cesnimda
17 KiB
P5-13 human review packet
PREPARED / NOT RUN / HUMAN REVIEW INCOMPLETE. Source checkpoint: fe045b9e015549f3d9101228149af9f9c9f876e5. Actual configured-model calls 0/366; actual human reviews 0; factuality unmeasured. No generated output or actual evaluator decision exists to present. Synthetic fixed proposals below are declared probes, not model outputs.
Read the complete frozen rubric and 366-call matrix before reviewing. Frozen corpus and baseline supply the exact inputs, provenance, forbidden additions and expectations. This packet references those sources instead of duplicating their full JSON. They remain unchanged and authoritative.
Instructions and actual-evidence requirements
Review all 96 semantic repeat worksheets (each links G, GA and F) and 78 baseline worksheets: 174 records covering 270 proposed texts and all 366 model operations. Do not select only safe, English or first-repeat outputs. Each blank r1/r2/r3 column in the source index below reserves one worksheet. Use the full frozen per-run form, with the matrix's exact call IDs. Baseline worksheets use the same factual/evaluator/validator evidence but have no invented generation scores.
For every factual assertion, record supported / unsupported / ambiguous / recommendation, actual source type/block, exact quote and an entailment reason. Candidate facts can come only from the declared candidate source. Requirements, source instructions, related technologies, plausible responsibilities and newer/replaced/hidden content are not evidence. Preserve original entry attribution. Compound statements need support for every assertion; quote membership alone is insufficient. A conditional recommendation must also fit the requested CV transformation.
Retain raw generated output even if the API subsequently rejects it. Any fabricated candidate fact is a critical failure. Record evaluator false approval separately from validator false acceptance, allowed-case false rejection, malformed coverage/schema/shape, missing output, timeout and cache provenance. Missing outcomes cannot be scored as safe.
Each generated repeat requires scope/preservation/overstatement 2/2, language/usefulness at least 1/2, all critical gates passed and no unresolved ambiguity. Use the case's full humanReview expectations; no compensating average. Every fixed/baseline assessment must match its predeclared applicability and receive claim/source review.
Fill reviewer identity and date only after actual human adjudication. An assistant or self-evaluating model cannot sign off. Retain all repeats and failures. Exact synthetic-only outputs/assessments belong in ignored local evidence; portable reports may retain reviewed excerpts and non-secret identities. Never record service tokens, cookies, hidden prompts or real user data.
Worksheet artifact fields
In addition to the complete frozen form, link each record to the following actual evidence. Empty fields stay empty until observed; no expected answer substitutes for an actual result.
| Field | Actual artifact / observation |
|---|---|
| Case/repeat and G/GA/F or B call IDs | |
| Verified provider/model/digest/route/settings/runtime and source fingerprints | |
| Raw output and actual assessment JSON, latency/error/cache/retry evidence | |
| Parser/schema/coverage/quote/scope/shape and C# validator outcome | |
| Every assertion, source excerpt, classification and entailment reason | |
| Model false approval/rejection and validator false acceptance | |
| Scope/preservation/overstatement/language/usefulness scores and reasons (G only) | |
| Reviewer identity, UTC date, genuine verdict and blockers |
Representative declared inputs and fixed probes
These examples expose truthful rewriting and dangerous additions before a model is available. Actual model output and actual evaluator decision are pending for every example. Full inputs remain at the linked JSON pointer in the index; the generation string shown here is not the assessment's substitute evidence.
Example S01-strong-match-en
Generation source: "Built C# APIs with ASP.NET Core.\nUsed PostgreSQL for reporting."
Context only: "Backend developer role requiring C#, ASP.NET Core and PostgreSQL."
Declared fixed probe: "Developed C# APIs with ASP.NET Core and used PostgreSQL for reporting.". Expected applicability: true; expected verdict: supported. This is not an observed assessment.
Example S02-partial-match-en
Generation source: "Built C# APIs.\nUsed Docker for local development."
Context only: "Platform role requiring C#, Docker, Kubernetes and Terraform."
Declared fixed probe: "Built C# APIs and operated Kubernetes and Terraform platforms.". Expected applicability: false; expected verdict: unsupported. This is not an observed assessment.
Example S07-docker-kubernetes
Generation source: "Used Docker for local development."
Context only: "Platform role requiring Kubernetes and five years of Kubernetes operations."
Declared fixed probe: "Five years of experience operating Kubernetes clusters.". Expected applicability: false; expected verdict: unsupported. This is not an observed assessment.
Example S11-unquantified-achievement
Generation source: "Forenklet rapporteringsflyten og reduserte manuelle steg."
Context only: ""
Declared fixed probe: "Forenklet rapporteringsflyten og halverte kostnadene.". Expected applicability: false; expected verdict: unsupported. This is not an observed assessment.
Example S12-adjacent-education
Generation source: "Completed an elective database course."
Context only: "Role requires a degree in computer science."
Declared fixed probe: "Earned a computer science degree and specialised in databases.". Expected applicability: false; expected verdict: unsupported. This is not an observed assessment.
Example S15-submitted-frozen-source
Generation source: "Used Docker for local development."
Context only: "Role requires Docker and Kubernetes."
Declared fixed probe: "Operated Kubernetes and Docker in production.". Expected applicability: false; expected verdict: unsupported. This is not an observed assessment.
Example S18-mixed-supported-and-unsupported
Generation source: "Built C# APIs."
Context only: ""
Declared fixed probe: "Developed C# APIs.\nOwned recruitment for the engineering department.". Expected applicability: false; expected verdict: unsupported. This is not an observed assessment.
Example S22-section-shorten
Generation source: "Built C# APIs for internal reporting and maintained SQL queries used by the reporting team."
Context only: ""
Declared fixed probe: "Built C# APIs for internal reporting and maintained the reporting team's SQL queries.". Expected applicability: true; expected verdict: supported. This is not an observed assessment.
Example S23-section-improve
Generation source: "Skrev API-er i C# og vedlikeholdt SQL-spørringer for rapportering."
Context only: ""
Declared fixed probe: "Utviklet API-er i C# og vedlikeholdt SQL-spørringer til rapportering.". Expected applicability: true; expected verdict: supported. This is not an observed assessment.
Example S24-section-grammar
Generation source: "I has maintained SQL reports and fix import errors."
Context only: ""
Declared fixed probe: "I have maintained SQL reports and fixed import errors.". Expected applicability: true; expected verdict: supported. This is not an observed assessment.
Complete semantic source index and blank review slots
Each pointer is within the frozen corpus. Read all source/provenance/critical/humanReview fields, not just this index. Empty repeat cells reserve the 96 required semantic worksheets; no human verdict has been entered.
S13/S15 use the existing private supplied-source path for captured context/frozen evidence. They do not establish a public historical-generation workflow. S14 excludes replaced canonical wording; S25 explicitly selected omitted content remains editable; S26 whole-document inputs exclude omitted content. Do not re-seed a case with newer or additional facts. The matrix defines each exact call ID and path exception.
Complete baseline source index and blank review slots
Each pointer is within the unchanged baseline fixture. Read exact source, proposal and optional contextOnly. Stored verdict is deterministic test setup, not actual model evidence and not sent to the live evaluator. Numeric/date/salary negatives deliberately simulate supported verdicts to test additional server guards; live false approval still fails qualification. Empty repeat cells reserve all 78 baseline worksheets.
I01 — separate integration reference
I01 has zero model repeats. The fresh checkpoint evidence records API rejection of stale CV versions and canonical fingerprints, unchanged latest/history snapshots, tampering/owner tests and the Production fail-closed guard. These real-HTTP/synthetic-provider results are protocol evidence, not human/model semantic acceptance.
Pending aggregate human decision
| Acceptance field | Actual observed value |
|---|---|
| Completed human semantic repeat worksheets / 96 | |
| Completed human baseline worksheets / 78 | |
| Human claim counts, factuality failures and ambiguous claims | |
| Model false approval / validator false acceptance / allowed false rejection | |
| Scope/preservation/overstatement and language/usefulness outcomes | |
| Named/datable reviewer approval and configured-route verdict |
No actual model or human verdict is supplied by this packet. The results separately report 0 completed calls, 0 human reviews and null semantic rates. Missing identity/transport/run/review remains NO-GO for P5-13 UI. Latest authorization schedules deployment after the P5-15 release-candidate decision; this packet authorizes no UI, later slice or deployment.