0 Fork 0

Files

T

dennisthiessenandClaude Opus 5 ce8619b8e3 fix(F14): call keyword relevance what it is, and teach it German

The scout presented a keyword count as "Strong fit" / "Medium fit". It is
neither: it measures how many profile keywords a posting echoes, and knows
nothing about his evidence, the hard requirements, the level, the language or
the commute.

- Buckets renamed to keyword relevance, with "NOT a fit judgement" in the
  heading, and score_job documents the limit with the review's own example:
  "Senior AI Engineer, Generative AI / LLM / RAG" scores 12 against "Senior Data
  Engineer" at 3, while the second is much closer to what he can evidence.
- German job vocabulary added. Every German-language posting used to score near
  zero regardless of fit, which is the real reason RUAG, SBB, BKW, PostFinance
  and Kdo Cy all needed a _score_floor. The floor was a workaround; the missing
  terms were the defect. "Senior Dateningenieur" 1 -> 3, level with its English
  twin.
- Platform/DevOps vocabulary was simply absent: devsecops, platform engineer,
  platform automation, infrastructure as code, observability, site reliability.
  The Kdo Cy "ICT-System Engineer (DevSecOps / Platform Automation)" scored ZERO
  before this - a posting whose requirement lines match his canonical evidence
  almost line for line. It now scores 10.
- Noise is unaffected: "Office Manager:in Administration" still scores 0.

Floors were left in place so this diff stays attributable; they can likely be
reduced now that the vocabulary exists.

34/34 tests pass; validated on a live 37-company run.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qej5SBvhcbGXHwLqNWiskM

2026-09-11 19:54:17 +02:00

8.3 KiB

Raw Blame History

Review remediation — 2026-09-11

User authorized all recommendations in output/Project_Review_2026-09-11/project_review.tex. Work sequentially; update this file after each verified fix and before stopping. Do not rewrite submitted application archives or invent missing evidence.

Resume point

F01–F14 complete. Next: F15 — typed dates and retained contact/compensation metadata (job_scout/scout.py: the Taleo parser writes a deadline into posted, and the Prospective path's unit/Lohnklasse/contact are folded into description, which _process_company discards). Read this file and inspect git log/git diff before resuming; do not repeat completed work.

Checkpoints: afdef7e (F01+F02) · fc5f3e4 (F03–F06) · ce5b7f5 (F07) · 3519b2f (F08–F10) · afd8174 (F11–F13) · plus F14.

User rulings — 2026-09-11

Aker BP application date: 2026-07-29 (confirmed; the CLAUDE.md cell saying 07-30 was wrong).
Aker BP Adjacent-despite-79: user does not remember and considers it immaterial, role closed and rejected — "mach wie du es am saubersten findest". Resolved as: keep the recorded class, record an honest override reason saying the rationale is unrecoverable, and do not reclassify. Silent reclassification was the review's own objection.
F18/F19 document layout changes: authorized.

Checklist

Finding	Status	Scope / verification
F01	DONE	Six shared workflows; both agent entrypoints route identically; parity/link regression test
F02	DONE	26 registry records, preserved original tables, common current view, interview events independent of outcome
F03	DONE	Quota gone (tracker is uncapped, `cohort_tracker.py` is a shim); Aker BP date and override recorded; validation enforces unique IDs, ISO dates, score range and gate/class consistency
F04	DONE	Three-level rule hierarchy in AGENTS.md + CLAUDE.md; deliverable vs QA separated; `resume_reference.md` Capgemini contradiction removed
F05	DONE	Per-claim forbidden wording enforced, identity/employer/placeholder/exclusion checks, `document_exclusions` given a canonical home, PASS no longer claims to be a truth certificate
F06	DONE	LaTeX comments, markup, accent macros and umlauts normalized; negations respected; German `und` in the cadence hint; advisory warnings labelled as never a gate
F07	DONE	Overstatement narrowed (BS-1, zero-downtime, retrieval, evaluation, PyTorch context); undervaluation fixed (BS-6 reached no bundle at all; Linux/Ansible/Bash missing from taxonomy)
F08	DONE	`evidence_schema` + per-claim `evidence` block on all 26; 8 dates recovered from existing text, 18 written `unrecorded` rather than guessed; coverage reported as an advisory hint, never a gate
F09	DONE	`12+ years` -> `over ten years` with the reasoning inline; `cv_template.tex` banner-marked as the academic legacy path
F10	DONE	US figures renamed `*_orientation_only`; the Bavarian formula runs foreign-INTO-German, so inverting it is not a recognised conversion. German grades stay canonical
F11	DONE	Requisition-keyed; same-req locations merge; shared titles warn instead of deleting. Live run recovered 57 roles (758→815), including a second RUAG Software Engineer C5I req that was invisible
F12	DONE	Whole-word place matching (Bernau no longer reads as Bern); ~25 Swiss towns incl. Thun and Zimmerwald; Danish cities; pan-regional scope no longer beats an explicit onsite/one-country restriction. All 14 review cases correct
F13	DONE	Identifying query params kept in the key; read and write paths share `lookup_decision`. Live run recovered 7 reposted RUAG decisions, labelled RELATED rather than applied as identical
F14	DONE	Buckets renamed to keyword relevance with an explicit 'NOT a fit judgement'; German titles and platform/DevOps vocabulary added. Kdo Cy ICT-System Engineer (DevSecOps/Platform Automation) went 0→10; Senior Dateningenieur 1→3, level with its English twin
F15	TODO	Typed dates and retained contact/compensation metadata
F16	TODO	Run-scoped reports, transactional persistence and scan coverage
F17	TODO	Reproducible new builds and safe historical archive classification
F18	TODO	Evidence earlier on page one; rendered skills budget
F19	TODO	Education layout, phrasing guidance, reproducible word counts

Sequencing note: F11–F16 all touch scout.py. Fix behaviours in place and keep each finding's diff separately verifiable. Do not start the ~2400-line file split F16 suggests mid-remediation — it would destroy the ability to attribute a regression to a specific fix.

Verification log

Review baseline: 9 existing tests passed; 43/45 LaTeX builds succeeded. See the report for limitations.
F01: moved current workflows into resume_builder/reference/workflows/; preserved skill frontmatter and Claude invocation support. python -m unittest discover -s tests -v: 10/10 pass.
F01 limitation: external quick_validate.py could not start because PyYAML is absent; repository tests validate frontmatter and both routing links.
F02: imported 26 applications from the newer CLAUDE table and seven structured cohort records. Originals preserved in job_scout/state/history/application_status_before_2026-09-11.json.
F03: application_tracker.py summary runs uncapped and reports "interview reached: 1" for SBB despite the withdrawal — the F02 separation working. Aker BP written through records.transaction(), so validation and view regeneration ran; --check reports views current.
F04: resume_reference.md said "Capgemini/Bundeswehr: include only when directly relevant", contradicting the standing never-include rule. Corrected, and the two rules separated.
F05/F06 verification — the counter-proof document is the evidence:
- All six counter-proofs from the report are now caught (9 errors on one synthetic document).
- % Never claim Terraform (a comment) and an honest "no Terraform experience" — English and German — correctly produce no finding.
- All six packages built under current rules scan clean: SBB, NVIDIA, RUAG, Kdo Cy, Citadel, Aker BP — 0 errors each. Four false positives of my own making were found and fixed along the way (title collision across employers, education institution vs employment entry, umlauts, LaTeX accent macros). Each is now a regression test.
- 13 older packages DO report findings (Apple 9, Microsoft Principal FDE 8, …). These look like genuine historical violations under today's stricter rules — "petabyte scale", "own the AWS data platform", uncontextualised PyTorch/TensorFlow. Not fixed: F17 governs historical archives and they must not be silently rewritten. Feed this list into F17.
- tests/test_document_scanner.py added: 8 must-fail, 4 must-not-fail, plus a guard that current packages stay clean. Full suite 31/31 pass.
F07: raw history preserved; only selection rules and safe wording changed. The NVIDIA/RUAG "line could not be stopped" phrasing is left in the submitted letters and recorded as an OPEN ITEM in experience_bosch.md - plausible, unevidenced, to be confirmed before reuse.
F08: unrecorded is deliberately a real value. Failing the build on it would push a future session to invent dates to get a green run.
F10: no stale final_grade_us_4pt references remain anywhere in the repo.
F11–F14 validated on a live 37-company run: 815 matches vs 758 the day before, the delta being requisitions the title dedupe had been deleting. Seven employers reported shared-title collisions that are now kept and flagged (Grafana 5, Axpo 6, AWS 3, RUAG 2, …).
F14 note: the score floor on RUAG/SBB/BKW/PostFinance/Kdo Cy was a workaround for missing German vocabulary. The vocabulary is now the fix; the floors can probably be reduced later, but they were left in place to keep this diff attributable.
Not re-run this session: the 45-document LaTeX build sweep and the visual sampling.

Completion rules

DONE requires implementation plus relevant checks. Record partial work and exact next action. Missing factual evidence must remain explicitly unresolved, never silently promoted to verified. Historical packages are preserved; fixes apply to active sources and future generation unless specifically noted.

8.3 KiB Raw Blame History Unescape Escape