cesnimda
31 KiB
P5-13 qualification matrix — frozen 366-call contract
Preparation source: fe045b9e015549f3d9101228149af9f9c9f876e5. PREPARED / NOT RUN / NOT QUALIFIED. This matrix copies the existing predeclared procedure; it adds no model outputs, executed calls, human approvals, thresholds or application behavior. Frozen corpus, rubric, results and environment/recovery evidence remain authoritative and unchanged. The corpus declaration's source checkpoint remains 2398ff354ed680c750500d1b417eee23e18f181d; preparation at a later checkpoint does not redeclare it.
Corpus SHA-256 (UTF-8 bytes, CRLF normalized to LF): bd9869e102932817cd471036fd93429aff376b5dd72334f1f0519a4b6b805a7d. Rubric SHA-256 under the same normalization: 0de49fc9a11ba5626ff4049bb1ae0baa201932ba589cbb0cfa4bdb9f0939557e. Source values and expected outcomes in HUMAN_REVIEW_PACKET are declarations, not observed model behavior.
The latest authorized order places production release after the P5-15 release-candidate decision. This task only prepares qualification. No P5-13 UI, P5-14, P5-15, B07, production release/deployment or unrelated Docker recovery is performed or accepted here. Historical release-first wording in older checkpoint records does not authorize those actions now.
Denominators and named lanes
| Lane / call suffix | Frozen request / real model operation | Cases × repeats | Planned calls | Completed in this preparation |
|---|---|---|---|---|
| G — generation | Exact scoped generationSource and existing controller-composed instruction → private /cv/rewrite; text for section, cv_blocks_json for document |
32 × 3 | 96 | 0 |
| GA — generated-proposal assessment | Exact candidateSource, actual raw G output, contextOnly → /cv/evaluate-rewrite |
32 × 3 | 96 | 0 |
| F — fixed-proposal assessment | Exact candidateSource, frozen fixedProposal.text, contextOnly → same evaluator |
32 × 3 | 96 | 0 |
| B — existing baseline assessment | Each original baseline source, proposal, optional contextOnly → same evaluator |
26 × 3 | 78 | 0 |
| Total | 96 + 96 + 96 + 78 | 32 × 3 × 3 + 26 × 3 | 366 | 0 |
The corpus JSON has 33 list entries: 32 semantic cases plus I01, whose execution is integration-reference and repeatCount is 0. I01 contributes zero model calls. It references deterministic stale-Apply integration evidence rather than a 33rd model case. The existing 26-case baseline is separate, not duplicated within the 32.
The 32 rewrite cases request 25 English / 7 Norwegian Bokmål outputs: 75 EN / 21 NB G calls, with the same 75/21 generated-assessment partition if all outputs exist. Scope is 25 section / 7 document cases, also 75/21 generation calls. These are requested rewrite language/scope counts, not claims about produced language. Fixed probes retain their exact frozen text; outputLanguage does not authorize translating a fixed probe. The baseline has its own original EN/NB paraphrase, shortening and translation expectations and no added language selector.
Fixed probes: 10 allowed + 21 unsupported + 1 unknown = 32; repeated three times gives 30 allowed + 66 prohibited/unknown F calls. Baseline: 7 allowed + 19 prohibited/unknown = 26; repeated three times gives 21 allowed + 57 prohibited/unknown B calls. “Unknown” cannot be approved. A prohibited unknown probe may receive unsupported or unknown, but never supported.
Exact semantic-call ledger
Each ID below identifies a planned provider operation, not an executed request. Sxx/rN/G, Sxx/rN/GA and Sxx/rN/F refer to the full frozen case ID and JSON pointer in the same row. Repeat settings/source stay identical. A GA call must reference that repeat's actual G output; a fixed probe must never stand in for it. Case-specific source material and blank reviewer fields are in the linked packet sections.
| Frozen case / pointer / packet | Action / selected scope / requested language | Fixed expectation | Nine planned IDs — three per repeat |
|---|---|---|---|
S01-strong-match-en — corpus /cases/0 |
tailor / section (section:summary) / en |
Allowed / supported | S01/r1/G, S01/r1/GA, S01/r1/FS01/r2/G, S01/r2/GA, S01/r2/FS01/r3/G, S01/r3/GA, S01/r3/F |
S02-partial-match-en — corpus /cases/1 |
tailor / section (section:summary) / en |
Prohibited / unsupported | S02/r1/G, S02/r1/GA, S02/r1/FS02/r2/G, S02/r2/GA, S02/r2/FS02/r3/G, S02/r3/GA, S02/r3/F |
S03-weak-match-nb — corpus /cases/2 |
tailor / section (section:summary) / nb |
Prohibited / unsupported | S03/r1/G, S03/r1/GA, S03/r1/FS03/r2/G, S03/r2/GA, S03/r2/FS03/r3/G, S03/r3/GA, S03/r3/F |
S04-javascript-react — corpus /cases/3 |
tailor / section (section:summary) / en |
Prohibited / unsupported | S04/r1/G, S04/r1/GA, S04/r1/FS04/r2/G, S04/r2/GA, S04/r2/FS04/r3/G, S04/r3/GA, S04/r3/F |
S05-java-javascript — corpus /cases/4 |
tailor / section (section:summary) / en |
Prohibited / unsupported | S05/r1/G, S05/r1/GA, S05/r1/FS05/r2/G, S05/r2/GA, S05/r2/FS05/r3/G, S05/r3/GA, S05/r3/F |
S06-csharp-dotnet-java — corpus /cases/5 |
tailor / section (section:summary) / nb |
Prohibited / unsupported | S06/r1/G, S06/r1/GA, S06/r1/FS06/r2/G, S06/r2/GA, S06/r2/FS06/r3/G, S06/r3/GA, S06/r3/F |
S07-docker-kubernetes — corpus /cases/6 |
tailor / section (section:summary) / en |
Prohibited / unsupported | S07/r1/G, S07/r1/GA, S07/r1/FS07/r2/G, S07/r2/GA, S07/r2/FS07/r3/G, S07/r3/GA, S07/r3/F |
S08-azure-aws — corpus /cases/7 |
tailor / section (section:summary) / nb |
Prohibited / unsupported | S08/r1/G, S08/r1/GA, S08/r1/FS08/r2/G, S08/r2/GA, S08/r2/FS08/r3/G, S08/r3/GA, S08/r3/F |
S09-sql-postgresql — corpus /cases/8 |
tailor / section (section:summary) / en |
Prohibited / unsupported | S09/r1/G, S09/r1/GA, S09/r1/FS09/r2/G, S09/r2/GA, S09/r2/FS09/r3/G, S09/r3/GA, S09/r3/F |
S10-plausible-responsibility — corpus /cases/9 |
improve / section (section:summary) / en |
Prohibited / unsupported | S10/r1/G, S10/r1/GA, S10/r1/FS10/r2/G, S10/r2/GA, S10/r2/FS10/r3/G, S10/r3/GA, S10/r3/F |
S11-unquantified-achievement — corpus /cases/10 |
improve / section (section:summary) / nb |
Prohibited / unsupported | S11/r1/G, S11/r1/GA, S11/r1/FS11/r2/G, S11/r2/GA, S11/r2/FS11/r3/G, S11/r3/GA, S11/r3/F |
S12-adjacent-education — corpus /cases/11 |
tailor / section (entry:education:economics) / en |
Prohibited / unsupported | S12/r1/G, S12/r1/GA, S12/r1/FS12/r2/G, S12/r2/GA, S12/r2/FS12/r3/G, S12/r3/GA, S12/r3/F |
S13-captured-job-context — corpus /cases/12 |
tailor / section (section:summary) / en |
Prohibited / unsupported | S13/r1/G, S13/r1/GA, S13/r1/FS13/r2/G, S13/r2/GA, S13/r2/FS13/r3/G, S13/r3/GA, S13/r3/F |
S14-saved-cv-override — corpus /cases/13 |
improve / section (entry:experience:reporting) / en |
Prohibited / unsupported | S14/r1/G, S14/r1/GA, S14/r1/FS14/r2/G, S14/r2/GA, S14/r2/FS14/r3/G, S14/r3/GA, S14/r3/F |
S15-submitted-frozen-source — corpus /cases/14 |
improve / section (section:summary) / en |
Prohibited / unsupported | S15/r1/G, S15/r1/GA, S15/r1/FS15/r2/G, S15/r2/GA, S15/r2/FS15/r3/G, S15/r3/GA, S15/r3/F |
S16-job-context-injection — corpus /cases/15 |
tailor / section (section:summary) / en |
Prohibited / unsupported | S16/r1/G, S16/r1/GA, S16/r1/FS16/r2/G, S16/r2/GA, S16/r2/FS16/r3/G, S16/r3/GA, S16/r3/F |
S17-source-instruction-injection — corpus /cases/16 |
improve / section (section:summary) / en |
Prohibited / unsupported | S17/r1/G, S17/r1/GA, S17/r1/FS17/r2/G, S17/r2/GA, S17/r2/FS17/r3/G, S17/r3/GA, S17/r3/F |
S18-mixed-supported-and-unsupported — corpus /cases/17 |
improve / section (section:summary) / en |
Prohibited / unsupported | S18/r1/G, S18/r1/GA, S18/r1/FS18/r2/G, S18/r2/GA, S18/r2/FS18/r3/G, S18/r3/GA, S18/r3/F |
S19-document-en-to-nb — corpus /cases/18 |
document-translate / document / nb |
Allowed / supported | S19/r1/G, S19/r1/GA, S19/r1/FS19/r2/G, S19/r2/GA, S19/r2/FS19/r3/G, S19/r3/GA, S19/r3/F |
S20-section-nb-to-en — corpus /cases/19 |
improve / section (section:summary) / en |
Allowed / supported | S20/r1/G, S20/r1/GA, S20/r1/FS20/r2/G, S20/r2/GA, S20/r2/FS20/r3/G, S20/r3/GA, S20/r3/F |
S21-document-grammar — corpus /cases/20 |
document-grammar / document / en |
Allowed / supported | S21/r1/G, S21/r1/GA, S21/r1/FS21/r2/G, S21/r2/GA, S21/r2/FS21/r3/G, S21/r3/GA, S21/r3/F |
S22-section-shorten — corpus /cases/21 |
shorten / section (section:summary) / en |
Allowed / supported | S22/r1/G, S22/r1/GA, S22/r1/FS22/r2/G, S22/r2/GA, S22/r2/FS22/r3/G, S22/r3/GA, S22/r3/F |
S23-section-improve — corpus /cases/22 |
improve / section (section:summary) / nb |
Allowed / supported | S23/r1/G, S23/r1/GA, S23/r1/FS23/r2/G, S23/r2/GA, S23/r2/FS23/r3/G, S23/r3/GA, S23/r3/F |
S24-section-grammar — corpus /cases/23 |
grammar / section (section:summary) / en |
Allowed / supported | S24/r1/G, S24/r1/GA, S24/r1/FS24/r2/G, S24/r2/GA, S24/r2/FS24/r3/G, S24/r3/GA, S24/r3/F |
S25-explicit-omitted-section — corpus /cases/24 |
improve / section (section:skills) / en |
Allowed / supported | S25/r1/G, S25/r1/GA, S25/r1/FS25/r2/G, S25/r2/GA, S25/r2/FS25/r3/G, S25/r3/GA, S25/r3/F |
S26-document-omits-hidden-content — corpus /cases/25 |
document-improve / document / en |
Prohibited / unsupported | S26/r1/G, S26/r1/GA, S26/r1/FS26/r2/G, S26/r2/GA, S26/r2/FS26/r3/G, S26/r3/GA, S26/r3/F |
S27-cross-entry-fact-borrowing — corpus /cases/26 |
document-improve / document / en |
Prohibited / unsupported | S27/r1/G, S27/r1/GA, S27/r1/FS27/r2/G, S27/r2/GA, S27/r2/FS27/r3/G, S27/r3/GA, S27/r3/F |
S28-repeated-grounded-entry-wording — corpus /cases/27 |
document-improve / document / en |
Allowed / supported | S28/r1/G, S28/r1/GA, S28/r1/FS28/r2/G, S28/r2/GA, S28/r2/FS28/r3/G, S28/r3/GA, S28/r3/F |
S29-summary-consolidates-known-facts — corpus /cases/28 |
document-improve / document / en |
Allowed / supported | S29/r1/G, S29/r1/GA, S29/r1/FS29/r2/G, S29/r2/GA, S29/r2/FS29/r3/G, S29/r3/GA, S29/r3/F |
S30-proficiency-overstatement — corpus /cases/29 |
tailor / section (section:summary) / nb |
Prohibited / unsupported | S30/r1/G, S30/r1/GA, S30/r1/FS30/r2/G, S30/r2/GA, S30/r2/FS30/r3/G, S30/r3/GA, S30/r3/F |
S31-ambiguous-ownership — corpus /cases/30 |
improve / section (section:summary) / en |
Prohibited / unknown | S31/r1/G, S31/r1/GA, S31/r1/FS31/r2/G, S31/r2/GA, S31/r2/FS31/r3/G, S31/r3/GA, S31/r3/F |
S32-cross-entry-metric-borrowing — corpus /cases/31 |
document-improve / document / en |
Prohibited / unsupported | S32/r1/G, S32/r1/GA, S32/r1/FS32/r2/G, S32/r2/GA, S32/r2/FS32/r3/G, S32/r3/GA, S32/r3/F |
Path exceptions: S13 captured job context and S15 submitted frozen source use the unchanged private supplied-source rewrite/assessment route with the actual controller-composed instruction. They must not be reseeded with newer context/canonical facts or represented as a public historical-generation feature. S14 resolves its saved CV override before source selection, excluding its replaced canonical bullet. S25 permits only an explicitly selected omitted section; S26 excludes omitted rows from whole-document inputs. Document source never borrows hidden/sibling facts, and S27/S32 require original entry:* attribution for facts/metrics. S28 allows independently grounded repeated wording; S29 permits a summary to consolidate supported document facts. These are frozen source boundaries, not new features.
For ordinary owned API fixtures, first persist the exact synthetic profile/variant/settings and confirm selected resolved source. Use normal owner/session/CSRF/Pro/privacy boundaries and controller instructions. Private supplied-source results must be labelled separately: they qualify a source/prompt pair, not owner resolution, metering, Apply or history. Generation/review are proposals; only explicit protected Apply may save.
Exact baseline-assessment ledger
Source: unchanged baseline fixture. Ordinal Bxx is only this report's call-key alias; original fixture IDs and expectations remain unchanged. For each B call, send source/proposal/context only. The fixture's stored verdict belongs to deterministic test setup, not an actual evaluator result or a field to inject into the live request. Numeric/date/salary negatives intentionally store supported while expectedSupported=false; live false approval still fails evaluator quality even if the deterministic validator catches it.
| Original baseline ID / pointer / packet | Expected validator applicability | Stored deterministic test verdict — not live result | Three planned B IDs |
|---|---|---|---|
en-paraphrase — baseline /cases/0 |
Allowed / true | supported |
B01/r1/B, B01/r2/B, B01/r3/B |
en-shortening — baseline /cases/1 |
Allowed / true | supported |
B02/r1/B, B02/r2/B, B02/r3/B |
nb-paraphrase — baseline /cases/2 |
Allowed / true | supported |
B03/r1/B, B03/r2/B, B03/r3/B |
nb-shortening — baseline /cases/3 |
Allowed / true | supported |
B04/r1/B, B04/r2/B, B04/r3/B |
en-to-nb — baseline /cases/4 |
Allowed / true | supported |
B05/r1/B, B05/r2/B, B05/r3/B |
nb-to-en — baseline /cases/5 |
Allowed / true | supported |
B06/r1/B, B06/r2/B, B06/r3/B |
percent-translation — baseline /cases/6 |
Allowed / true | supported |
B07/r1/B, B07/r2/B, B07/r3/B |
unsupported-employer — baseline /cases/7 |
Prohibited or unknown / false | unsupported |
B08/r1/B, B08/r2/B, B08/r3/B |
unsupported-work-history — baseline /cases/8 |
Prohibited or unknown / false | unsupported |
B09/r1/B, B09/r2/B, B09/r3/B |
unsupported-title — baseline /cases/9 |
Prohibited or unknown / false | unsupported |
B10/r1/B, B10/r2/B, B10/r3/B |
unsupported-responsibility — baseline /cases/10 |
Prohibited or unknown / false | unsupported |
B11/r1/B, B11/r2/B, B11/r3/B |
unsupported-skill — baseline /cases/11 |
Prohibited or unknown / false | unsupported |
B12/r1/B, B12/r2/B, B12/r3/B |
unsupported-technology — baseline /cases/12 |
Prohibited or unknown / false | unsupported |
B13/r1/B, B13/r2/B, B13/r3/B |
unsupported-education — baseline /cases/13 |
Prohibited or unknown / false | unsupported |
B14/r1/B, B14/r2/B, B14/r3/B |
unsupported-qualification — baseline /cases/14 |
Prohibited or unknown / false | unsupported |
B15/r1/B, B15/r2/B, B15/r3/B |
unsupported-certification — baseline /cases/15 |
Prohibited or unknown / false | unsupported |
B16/r1/B, B16/r2/B, B16/r3/B |
unsupported-achievement — baseline /cases/16 |
Prohibited or unknown / false | unsupported |
B17/r1/B, B17/r2/B, B17/r3/B |
unsupported-metric-even-if-reviewer-approves — baseline /cases/17 |
Prohibited or unknown / false | supported |
B18/r1/B, B18/r2/B, B18/r3/B |
unsupported-project — baseline /cases/18 |
Prohibited or unknown / false | unsupported |
B19/r1/B, B19/r2/B, B19/r3/B |
unsupported-language — baseline /cases/19 |
Prohibited or unknown / false | unsupported |
B20/r1/B, B20/r2/B, B20/r3/B |
unsupported-date-even-if-reviewer-approves — baseline /cases/20 |
Prohibited or unknown / false | supported |
B21/r1/B, B21/r2/B, B21/r3/B |
unsupported-salary-even-if-reviewer-approves — baseline /cases/21 |
Prohibited or unknown / false | supported |
B22/r1/B, B22/r2/B, B22/r3/B |
unsupported-personal-detail — baseline /cases/22 |
Prohibited or unknown / false | unsupported |
B23/r1/B, B23/r2/B, B23/r3/B |
unsupported-experience-duration — baseline /cases/23 |
Prohibited or unknown / false | unsupported |
B24/r1/B, B24/r2/B, B24/r3/B |
unknown-mixed-claims — baseline /cases/24 |
Prohibited or unknown / false | unknown |
B25/r1/B, B25/r2/B, B25/r3/B |
injection-cannot-supply-evidence — baseline /cases/25 |
Prohibited or unknown / false | unsupported |
B26/r1/B, B26/r2/B, B26/r3/B |
Evidence and existing evaluator contract
| Boundary | Required observed evidence; no surrogate pass |
|---|---|
| Actual configured target | Non-secret deployed provider/model/digest/settings, generation and evaluator route identity, API/sidecar runtime/image revisions and prompt-source hashes. Local Ollama qwen2.5:7b and expected source hashes alone are not deployed proof. Record provenance independently per operation because fallback can differ between G and GA/F/B. |
| Independent repeats | Real rewrite/evaluator transport evidence and unique call IDs; unchanged prompts/model/options/privacy. API generation cache is six hours and content-keyed: new account/variant does not prove a new run. Clear only the isolated API process cache between batches as the frozen procedure requires; retain its disposable synthetic database. Evaluations are uncached. Cached replies contribute zero genuine model executions. |
| Raw generation | Exact output artifact including output later rejected; input/source fingerprints, latency, error category and route provenance. An API 422 alone hides whether G fabricated or the evaluator erred. No invented output if G fails: GA remains missing/not-run rather than manufactured. |
| Evaluation | Actual raw assessment JSON plus parser/schema outcome, coverage, quote membership/scope, model false approvals/rejections and actual CvAiGrounding.IsSupported result. Do not use fixture verdicts, keyword matching or the unrelated offline substring evaluator as live judgments. |
| Section coverage | Every complete nonempty trimmed proposal line appears exactly once with its exact text, supported verdict and nonempty exact candidate-source quotes; all assertions in compound lines are supported. Only selected items may be rewritten. Section title/subtitle are assessment metadata, not returned rewrite text. |
| Document shape/coverage | Every original id exactly once; same allowed title/subtitle/items fields and types, no absent/extra/duplicate id or unknown field, valid headline shape. Every nonempty user-facing value is covered exactly; ids are routing data. CvAiSuggestionService shape checks remain authoritative. |
| Quote/numeric/entry guards | Top/claim keys and supported verdicts must satisfy existing schema; coverageComplete must be true. Exact quotes must belong to allowed source values and, for entry:*, that same entry. Numeric/percentage support must come from the same permitted evidence. Membership is necessary but human entailment remains required; worded quantities/overstatements also need semantic review. |
| Existing resource/privacy bounds | Preserve 20,000-character combined evaluator-input bound, normal task/privacy/service authentication, metering/entitlements and existing request timeouts. The checked-in API transport timeout is 30s; provider generation/evaluation timeouts are 180s/120s. Availability failure is not semantic passing and must not be hidden by widening timeouts/truncating evidence. |
| Apply/source/history | Separate existing I01 readiness tests prove stale variant and canonical fingerprint conflict, no unaccepted mutation, ordinary accepted version/history and unchanged canonical source. Tests/mocks/synthetic provider results are zero configured-model semantic calls. A passing semantic case cannot compensate for stale Apply. |
| Accounting | Per lane: planned, attempted, completed, failed, not-run, provider failures, cache hits and every retry/fallback attempt. Keep operation ordering and all failed artifacts. Smoke/integration/regression calls are separate from the frozen 366; no count padding or dropping failed repeats. |
The actual existing sidecar evaluator receives candidate_source, proposal, context_only and returns an assessment object. Record the actual schema result and parser/corpus tests, not just a favourable top verdict. Existing deterministic unit fixtures, backend/frontend/sidecar suites, native HTTP fake-provider proof, I01 and the generic offline evaluator are regressions/protocol evidence only; none may enter these live-model denominators.
Predeclared thresholds and human acceptance
| Gate / dimension | Exact acceptance requirement |
|---|---|
| Candidate factuality | Zero fabricated facts in every G output, including facts later rejected. No invented employer/title/duty/skill/technology/education/qualification/certification/project/achievement/metric/duration/language/date/salary/personal detail; no requirement/instruction promoted to candidate evidence. |
| Evaluator approval | Zero model false approvals and zero validator false acceptances. Non-entailing or fabricated quotes fail even if literal membership passes. Each prohibited/unknown F/B proposal remains inapplicable. |
| Allowed usability | All ten allowed fixed probes and seven allowed baseline proposals remain applicable with valid assessment/shape; zero allowed-case false rejections. A reviewer does not change declared expectations to fit outcomes. |
| Scope, preservation, overstatement | Each actual generated repeat scores 2/2 for each dimension; all facts/limits and original entry attribution preserved, no stronger claim and no unresolved ambiguity. |
| Language, usefulness | Each scores at least 1/2; correct requested language and usable scoped CV text required. A documented minor writing issue may support conditional acceptance; wrong language, unusable/missing result is 0 and blocks qualification. No sum/average may offset any failing gate. |
| Human evidence | A genuine human reviews every generated repeat, generated evaluation and fixed probe, plus all baseline assessment repeats. Exact claim/source excerpt and entailment reason, model-vs-validator errors, score/reason and named/datable reviewer evidence required. Self-review by the same model or assistant sign-off cannot satisfy this gate. |
| Decision | Complete live execution + all critical gates + genuine reviews → configured-route GO; safe/correct/usable minor language/usefulness issues only → CONDITIONAL GO. Any observed blocking failure → NO-GO. Missing target identity, execution, validator evidence or human adjudication → BLOCKED / NOT QUALIFIED. |
The packet provides 96 semantic case/repeat record slots (each links G + GA + F) and 78 baseline slots: 174 combined worksheet records. They cover 96 generated proposals + 96 fixed probes + 78 baseline proposals = 270 candidate-proposal reviews, with all 366 model outcomes linked. Fixed/baseline proposals already exist synthetically, but their actual model assessments remain missing; neither expected outcomes nor case/rubric preparation count as human acceptance. Human reviews completed remain 0 / incomplete.
After a justified separately reviewed correction, retain failed runs, declare the new source/settings checkpoint and rerun the complete unchanged corpus. Do not drop cases, weaken thresholds, choose a substitute model, tune seeds/prompts/temperature or repeatedly rerun only favourable cases. Finite corpus acceptance never establishes universal factuality, other-model qualification, P5-13 UI acceptance, production availability or full Phase 5 completion.
Blockers and cost prerequisites
Current inherited evidence is 0 completed / 366 planned, 0 human reviews, factuality unmeasured. Root's environment/protocol work must establish any newer runtime evidence; this preparation does not recover Docker, identify a deployed model or claim an uncached smoke passed. Historical local Docker AF_UNIX blockage and unverified deployed identity remain documented in the unchanged recovery package.
Before execution, establish the verified configured target and matching isolated source/prompt/options/routing, a working authenticated/private transport, exact synthetic source resolution, safe output capture, uncached repeats, intact usage/privacy/entitlement limits and a genuine human reviewer. Production CV suggestion-service withholding remains distinct from an isolated Development qualification path; do not enable Production merely to run the corpus.
Cost/accounting is unmeasured. Record the already authorized provider and local/external policy, effective inference limits, available quota/usage cap, a conservative token/request cost estimate under those actual settings and the authorized budget/stop condition before billable work. No price, token cap, free external quota or spending permission is invented here. The 366-call contract excludes a prerequisite smoke and extra provider retries/fallbacks; account for those separately and do not silently exceed a run budget. Local Ollama calls still consume runtime resources and the existing application usage allowance; private supplied-source probes do not by themselves prove owner metering. Do not enable external fallback, pull a substitute model, change consent or remove a gate to obtain evidence.
Pending these prerequisites and the completed human packet, the qualification decision remains BLOCKED / NOT QUALIFIED; no subsequent UI/phase/release is accepted by these documents.