be re-evaluated from the corrected conclusion, or its residual effect must be included in an explicit final limitation. Changing CURRENT CANDIDATE alone does not revalidate STALE downstream steps. FINAL STOP is inadmissible while any material STALE downstream step remains unreviewed." PROBLÉM Dependency graph: A → D1 → D2 → D3 → FINAL DECISION A je intermediate conclusion. D1 přímo použil A. D2 použil pouze výstup D1. D3 použil pouze výstup D2. Všechny kroky jsou decision-relevant. A je následně accepted findingem invalidováno a opraveno na B. Při původním textu existuje spor mezi: 1. DIRECT: STALE pouze u přímých konzumentů A. 2. TRANSITIVE: STALE u celého downstream dependency closure. 3. ITERATIVE: nejprve D1 STALE; po jeho re-evaluaci může změna D1 vyvolat další STALE na D2 atd. NYNÍ POSUZUJ TYTO TŘI VARIANTY. VARIANTA A — CURRENT Ponechání původního textu beze změny: "If an accepted finding invalidates an intermediate conclusion already used by any performed step relevant to the decision, every downstream step or artifact that consumed that conclusion becomes STALE. A STALE downstream step must be re-evaluated from the corrected conclusion, or its residual effect must be included in an explicit final limitation. Changing CURRENT CANDIDATE alone does not revalidate STALE downstream steps. FINAL STOP is inadmissible while any material STALE downstream step remains unreviewed." VARIANTA B — EXPLICIT TRANSITIVE DEPENDENCY Jediná podstatná změna je explicitní rozšíření scope: "If an accepted finding invalidates an intermediate conclusion already used by any performed step relevant to the decision, every downstream step or artifact that directly or transitively depends on that conclusion becomes STALE. A STALE downstream step must be re-evaluated from the corrected conclusion, or its residual effect must be included in an explicit final limitation. Changing CURRENT CANDIDATE alone does not revalidate STALE downstream steps. FINAL STOP is inadmissible while any material STALE downstream step remains unreviewed." VARIANTA C — EXPLICIT TRANSITIVE + ITERATIVE PROPAGATION "If an accepted finding invalidates an intermediate conclusion already used by any performed step relevant to the decision, every downstream step or artifact that directly or transitively depends on that conclusion becomes STALE. STALE status propagates through the dependency chain: any step or artifact that consumed the output of a STALE step is itself STALE, and this propagation continues until no unreviewed consumer relevant to the decision remains. A STALE downstream step must be re-evaluated from the corrected conclusion, or its residual effect must be included in an explicit final limitation. Changing CURRENT CANDIDATE alone does not revalidate STALE downstream steps. FINAL STOP is inadmissible while any material STALE downstream step remains unreviewed." TEST CONDITIONS Assume: • A is invalidated and corrected to B. • D1, D2 and D3 were already performed before correction. • D1 directly consumed A. • D2 consumed only D1's old output. • D3 consumed only D2's old output. • All three are material and decision-relevant. • No new independent user task appears. • No external Web Search is needed. • No hidden provenance ledger may be assumed unless the text explicitly requires it. • Do not rescue a deficient variant by appealing only to author intent or "what the rule obviously meant". TEST 1 — DIRECT COUNTEREXAMPLE Can the following state legally survive? A→B D1 = STALE → reviewed from B D2 = old output derived from D1 D3 = old output derived from D2 FINAL STOP For each variant, classify: BLOCKED / STILL POSSIBLE / NOT DETERMINABLE. TEST 2 — TRANSITIVE CLOSURE After A→B, must D1, D2 and D3 all become STALE before FINAL STOP? Classify each variant: COVERED / PARTIALLY COVERED / GAP / NOT DETERMINABLE. TEST 3 — MULTI-HOP CASCADE Use A→D1→D2→D3→D4→FINAL. Determine whether the wording guarantees that invalidation of A reaches D4 without requiring an unstated new rule. TEST 4 — RE-EVALUATION CHANGE Suppose D1 is re-evaluated from B and its output changes from D1_old to D1_new. Must D2 become STALE? Then D2 is re-evaluated and changes from D2_old to D2_new. Must D3 become STALE? Distinguish carefully between: • initial propagation from A; • propagation caused by change of a previously stale step's output; • a new accepted finding; • mere re-evaluation. TEST 5 — LIMITATION ESCAPE Suppose a stale step cannot be fully re-evaluated and the orchestrator instead uses the "explicit final limitation" branch. Can that branch be used to terminate while a downstream consumer still contains materially contaminated information? For each variant state exactly what the limitation must cover, and whether a downstream escape remains possible. TEST 6 — OVER-INVALIDATION Construct the smallest safe graph containing: A → D1 A → X D1 → D2 where X is parallel to D1 and does not depend on A. Does each variant correctly stale D1/D2 while leaving X non-stale? Identify any wording that risks making unrelated or parallel work stale. TEST 7 — NO FALSE CASCADE Construct a graph where D1 consumes A but its output is proven unchanged after re-evaluation from B. Does each variant nevertheless force unnecessary re-evaluation of every downstream node? Distinguish safety requirement from unnecessary work. TEST 8 — LOOP / TERMINATION Can any variant create an unbounded revalidation loop? If yes, determine whether that is: • a safety failure; • a liveness/efficiency risk only; • blocked by the existing explicit-limitation termination path. Do not label a liveness risk as a safety failure without proving that final output can become unsafe. TEST 9 — MINIMALITY For each variant, identify the smallest semantic change relative to CURRENT. Determine whether B or C introduces any requirement that is not necessary to close the observed FM-ADJ-001a ambiguity. TEST 10 — REGRESSION For each variant, answer: • Does it block the original direct counterexample? • Does it force transitive closure? • Does it force propagation after changed re-evaluation output? • Does it preserve safe non-dependencies? • Does it preserve the limitation branch? • Does it create a new material failure mode? FINAL ADJUDICATION Rank the three variants: 1 = best 2 = second 3 = worst The ranking must be based on: • epistemic determinacy; • ability to block the counterexample; • completeness of dependency invalidation; • resistance to semantic escape routes; • avoidance of over-invalidation; • termination/safety; • minimality of intervention. Do NOT rank by stylistic elegance. Do NOT rank by majority expectation. Do NOT assume that more text is better. MANDATORY OUTPUT VARIANT A: VERDICT: TEST 1: TEST 2: TEST 3: TEST 4: TEST 5: TEST 6: TEST 7: TEST 8: TEST 9: TEST 10: RESIDUAL GAP: CRITICAL TEXT: WHY: VARIANT B: VERDICT: TEST 1: TEST 2: TEST 3: TEST 4: TEST 5: TEST 6: TEST 7: TEST 8: TEST 9: TEST 10: RESIDUAL GAP: CRITICAL TEXT: WHY: VARIANT C: VERDICT: TEST 1: TEST 2: TEST 3: TEST 4: TEST 5: TEST 6: TEST 7: TEST 8: TEST 9: TEST 10: RESIDUAL GAP: CRITICAL TEXT: WHY: FINAL RANKING: 1. 2. 3. PREFERRED VARIANT: NONE / A / B / C MINIMAL REPAIR: State in at most 3 sentences what exact semantic mechanism should be added or clarified. CONFIDENCE: LOW / MEDIUM / HIGH IMPORTANT If a variant is safe only under an interpretation that is not explicitly forced by its text, mark that dependency explicitly. If two variants are semantically equivalent for all tested cases, say so rather than inventing a distinction. If a proposed repair closes one escape route but opens another, report both.