F03 — application records
- Quota already removed with F02's tracker; verified uncapped and that
"interview reached" survives a later withdrawal.
- Aker BP: application date confirmed 2026-07-29 by the user, settling the
one-day conflict. Its Adjacent-despite-79 classification is kept, NOT
reclassified: the user does not remember the rationale and judged it
immaterial on a closed role, and silently reclassifying to make the
arithmetic tidy is the exact objection the review raised. Recorded as an
explicit override reason with a rubric version instead.
F04 — rule hierarchy
- Three levels stated in AGENTS.md and CLAUDE.md: immutable evidence rules,
standing user preferences, market/employer presentation.
- Deliverables separated from QA. The .tex stays the artefact of record; you
still compile, into the application's own output/ folder, because the user
cannot approve what he cannot open.
- `resume_reference.md` said "Capgemini/Bundeswehr: include only when directly
relevant", contradicting the standing never-include rule. Split and fixed.
F05 — a green validator now means something
- Per-claim `forbidden` wording is enforced. 14 of 26 claims carried such a
list and NONE of it was checked, which is why "owned the full ML lifecycle"
passed. Matching tolerates inserted adjectives: the real regression read
"Built a decentralized Data Mesh" and one word defeated the literal list.
- Added identity (e-mail/phone), employer-date, placeholder and
document-exclusion checks. Capgemini's omission rule now has a canonical
home in claims.json instead of living only in memory and config.md.
- Registered Groovy and Snowflake as forbidden skills. Both were absent from
claims.json entirely, so "forbidden" existed in prose with nothing enforcing
it.
- PASS no longer reads as a truth certificate: it prints what it did not check
— scope implications, bullet-to-claim support, whether a number is real.
F06 — the scanner stopped fighting the documents
- LaTeX comments are no longer claims; `% never claim Terraform` was a finding
before, which taught the workflow to write around the checker.
- Markup, accent macros and umlauts are normalized, so `\textbf{Snow}flake`
cannot smuggle a forbidden skill past a substring test and
`Universit\"at der Bundeswehr M\"unchen` is recognisably a school.
- Honest denials pass in English and German.
- The triple-list cadence hint knew only English "and", so it silently did
nothing on every German package. Advisory only, and now labelled as such.
Verification
- The six counter-proofs from the report are caught: 9 errors on one synthetic
document, where the old scanner found none.
- All six packages built under current rules scan clean — SBB, NVIDIA, RUAG,
Kdo Cy, Citadel, Aker BP. Four false positives of my own were found and fixed
on the way there; each is now a regression test.
- 13 older packages do report findings. Those look like genuine historical
violations and are deliberately NOT fixed here: F17 governs archives and they
must not be silently rewritten.
- `tests/test_document_scanner.py` added. Full suite 31/31 pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qej5SBvhcbGXHwLqNWiskM
Scout run 2026-09-04 (36 companies, 778 eligible) plus the NVIDIA package it
surfaced, and the tooling and policy fixes found along the way.
Application
- NVIDIA Senior DevOps Engineer, Zurich (JR2024256): submitted 2026-09-05.
Evidence Fit 64/100 Stretch, hard gate FAIL x2, channel Weak. Sent on an
explicit user override with the working model unresolved, so hybrid-vs-onsite
is a live screening topic. 2pp resume + 1pp cover letter, both validators
PASS, 0 boxes, 0 Tier 1 findings open.
Policy
- Drop the 7 Core / 2 Adjacent / 1 Stretch cohort caps. They had no derivation
and were overridden whenever they bound, and the cohort was confounded: it
tested fit class while holding channel constant at Weak across all six
applications. Keep the recording, drop the rationing.
- Channel replaces the slot budget as the rationed variable: no application
goes out cold when a named contact is published and reachable.
- Propagated through shared_ops, session_file_template, cohort_tracker
(counts, never counts-against-a-target) and application_cohort.json.
Tool fixes
- fetch_workday emitted 404 URLs for every Workday company (NVIDIA, Novartis,
Equinor): externalPath is site-relative and needs the locale + site prefix.
6/6 short-form 404, 3/3 long-form 200. 773 dead links repaired across 42
archived reports.
- find_decision_url silently de-linked decisions on two boards, which is why
ruled-out roles kept resurfacing as undecided. RUAG rotates its UUID on
repost; BIS migrated its vacancy URL scheme, orphaning 5 decisions including
a live application. Fixed as a two-tier match, not a slug key -- slugs
collide (two live RUAG reqs shared one).
Knowledge base
- claims.json: add Bash / shell scripting, the third canonical gap of this
shape after Ansible and Linux administration.
- bundle_data_platform: BS-4 was framed as a delivered observability stack over
what claims.json calls a proof of concept -- the defect the 2026-08-21 sweep
was told to look for.
- CLAUDE.md and AGENTS.md: correct the RUAG Data Lakehouse shortlist claim,
which the JD does not support and which both files still asserted.
- resume_template: comment-guard the missing babel/ngerman language block.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QFb8tTCig5r1KzkQiNXcYt
RUAG C5I - AI Engineer C5I (Thun, app ID 18027): SUBMITTED 2026-08-27.
Evidence Fit 74/100, fit class Stretch, title-capability hard gate FAIL
(R1 build/operate the AI-LLM platform and R2 in-house LLM/retrieval are
Adjacent, not Direct) - all knowingly accepted under the user's explicit
Phase 0 override. Consumes the cohort's sole Stretch slot: cohort is now
6/10, Stretch 1/1.
Critique ran twice. Round 1 scored the documents 85/100 and raised three
Tier 1 items; Round 2, after the fixes, scores 93/100 with truth and
provenance 24/25 and zero Tier 1 findings. Evidence Fit did not move and
was not allowed to: PP-3 is a personal project and cannot convert an
Adjacent title-capability into a Direct one.
Documents (.tex is the deliverable; PDFs are gitignored):
- resume 2pp/13 bullets, letter 1pp/295 words, both validators PASS
- headline is now the canonical title "Staff Data, Analytics & AI
Engineer", which puts the req's own word above the fold
- PP-3 added as a Projekte section - the only current Linux/hardening
evidence, since BS-6 ended Dec 2022
- the letter names the AI-platform gap in one sentence, then connects
SW-5 to C5I's stated DevSecOps model
- "governte" -> "governance-konforme"; JD term coverage 18/22 -> 22/22
claimable, with all six gap terms still correctly absent
Two defects found and fixed that the first critique missed:
- the resume never loaded babel, so a German document was hyphenated
with English patterns (Hal-bleiterfertigung, Tran-skription). Fixed
with babel[ngerman] plus a Transkription exception; this also cleared
an overfull box. Check this on every future German package -
resume_template.tex likely has the same omission.
- IBM AI Engineering had no primary-source record. Certificate read
directly (IBM via Coursera, 4 Jun 2020, verify 3ZBZFVAL6A34), recorded
as entry #8; it also proves PyTorch, now evidence: certification.
Also included: the submitted Kdo Cy DevOps Engineer III package, the
Capgemini omit rule promoted into config.md, BW-1 canonicalized in
experience_bundeswehr.md, and a scout.py comment correcting the
telenorgroup slug from "near-empty" to a claimed board serving stale
phantom listings that DEMO_TITLES would not catch.
Compensation, PSP/project eligibility and role level were never resolved
before sending and are now live screening topics.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JMcHCsTKWvVzqyLChF5ckk
ECTS grade B is the next 25% after the top 10%, so the accurate phrasing
is 'top 35%'. The record's original value was right; the 'top 30%'
considered yesterday was the error, now removed rather than left as an
open question.
The field also now says never to narrow B to a point estimate. B is a
band, and quoting 'top 30%' or 'top 25%' overstates the standing - the
kind of small, checkable claim that costs credibility for no gain.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
A Citadel application form asked for GPA and neither degree's
Abschlussnote existed anywhere in claims.json - only the M.Eng. thesis
grade. The single lead was an old CV extraction whose own provenance
note disclaims it and which already carries a known error, so it was not
safe to convert from. User confirmed both directly: M.Eng. 1.6 and
B.Eng. 2.4 on the German 1.0-4.0 scale.
US 4.0 equivalents recorded via the modified Bavarian formula, which
simplifies to 5.0 - G: 3.4 and 2.6. Both fields carry usage rules,
because a converted grade is easy to misuse. They are for application
forms and credential checks only, never to be presented as a natively
awarded GPA, and the German scale must always be stated alongside -
particularly for the bachelor, where the conversion makes a solid "gut"
read weaker to a US audience than it is.
Also flagged a percentile that does not agree with itself. The relative
grade is ECTS B, but this record said top 35%, the user said top 30%,
and ECTS B is formally the next 25% after the top 10%. Rather than pick
one, the field now says to check the certificate and notes that quoting
"ECTS grade B" alone is always safe. A specific percentile on an
application is exactly the kind of small number that a credential check
can contradict.
New grade_output_rule: neither grade belongs on a resume or in a cover
letter by default. Industry resumes omit grades, and the thesis 1.0 is
the only figure worth surfacing where academic performance is valued.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
claims.json had zero hits for trading, finance or quant, so a real
capability could not legally reach any document: the file is the
highest-authority source and nothing outside it is claimable. Surfaced
while assessing the Citadel Securities Platform Engineer req (Zurich),
where domain fluency is exactly what was missing.
PP-1 is the self-built investing/signal platform: US-equity price,
fundamental and LLM-assisted sentiment ingestion, a long-only
cross-sectional residual 12-1 momentum book with ATR stop/trail and max
15 concurrent names, scheduled scan and backtest pipelines, web
dashboard and Telegram alerts. PP-2 is the Udacity AI for Trading
nanodegree. Two allowed-with-context skills reference both.
Written defensively, because this is the kind of entry a later run
inflates. Each carries an explicit forbidden list: no PnL, Sharpe,
return or backtest figure may EVER be quoted (none is verified); PP-1 is
single-user and self-hosted, so never distributed, scalable, production
or multi-user; neither implies professional quantitative, trading or
finance experience; the nanodegree is never an academic or quant
credential. Legitimate use is self-directed domain fluency plus an
end-to-end ingestion/backtest/alerting build outside work.
PP-1 is also the one claim where full-ownership verbs are correct - it
is genuinely solo, unlike the employer work that Scope Discipline
governs. Stated in the scope so the two rules do not collide later.
JSON valid, validator PASS, 9/9 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
Explore SDU Center for Energy Informatics PhD, Theme 2 (Predictive
Maintenance and Asset Management of Smart Energy Networks), Odense DK,
deadline 2026-08-20. JD scraped live via Playwright.
Verified the master's thesis against the source PDF rather than the
stored title alone:
- Vibration-based condition monitoring of CNC machine tools; features
are the dimensionless waveform, peak, pulse, margin and kurtosis
indices.
- Implemented a hybrid of rule-based reasoning and a 7-10-3 ANN.
- Evaluation measured throughput and latency, not model accuracy:
~500 SPS pipeline against 72.9 kSPS sensors.
- Corrects an earlier note in the same session: PSO was surveyed only
and never implemented. Recorded under thesis_limits alongside the
absence of real operational data and accuracy figures.
Also records the ECTS relative grade (B, top 35%), the existence of
English-language transcripts, and the concrete Bosch sensor-data types
(defect management records, wafer inspection images, PCM electrical
parameters) under BS-2.
Adds the enquiry email sent to Prof. Bo Norregaard Jorgensen, which
asks whether an industry candidate is viable and whether non-academic
letters of recommendation are accepted.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Aker BP ASA — Data Product Architect (FINN 469067315): archive the finalized
resume/CL sources, JD, session and critique. Submitted 2026-07-30 ahead of the
2026-08-02 deadline; logged as applied in the scout decision log and marked
SUBMITTED in Active Sessions.
KB: record Atlassian Compass as Swisscom's metadata/catalogue platform for data
products and lineage (user-confirmed 2026-07-29). Practitioner use only —
claims.json forbids claiming administration, rollout or ownership, and it is
explicitly not a substitute claim for Purview, Collibra or Alation.
Also: close out the Google Business Home session (no interview), extend the job
scout, and allow the job-board domains used during the Aker BP research.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
dennisthiessenandClaude Opus 5