AI CareerOperating System

The saved gpt-oss:20b baseline had six failures: four sentence-limit failures and two missing expected signals. Seven routing-eligible questions passed 7/7 before routing.

Limit: These are preserved baseline failures from one local answer run, not a production benchmark or semantic-quality percentage.

Saved 43-case run; uncommitted local snapshot.

An OpenAI Decisions API layer using gpt-6-luna selected topics, depth and evidence needs before Responses API answer generation. Fixtures covered routed questions and direct-question bypasses.

Limit: Routing behavior does not establish answer quality or production readiness. This remained a locally tested prototype.

Routing version 2026-10-08.v1; saved result October 9 at 03:50 UTC.

Twelve questions ran three times per arm: 72 answers using gpt-5, gpt-6-luna routing and a source-aware gpt-5-mini judge. Helpfulness averaged 3.97/5 off and 4.28/5 on, while combined check-and-judge passes fell.

Limit: Routing alone was not an overall quality win. Concision and source-limit flags remained. Both arms used matched loaded code and sources; later edits are a different snapshot.

Stored routing-file fingerprint d0457ca5c44a; complete file fingerprints retained locally.

Answer processing kept qualifiers with proof, bounded sentence counts and preserved question scope. A separate ownership follow-up passed 8/8 trials across four models, twice each.

Limit: The 14-case smoke and eight ownership trials are different datasets. Targeted repair checks do not replace a complete regression after the final change.

Saved local smoke and ownership artifacts; not a committed release snapshot.

A linked skills bank registered 31 skill and tool labels, with supported labels connected to approved bullet variants and four mappings explicitly unresolved. Drafting and rendering checked proof visible in the resume. A synthetic PDF retained the Decisions API local-prototype skill and its qualifier.

Limit: Four mappings remained explicitly unresolved. Structural/provenance and rendering tests do not measure resume persuasiveness or establish deployment.

Local ResumeCustomizer commit 25892de; no model evaluation is included in this test count.

Six synthetic, public-safe scenarios ran once per routing arm using gpt-5.6-terra, Luna routing and a source-aware gpt-6.1-sol paired judge. The enabled ownership answer lost its AI-assisted implementation distinction during sentence bounding, even though the raw answer contained it. The 41 API requests price to about $0.57 from reported token/cache usage at published standard rates; the invoice was not independently reconciled.

Limit: Routing worked on this small authored set, but enabled answer quality was not a win. A deterministic repair replayed all 12 unchanged raw answers with byte-identical prompts and changed only that ownership answer. The original failed grade is retained; the repair has no new paid quality score. Complete current local validation, service compatibility and deployed configuration remain separate gates.

Canary snapshot 7371d60; original answers, grades, usage and the separate repair replay are retained locally.

A complete original 43-case pass at e2a72a4 used installed gpt-oss:20b answers and the same source-aware qwen2.5:14b judge. Two hard failures were wording or typography mismatches, but manual review still found a lost checkout estimate qualifier and AI Career measurement and release-status limits. Three deterministic probes exposed related output-processing gaps.

Limit: The two hard failures remain in the recorded 41/43 result. Judge passes did not override the manual findings, and the local models did not establish Terra or production quality. This snapshot remains part of the record after repairs.

Frozen e2a72a4 answers, original rubrics, grades and file hashes retained separately from later repairs.

Captured raw answers and supplied canonical evidence drove bounded repairs for estimate qualifiers, unspecified time units, exact static-inventory figures, unlaunched integration scope, empty assistance credit, adjacent media-outcome claims and retrieval typography. Focused tests covered missing evidence, already-qualified answers and unrelated measured runtime facts; four existing tests stayed skipped.

Limit: These are regression checks, not a semantic-quality score. The complete 43-case local answer and same-judge rerun is separate evidence and retains its original rubric. All prior comparisons, including the model bake-off and the paid enabled failure, remain preserved; no additional paid model score or production verification is claimed.

Repair and full-run manifests identify their own code and source snapshots; release remains held for review and owner configuration verification.

A complete follow-up at bcf571e retained all 43 original questions and rubrics. Manual review still found generic retrieval wording without its disabled-release limit, stronger survey agency and broader platform-delivery claims. The one judge failure alleged metrics absent from the displayed answer; separate source review found a real helped-lead qualifier to retain.

Limit: That combined result remains 41/43. A further source-gated repair passed 285 deterministic tests with four existing skips, including absent-source and already-qualified negatives. Its complete local rerun and actual-answer review are separate evidence; no old answer, grade, paid result or bake-off comparison is replaced.

Frozen bcf571e outputs and grades are retained. The later repair and full-run manifests identify their own snapshot; Production configuration remains unverified after the denied metadata read.

The complete 3cec994 run preserved one literal professional-grounding mismatch and one recommendation advisory. Source review found that approved career evidence already confirmed February 2026 platform production, while the Guide packet emphasized prototype framing. A conservative follow-up carried the timeline and original-vision caveat into the packet and retained helped-lead and worked-on agency.

Limit: The recorded combined result remains 42/43. Recommendation coverage alone is not a factual release blocker, and the source cross-check does not establish sole engineering ownership, preservation of the full original vision or stronger performance outcomes. The new snapshot receives its own complete local run; original scores, rubrics and the model bake-off remain unchanged.

Frozen 3cec994 answers and grades remain separate from the source-alignment revision and its later validation; Production configuration is still a separate gate.

The frozen a611b35 run completed all 43 original questions with gpt-oss:20b answers and qwen2.5:14b judging. Actual-answer review still found an incorrect 16-month timeline, unsupported measurable-business-outcome language and a missing helped-lead qualifier. Each claim originated in raw local generation and survived processing; the combined recorded result remains 38/43.

Limit: Four literal or coverage failures and one recommendation advisory remain separately recorded. These local failures are not evidence that Terra made the same claims. The earlier paid Terra canary retained its own 6/6 disabled and 5/6 enabled grades; the enabled ownership failure was a demonstrated processing loss of credit.

Original a611b35 questions, raw answers, displayed text, rubrics and grades remain immutable alongside every prior comparison and bake-off.

Daniel approved replacing expanding prose repairs with selection of approved claim-block IDs. The renderer keeps each source claim and its qualifications together, rejects invented IDs and free-form answer text, and uses a canonical fallback. The existing budget guard reconciles completed prior usage before admitting a small Terra/Luna regression comparison.

Limit: Deterministic contracts do not establish Terra answer quality or Production readiness. The comparison uses synthetic questions and public-safe approved evidence, with remote career retrieval disabled and no automatic paid retries. Its actual outcomes are recorded separately after completion.

All historical stages and the separate 72-trial model bake-off are retained. Owner Production configuration confirmation remains pending.

A new four-question comparison at 09db575 used eight Terra answer calls, four blinded Sol paired judgments and one Luna routing request shared across identical comparison inputs. It covered the three latest local claim failures and the original ownership-credit question. Terra selected valid approved block IDs in all four candidate answers, retaining their qualifications and implementation credit.

Limit: The candidate still failed relevance and concision: the judge preferred the baseline in three cases and the candidate in one. All three original page cases exceeded the unchanged four-sentence limit, and the impact answer missed one requested proof signal. Both ownership answers stayed on the current project despite the broader question. The small run established no universal quality gain or Production readiness; no automatic paid retry or regrade followed.

The a611b35 baseline code and 09db575 candidate used the same scoped public request per pair, with no remote career retrieval. Original rubrics, raw outputs, original paid 6/6-versus-5/6 grades and all historical bake-off evidence remain unchanged. Quality and owner Production configuration gates remain open.

A bounded local follow-up replaced broad bundles with concise authored claim-and-qualification blocks, pinned to the supplied approved source fields. Page catalogs now expose relevant proof blocks, role answers name the stated role, and across-portfolio ownership requires a supplied supporting project. The same four saved Terra ID outputs drove deterministic replay; no new model calls were made.

Limit: The original paid 2/4 candidate and 3/4 baseline judge results remain unchanged. Two saved selections use canonical fallback under the narrower catalog, so replay does not establish how Terra will select under the new schema or what a fresh judge would conclude. The spending-admission block and owner Production configuration gate remain open.

All paid raw answers, source/selection traces, original rubrics, earlier comparisons and the separate model bake-off are retained. New local checks are distinct from live semantic evidence.

Two new Terra selections used the concise claim catalog for role relevance and cross-portfolio ownership, with one Luna routing call and two Sol paired judgments against preserved baseline answers. Both selections were valid without canonical fallback. Complete qualifications stayed with the rendered proof, including prototype reuse intent, AI-assisted implementation credit, disabled retrieval and production-ownership limits.

Limit: Both baseline answers also passed. The judge preferred the baseline's more explicit role connection and the candidate's attributed cross-portfolio ownership comparison; the candidate was more concise in both cases. This is one observation per question against archived baselines, not a full 43-case Terra result or a universal quality gain. Impact and strategy retain replay-only coverage, and owner Production configuration remains unconfirmed.

Answer runtime 0f1ba7b with reviewed ledger correction be59b62; all five requests completed with known usage and no retries. Original four-question paid grades, source traces, historical stages and the separate model bake-off remain unchanged. The targeted semantic check passes; deployment remains a separate gate.