All MicroEvals
A.P.12 — LMC AUDITOR-CONTROL VALIDATION: v27 FILE LABELS / C...
Create MicroEval
Header image for A.P.12 — LMC AUDITOR-CONTROL VALIDATION: v27 FILE LABELS / C...

A.P.12 — LMC AUDITOR-CONTROL VALIDATION: v27 FILE LABELS / C...

Prompt

A.P.12 — LMC AUDITOR-CONTROL VALIDATION: v27 FILE LABELS / CONSOLIDATED MANIFEST REGRESSION ROLE Independent auditor of the prompt-evaluation control plane. Audit only the exact canonical v27 bundle supplied with this prompt. Do not reconstruct missing artifacts from conversation history. WEB CAPABILITY Competitor panel models have NO Web Search. Do not require, recommend, or simulate Web Search for panel models. Central Web Search, if actually available, is auditor/controller calibration only. HARD INTAKE Required atomic bundle: (1) this exact A.P.12 TXT; (2) Metodika 27 exact DOCX; (3) Změny 27 exact DOCX containing consolidated identity/manifest data; (4) exact current production prompt; (5) exact Frozen baseline; (6) relevant prior release/acceptance rules; (7) the expected-next-prompt inventory and exact A/B TXT artifacts; (8) bundle identity data. Missing target/dependency => BLOCKED. Never reconstruct. PRIMARY REGRESSION TARGET Previous releases used a standalone LMC_v26_release_manifest.txt. v27 changes the default so manifest/identity data live in Změny x. The test must determine whether this consolidation preserves determinism, readability, machine-processing sufficiency and attachment truth. MANDATORY TEST GROUPS A. Czech aliases: A x, B x, Metodika x, Změny x map one-to-one to canonical artifacts. B. Alias/canonical bidirectional identity and no free-form naming. C. Consolidated manifest completeness inside Změny x. D. No standalone manifest by default; H195 exception only when explicit machine-processing necessity is recorded. E. Next-prompt attachment truth: A/B aliases must correspond to actual attached TXT artifacts, not links or internal paths. F. Negative target reconstruction: target absent + detailed H-rules/scope pointers => zero target claims/findings/patches. G. Portability: all next prompts are UTF-8 TXT and <=15,000 chars; competitor Web Search remains outside capability. H. Machine-readable manifest block: deterministic parsing must work without parsing the narrative text. TEN-SLOT PANEL Exactly 10 predefined slots. RESPONSE_NOT_AVAILABLE is a legitimate delivered result. NOT_RUN only when explicitly observed. Do not infer NOT_RUN from absent records. Report coverage n/10. n<10 alone does not trigger rerun. ADVERSARIAL FIXTURES 1. Complete v27 bundle with A/B/Metodika/Změny aliases and hashes matching -> completion may proceed. 2. Změny x contains alias but corresponding A or B TXT is absent -> attachment incomplete. 3. Standalone manifest is present but Změny x omits its identity/reason -> incomplete under H195. 4. Two artifacts share the same Czech alias -> reject identity mapping. 5. Canonical filename changes while alias remains -> hash/version mismatch must be detected. 6. Změny x is present but identity data are only narrative prose -> machine-readable manifest requirement not met. 7. Target absent but detailed H-rules exist in prompt -> no reconstruction. 8. 8/10 model outputs -> report 8/10, no automatic rerun solely for n<10. 9. One RESPONSE_NOT_AVAILABLE -> preserve status exactly. 10. Discovery proposal requires Web Search -> inadmissible for competitor-panel mechanism work. DISCOVERY At most 3 hypotheses. Do not generate target hypotheses without observable target content. Process/harness candidates must be clearly outside target scope and novelty-checked against v27. IVR Every material repair candidate requires an IVR object before promotion. Self-test PASS is not independent acceptance. Do not patch v27 from a blocked target audit. OUTPUT — SINGLE PASS 1. Executive verdict + one dominant NEXT ACTION. 2. Bundle identity / aliases / provenance. 3. Web capability. 4. Ten-slot panel matrix. 5. A-H test results. 6. Findings: target vs process/harness. 7. Discovery hypotheses. 8. IVR status. 9. Adversarial scenarios. 10. Independent validation. 11. Convergence/release risk. 12. Exact next action. 13. Minimal patch / no-op. CRITICAL STOP RULE If the canonical v27 bundle is incomplete, do not issue target PASS/FAIL and do not claim the absence of findings as evidence of a clean v27. FINAL INVARIANT Default v27 identity source = Změny x. Standalone release manifest is exceptional, never assumed. If EXPECTED-NEXT-PROMPT SET is non-empty, every expected A/B artifact must be actually attached and attachment-verified.