User corrections, 2026-09-11.
KOTLIN was absent from claims.json entirely and was therefore unwritable under
the evidence-first rule. Bosch 2020-2022: backend services, Ansible plugin work
and extensions. This is the FOURTH real skill found missing this way, after
Ansible and Linux (2026-08-27) and Bash (2026-09-05).
Correction to an earlier statement in this session: I described Kotlin as the
fifth instance and named Groovy/Snowflake as the other two. That was wrong and
the user caught it. Groovy and Snowflake were registered the same day as
`forbidden` - skills he does NOT have, registered so the document scanner could
catch them. The opposite of this pattern. The Kotlin note records the
distinction so it does not propagate.
Language split per employer, none of it previously recorded:
- Fraunhofer: C#. FC-1 and FC-3 named no language at all.
- Vizrt: Python for the wrapper, C++ for the transcoding engine (FFmpeg).
Cross-referenced to the standing rule not to oversell C++.
- Bosch: primarily Python, with some Java and Kotlin (backend, Ansible plugins,
extensions) and C# for Spotfire plugins.
BS-3 / BS-5: he was TECHNICAL Application Owner for the KLA Defect Management
System (DMS) as well as for Spotfire. "Technical" is explicit so the line cannot
be read as people management or vendor-side ownership.
SW-6 (PySpark): the scope line carried "scale and performance metrics are not
verified", which duplicated the metrics field and read as doubt about the skill
rather than about the numbers - the skill is production-current and allowed. The
caveat now lives only where it belongs.
last_verified REMOVED from all 26 evidence blocks at the user's instruction:
every claim already carries its employer in the ID prefix, and for closed
employers a recency date measured nothing. Provenance (source, ownership) stays.
The schema records the reasoning so it is not reintroduced out of habit.
Validator advisory warnings 18 -> 0. 37/37 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qej5SBvhcbGXHwLqNWiskM
BS-1 operating constraint - the F07 open item, resolved by the user
The user clarified what the 24/7 constraint actually is: a defect in the
software could, worst case, halt production - an outcome that must never occur.
That is a FAILURE-CONSEQUENCE constraint, and the submitted NVIDIA cover letter
states something different: "a production line that could not be stopped for a
deployment" describes a scheduling window. The distinction matters in both
directions - the letter is inaccurate, and the truth is the stronger claim.
- BS-1 scope now records the real constraint and how to write it.
- Both variants of the letter's phrasing are added to BS-1 `forbidden`, so the
scanner catches them. Verified: it now flags the submitted letter.
- BS-1 evidence filled from a real source (this clarification) rather than left
unrecorded, including ownership - integration and operation, not model work.
- The submitted letter is NOT rewritten (F17). experience_bosch.md records the
resolution and the wording to use instead.
Scanner false positive fixed: the employer-date check now runs on CVs only. A
cover letter naming an employer in prose is not claiming a date.
Cleanup, authorised by the user
- 136 LaTeX build artefacts removed (already gitignored, regenerable).
- 7 superseded intermediate QA renders removed, 6 of them in NATO_AI_ENGINEER
which carried nine PDFs for two documents.
- Nothing evidential touched: 36 submitted PDFs, 41 e2e deliverables, 44 sources,
54 session/critique files and 20 JD files are intact. The Kraken PDF was
checked before deleting and turned out to be the saved job description, not a
render.
- output/ 16M -> 13M.
37/37 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qej5SBvhcbGXHwLqNWiskM
F18 - page-one weighting
- Across four compiled packages the first employment entry began at 45-55% of
page one. A reader reaching the middle of the page had seen a profile and a
skills block, and no evidence.
- The old skills rule counted `\skillline` MACROS, which does not control
density: six categories with long value lists wrap to twelve or more rendered
lines. Replaced with a rendered-line budget verified in the compiled PDF, plus
the rule that a skills line repeating what three bullets below already prove is
spending the first screen to say it twice.
- Employer chronology stays reverse-chronological; the ordering freedom is
WITHIN a position, where the bullet closest to the target role leads.
- New helper `page_metrics.py` measures page-one geometry from the PDF via
pdftotext -bbox.
F19 - education and phrasing
- Degree + focus + institution crammed into one cell with a right-aligned date
produced visible word spacing and mid-word breaks in SBB, NVIDIA and Aker BP.
Split the entry, as RUAG_AI_Engineer_C5I already did. Template carries the rule.
- Grades and thesis stay only where concretely relevant. claims.json
grade_output_rule already keeps grades off by default; the M.Eng. thesis earns
its place on condition-monitoring or sensor-data roles and otherwise does not.
Validated by measurement rather than assertion
- page_metrics.py first reproduced the review's own figures (NVIDIA 55.4%,
SBB 50.6%, RUAG 50.3%). That is what makes the after-figure meaningful.
- The NVIDIA CV was rebuilt FROM A COPY with the guidance applied: experience
moved 55.4% -> 44.0% of page one, page-one text lines 47 -> 39, still two
pages, zero overfull boxes, validator PASS.
- The submitted package itself was not modified. F17 classifies archives as
evidence, not as drafts.
The percentage is a diagnostic, not a target: no pixel goal and no rule that
page two must be filled.
All 19 review findings are now complete. 37/37 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qej5SBvhcbGXHwLqNWiskM
Build sweep run this session: 48 sources, 46 built, 2 failed. Infineon_AI_Engineer's
cover letter fails on font expansion (the font is not scalable); cv_template.tex
fails on unfilled placeholders, which is expected for the academic legacy path.
Overfull boxes: Microsoft ISE 25.33pt, Isovalent 0.88pt. All historical, all
left alone.
New: resume_builder/helpers/package_manifest.py records, per application
package, what a rebuild would ACTUALLY do - sources, class-file hashes, PDFs
present, and the reasons a rebuild would not reproduce the submitted document.
It is descriptive; nothing in it authorises rewriting a submitted file.
What it found across 21 packages:
- 17 are DATE-DEPENDENT. The cover letters use oday, so recompiling a
historical package silently restamps it with today's date.
- 1 has an INCOMPLETE SOURCE DISTRIBUTION: output/NATO_AI_ENGINEER ships no
resume.cls. It builds on this machine only because a copy exists elsewhere,
and the shipped PDF hid that. Not repaired by copying the newest class in -
the repo carries THREE different resume.cls variants (13/7/1 copies) and older
macros differ, so that "fix" could silently change a historical document.
- 1 has no PDF at all, so nothing is reviewable without a build.
Classification: 7 application packages were in no reuse category. Added to
historical_revalidate, and the register now records the default explicitly -
anything unlisted is treated as revalidate-before-reuse. Being recent is not the
same as being verified; none of the seven was re-read for this.
Prevention rather than repair: oday removed from the cover-letter template so
new packages stop adding to the pile. A test now fails if it comes back, if the
manifest goes stale, or if a package is left unclassified.
Also: scan_document gained a template mode. A placeholder in a TEMPLATE is what
makes it a template; the defect is one surviving into a generated document.
37/37 tests pass; counter-proof still catches all 9.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qej5SBvhcbGXHwLqNWiskM
F15 - dates and metadata kept their meaning
- The Taleo parser stored an application DEADLINE in `posted`, so a role closing
on 30 September looked as though it had been published then - backwards for a
freshness signal. It now populates `deadline`, and `posted` stays empty
because that board genuinely publishes none.
- The Prospective adapter read unit, Lohnklasse and a named contact, folded them
into `description` for scoring, and `_process_company` then discarded
`description` - losing exactly the fields that decide level, pay and whether a
cold submit is necessary. They are now carried as fields through to the report
and the JSON dump, alongside `date_provenance`.
- Named contacts now appear in report rows. Channel is the rationed variable
since 2026-09-05, so a published, reachable name is the difference between a
cold submit being a choice and being the only option.
F16 - writes stopped destroying delivered work
- Every run of a day wrote `<date>.md`/`.json`, so a restricted run replaced the
full daily report. This is not hypothetical: it happened twice in this session
and both times the 37-company run had to be repeated by hand. Restricted runs
now write `<date>.only-kdocy.md` and say so.
- `save_seen()` ran in a `finally:` BEFORE the report was written. A failed
report write therefore left roles marked seen, so a `--new-only` retry would
hide precisely the hits that were never delivered. Seen-state is now committed
only after report and dump are on disk - and never by a restricted run, which
did not look at the whole board.
- Atomic writes used a fixed `.tmp` name; two writers could share it and
os.replace the wrong content into place. Unique temp names now.
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
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
These three all cost roles BEFORE a human ever read them, which is the
expensive kind of defect: a job that never reaches the report cannot be
evaluated at all.
F11 - dedupe by requisition, not title
- The old rule collapsed everything sharing a normalised title within one
employer. Two different reqs called "Senior Data Engineer" in Bern and Zurich
became one row, and different teams hiring the same title vanished.
- Requisition identity is now the key, parsed from the URL with the same
patterns the decision log uses, employer-scoped so two boards cannot collide.
- The legitimate case the old rule existed for - one role republished per
location - is a SAME-REQUISITION case, so those merge their locations.
- A shared title across different requisitions is now a warning, never a
deletion.
F12 - location filters
- Substring matching accepted "Bernau bei Berlin" as Bern. Place matching is
now whole-word.
- Thun, Zimmerwald, Baar, Olten and ~25 more Swiss towns were rejected outright
when the adapter supplied no country. Zimmerwald is the Kdo Cy site he applied
to; Thun is where he lives.
- Aarhus, Odense and other Danish cities added alongside Copenhagen.
- A stray "Europe" no longer implies remote eligibility next to an explicit
onsite marker or a single-country residence restriction.
- All 14 cases named in the review now resolve correctly.
F13 - decision identity
- The generic fallback key dropped the query string, so /vacancy?job=1 and
?job=2 were the same key and one decision silently covered both. Identifying
query parameters are now kept; tracking noise is not.
- The READ path had no slug fallback although the WRITE path got one on
2026-09-04. Every RUAG repost therefore came back undecided despite a recorded
decision - which is what happened to Lead-Architekt Network- und Richtfunk on
2026-09-10, days after 63 RUAG rule-outs were written.
- Both paths now share `lookup_decision`. Verified on a live RUAG run: 7
reposted requisitions recovered their decisions.
- A slug match is reported as RELATED with the matched URL, never applied as if
it were the same requisition: a repost can differ materially, as RUAG
demonstrated when it moved a disqualifier into a job title.
tests/test_scout_matching.py added. Full suite 34/34 pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qej5SBvhcbGXHwLqNWiskM
F08 - evidence metadata
- Only 1 of 26 claims carried structured source/last_verified/ownership, which
is why recency and reference checks could not run. Added an evidence_schema
block and an evidence block on all 26 claims.
- 8 dates were recovered from each claim's OWN recorded text; the other 18 are
written as the literal "unrecorded". Nothing was inferred. "unrecorded" means
nobody wrote it down, NOT that the claim is unevidenced - most carry evidence
in their free-text scope.
- Coverage is reported as an advisory warning, never a gate: failing on it would
push a future session to invent dates to clear the run.
F09 - templates
- resume_template.tex asserted "12+ years of production software experience".
It does not hold: canonical employment starts 2014-11 (~11y10m) and the
VISIBLE chronology starts 2015-05 (~11y4m) because Capgemini is omitted.
Now "over ten years", with the arithmetic and the no-date-math rule inline.
- cv_template.tex still carries page targets, rendered-line budgets and 2L/3L
variants the project dropped. Banner-marked as the ACADEMIC legacy path so it
is not reached for whenever a posting says "CV".
F10 - US GPA
- The modified Bavarian formula converts FOREIGN grades INTO the German system.
Inverting it to produce a US GPA is not a recognised method, and the TUM
source calls even its forward calculator orientation only.
- Fields renamed to final_grade_us_4pt_orientation_only with that caveat. The
German grades remain the canonical facts. If a recipient publishes its own
conversion, use theirs and record which was used.
Validator PASS (1 advisory), 31/31 tests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qej5SBvhcbGXHwLqNWiskM
The review's key point is that the error ran BOTH ways.
Overstatement, narrowed to canonical scope:
- BS-1 variants claimed "eliminated manual wafer inspection", "no manual
intervention" and "unattended quality monitoring". Canonical BS-1 supports a
qualitative reduction whose size is unverified. Reworded; the scope limit is
now stated inline where the variants are written.
- bundle_data_platform called BS-1 a "zero-downtime constrained deployment".
Integration into a fab that runs continuously is not a measured zero-downtime
property.
- experience_swisscom carried "grounded retrieval" as a skill and ATS keyword
while SW-8 records retrieval implementation as explicitly unknown, and its
reframing note invited "evaluation discipline" although formal model
evaluation is a globally forbidden pattern. Both removed.
- PyTorch sat in bundle_ml_ai_engineer as a bold tool and an ATS must-include
with no certification context - the exact way the context gets lost.
Undervaluation, the half that quietly costs him roles:
- BS-6 (Linux administration, intensive Ansible incl. self-written extensions)
has been canonical since 2026-08-27 but appeared in NO bundle, so the normal
selection paths could not see it. Added to the data-platform and data-engineer
bundles with its currency limit (employer-side ended Dec 2022; PP-3 is
personal) and the standing GitOps prohibition.
- Linux, Ansible and Bash were missing from skills_taxonomy entirely.
- skills_taxonomy still said the IBM certification year was "not recorded" and
that PyTorch needed verification; both were settled on 2026-08-27.
Not rewritten, deliberately: the NVIDIA and RUAG cover letters state the
production line could not be stopped. That exceeds the canonical record but is
plausible and both are already submitted. Recorded as an OPEN ITEM in
experience_bosch.md to be confirmed with the user before any reuse.
Validator PASS, 31/31 tests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qej5SBvhcbGXHwLqNWiskM
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
Checkpoint for the 2026-09-11 project review remediation. Work by a previous
agent that ran out of context; committed unchanged as a rollback point before
the remaining findings land on top of it.
F01 — the two agent directories had drifted into different processes
- Six shared workflows under `resume_builder/reference/workflows/`; both
`.agents/skills/` and `.claude/skills/` are now thin entrypoints that route
to the same file instead of carrying full copies.
- The `.claude` copies were the stale side: fixed bullet line variants and
character budgets that the newer process had already dropped. Removing them
is the point of the change, not collateral.
- Parity/link regression test added; corrected the top-level file map, which
pointed at a directory that does not exist.
F02 — application status was stored in four places and disagreed
- New registry `job_scout/state/applications.json` (26 records) with
`application_records.py` (schema, validation, transactional write) and
`application_tracker.py` (add/update, no quota).
- AGENTS.md and CLAUDE.md no longer keep their own status tables; both point
at `current_applications.md`, generated from the registry.
- Interview-reached is now independent of outcome, so a withdrawal after an
interview can record both facts. The old tables are preserved verbatim in
`job_scout/state/history/application_status_before_2026-09-11.json`.
Verification: `python -m unittest discover -s tests -v` — 18/18 pass.
Spot-checked that the detailed SBB and Kdo Cy notes survived the migration
into the registry rather than being flattened.
Known unresolved, deliberately not invented (see FIX_PROGRESS.md):
Aker BP's application date and its Adjacent-despite-79 classification.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qej5SBvhcbGXHwLqNWiskM
Scout
- New `prospective` adapter for ohws.prospective.ch, wired up as `kdocy`
(jobs.admin.ch, filtered to Kommando Cyber). jobs.admin.ch is an S3-hosted
SPA that returns AccessDenied to any direct fetch, headless included, so the
site cannot be scraped -- but the SPA reads a public, unauthenticated API.
All 463 federal roles in ~2.4s, against RUAG's 35s Playwright scrape.
- This board is richer than any other source: Lohnklasse, application deadline,
a named contact per posting, and the full Aufgaben/Anforderungen body, so
scoring runs against the JD text rather than the title.
- Thread `deadline` through normalization, the markdown report and the JSON
dump. A board with hard deadlines needs them on the row.
- Match `verwaltungseinheit_*` by prefix: the sub-unit key is department
specific (VBS 1083398, EFD 1083364) and hardcoding one breaks reuse.
- MANUAL_CHECK entry for the Bundeswehr Reserve portal. The data is reachable
(SAP E-Recruiting OData, SearchCategory 0022) but the portal actively blocks
automation, and getting through needs UA spoofing plus masking
navigator.webdriver. Not wired up by choice -- the operator-sanctioned
Suchauftrag/job alert is the better channel. Reasoning is in the comment.
Applications
- SBB Data Engineer 103755: interviewed 2026-09-09, WITHDRAWN by the candidate
2026-09-11. Net downgrade on four axes (41h vs 40h, 25+Brueckentage vs 28+5,
3 vs 2 office days, and Anforderungsniveau K tops out ~145k against ~135k
inside a band running to ~160k -- he would enter K at its ceiling merely to
break even). The cohort's only interview, and it ended as his no, not theirs.
- Kdo Cy DevOps Engineer III: REJECTED 2026-09-10, 14 days, no interview.
Form letter with no diagnostic content. Identical on every recorded
dimension to SBB (core / 79 / gate pass / DQ 92 / weak channel) and the two
went opposite ways -- recorded as worth watching, not as evidence for the
channel thesis.
- Citadel row cleaned up: it claimed the cohort and decision-log entries were
deliberately unwritten while its own status line read SUBMITTED. Both exist.
The next-step cell still said "ask BEFORE building the package" two weeks
after submission.
Decision log
- 69 RUAG rows -> 3 undecided. 63 title/family rule-outs, each noting that it
is not a JD read, plus three that are: Test Manager C5I, System Engineer C5I,
ICT Solution Architect. Guarded against the shared F/A-18 slug so one write
could not silently swallow the other.
Note: the K->L/Kader rule is SBB-specific and was deliberately narrowed --
Anforderungsniveau is SBB's own system and does not map onto jobs.admin.ch
(Lohnklassen), RUAG or PostFinance. Thun-area employers sit outside the Bern
salary logic entirely.
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
The deciding question on 2026-09-09 is a number nobody has, so the
worksheet is a frame to complete before the call rather than a filled-in
answer: Anforderungsniveau K is not public, which is why Q1 exists.
Records the three things that make a Swiss base-to-base comparison lie:
the free GA, Regionalzulage and employer pension share are real money
invisible in a base figure, while a Swisscom bonus is real money SBB
will not match; the GA is a taxable Lohnausweis benefit so its net value
sits below list price; and the GA and the "commute saving" are the same
franken, so counting both flatters SBB by four figures.
K is a capped GAV band with no negotiation lever at the top, so this is
a yes/no against a fixed number, not an opening position.
User's read is that it will not clear. Taking the call regardless: the
questions are worth asking, it is interview practice while NATO JWC,
BIS and Microsoft Principal are still live, and the K figure is the
first real data point on what the entire Bern/Thun local tier pays -
RUAG, Swissgrid, BKW and BFH sit in comparable band structures. Worth
recording whatever the outcome.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JMcHCsTKWvVzqyLChF5ckk
SBB Data Engineer (Job ID 103755): invited 2026-08-28 for Wednesday
9 September 2026, 08:30, 45 min on Teams.
Per the posting's own process this is stage 2 of 4 - "Virtuelles
Kennenlernen mit HR und Fuehrungskraft" - so HR together with Andri
Wienandts, not a pure HR screen. 45 minutes shared between two people,
which is the practical constraint the brief is built around.
Two facts worth recording:
- This is the FIRST conversion of the evidence-first cohort (6
applications, 2 rejections), and it came from a COLD submit. The
channel plan called for phoning Wienandts before applying; that never
happened, so the published warm contact went unused.
- The resume never contained Snowflake, dbt or Power BI - the three
literals an ATS keyword screen would have keyed on, and a risk the
session had explicitly recorded as "accepted, not solvable". It did
not filter him out, so a human read the dossier and the honest
substitution framing survived first contact. One data point; not
enough to revise strategy on, but it is evidence against assuming
honest omissions are fatal at the screen.
Brief covers the four questions that must not go unasked (Anforderungs-
niveau K band, design/architecture scope vs current Staff + Component
Owner level, RAMSI's Java/Spring Boot/Angular share, Kidz Care rate),
the substitution answers for each named-but-non-canonical tool, the
swissTAMP/RIS/MUD context, and the thesis boundary - methods prototype,
no operational data, no accuracy figures, PSO surveyed only.
It also asks him to decide three things before the call rather than on
camera: what he does if K lands at 130-150k, if architecture is set
elsewhere, or if RAMSI turns out to be heavily Angular. Any one can be
a no, and the level question is the same shape as the declined BKW.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JMcHCsTKWvVzqyLChF5ckk
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
Rejected 2026-08-27, no interview, ~28 days after submission. Recorded
in all four places that track an outcome: cohort state, decision log,
the Active Sessions row and the session file's status block.
The cohort slot is not freed. A rejection consumes an attempt the same
way a submission does, so the Adjacent count stays 2/2 and the cohort
stays 4/10 - unlike the Google FDE III and AWS FDE no-gos, which were
declined before a package existed.
Recorded as rejected_no_interview on the evidence: nothing in the log
shows any contact after submission. Recruiter-vs-hiring-manager
disposition is a tracked cohort variable, so this is one edit away from
being corrected if a screen did happen.
The package was validator PASS with no Tier 1 truth findings and the
follow-up email to Per Olav Marthinsen was never actioned, making this a
pure cold submit. A cold-channel no-interview rejection does not
implicate document quality on its own. That is now 2 of 4 cohort
applications rejected without interview, both cold, which fits the
channel thesis already in the log - and application_strategy.md forbids
reacting to a single rejection, so no positioning or targeting changed.
The Norway lane is explicitly not closed. NATO JWC Stavanger is still
live and strong Norwegian employers remain in scope; both the row and
the decision note say so, because with Aker BP closed the earlier
Equinor note ("lane now reduced to the open Aker BP application") would
otherwise read as an implied closure.
Submission date left inconsistent on purpose: the contemporaneous
records say 2026-07-29, the retrospective ones 2026-07-30. Both now
carry a note pointing at the other rather than one being overwritten.
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
Submitted 2026-08-26. Cohort entry added (Adjacent 2/2, cohort now
4/10, channel weak x3) and decision logged.
The finding that matters is not the submission. The application form's
preferred-work-location selector offered no Zurich, despite the JD body
reading "in Miami, Zurich or New York", the posting header listing
Zurich, and the site's own Zurich location filter returning this exact
req when checked on 2026-08-25. The careers site footer reads Citadel
Enterprise Americas LLC, so the likeliest explanation is a US-entity
apply flow with European seats routed elsewhere, or a Zurich seat that
is closed while the multi-location posting stands.
This partly supersedes the working-model risk that blocked Phase 2. If
no Zurich seat is reachable through that form, whether Zurich runs
hybrid or five-day onsite is moot, and the application may be sitting
against a US requisition - a hard no under the standing no-relocation
constraint.
Carry-forward rule recorded in both CLAUDE.md and the session file:
check the application form's location selector before investing in a
package. Two independent signals - the posting's stated locations and
the employer's own location filter - both proved unreliable as evidence
that a seat is actually reachable. Phase 0 treated the filter result as
confirmation that the req was live in Zurich; that inference was weaker
than it looked.
Package itself is finished and reusable: both documents cleared the
validator with zero warnings and the Tier 1 fixes lifted JD term
coverage from 51% to 68%. Nothing about the documents is the weak point.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
Finalization per shared_ops.md. Both submitted documents re-validated:
resume and cover letter each PASS with zero warnings. Final PDFs copied
to Dennis_Thiessen_Resume.pdf and Dennis_Thiessen_Cover_Letter.pdf.
Steps 5 and 6 - submission date and cohort entry - are deliberately not
done. The user directed "apply as is" but has not confirmed the
application was actually sent, and stamping today's date on a submission
that may happen tomorrow corrupts the log that every cohort and
disposition-timing judgement is read from. Both commands are written out
in the session file, ready to run on confirmation. This is the same
discipline applied to the SBB Wienandts call: record what happened, not
what was intended.
The critique stays STALE by choice rather than being re-run: it scored
the pre-Tier-1 resume and predates the cover letter. The estimated ~91
post-edit Document Quality is flagged in both the session file and here
as an estimate that must never be quoted as a score.
The working-model question is unresolved and now travels with the
application rather than blocking it, which is the user's call and a
reasonable one now that the package exists: clarifying first only pays
off in the branch where a cold application gets a call. It must be asked
at first recruiter contact. If Zurich is 5-day onsite this is a decline
on the Bern constraint no matter how the process goes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
Cover letter: 1 page, 270 words, 3 paragraphs, validator PASS. Carries
the one thing the resume structurally cannot - why a Swisscom data
engineer credibly wants a research-platform seat at a market maker,
which against an ex-FAANG field is the main differentiator.
Deliberately uses no external hooks. cl_reference.md says to omit an
unnecessary hook rather than spend words proving company familiarity, so
the only hook is the JD's own language, scraped verbatim and first-party.
No named executives: nothing to verify, and nothing that reads as
name-dropping. The letter also states the R1 gap plainly rather than
hiding it, because a technical reviewer will find it in the first
question anyway and owning it is stronger than being caught by it.
PP-1 appears with its personal-project label and an explicit "I make no
claims for its results".
Separately, loading the bundles for this letter surfaced two live scope
traps. The 2026-08-21 correction that fixed the SW-1 violation in
bundle_data_engineer.md ended with "check the other four bundles for the
same pattern". That sweep was never done, and two of them carried it:
bundle_data_platform.md - the S5 cover-letter hook read "migrating
Swisscom's legacy ETL stack to a cloud-native AWS platform", and its
narrative thread said "migrating an entire ETL infrastructure". Both
pair a full-ownership framing with a company-scale object, which
claims.json SW-1 forbids outright.
bundle_ml_ai_engineer.md - the SW-1 reframing row read "Built
cloud-native data infrastructure on AWS ... the scalable data layer",
which is both a full-ownership verb on an org-scale object and an
unverified scale claim.
Both rewritten to the scoped form with inline scope warnings, matching
the data_engineer fix. bundle_analytics_engineer.md and
bundle_semiconductor.md were checked and are clean.
These were loaded traps: any future cover letter built from either
bundle would have started from a sentence claims.json forbids.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
All three Tier 1 items from the critique, all honestly available, no
claim altered.
The R2 vocabulary now appears where R2 is actually evidenced: the SW-7
bullet carries ingestion and lifecycle, and SW-1 carries transformation
and storage. R2 was the single core responsibility scored Direct, and
the document had been proving it in substance while missing it in the
JD's own words. SW-2 now names the Oracle/Kafka/Python/Teradata estate a
distributed system he operates - not one he designed, which stays a gap.
The second skills line is relabelled "Distributed systems" and names
distributed databases explicitly, closing the fifth preferred-technology
hit that Phase 0 counted as Direct but the resume never actually wrote.
Recorded what was deliberately not added, because raising keyword
coverage is exactly where a later run would be tempted to overreach:
"scalable" is an unverified scale claim, and SDK, high-throughput,
data-intensive, simulation, model development and self-service have no
canonical evidence. Verified still absent after the edit.
Scope kept to what was approved. The user said "tier 1 fixes", so the
Tier 2 headline retitle - "Production Data Platforms on AWS", the line
most likely to draw the opening challenge from a technical reviewer -
was left in place for them to decide.
Validator PASS with zero warnings: the 58% cadence warning cleared as a
side effect, because restructuring two bullets with em-dashes broke the
comma-list pattern the checker keys on. Still exactly 2 pages, zero
overfull/underfull boxes, PDF re-inspected.
Critique marked STALE - it scored the pre-edit document.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
No Tier 1 truth findings, validator PASS, and the claim audit clears
every material claim including the PP-1 block, which carries its
personal-project label three times and no performance figure.
The one real weakness is that the document under-uses the JD's own
language for the responsibility it matches best. R2 - ingestion,
transformation, storage and lifecycle management of large datasets - is
the single core responsibility scored Direct and is close to a plain
description of SW-1 and SW-7, yet none of those four words appears in
the resume body. Two related misses: "distributed systems" never appears
as a phrase even though Q4 is a required qualification scored Direct, so
the nearest exact match sits on an eight-year-old Vizrt bullet rather
than current work; and "distributed databases" was called a Direct
preferred hit in Phase 0 but is only ever implied through product names.
Three Tier 1 fixes, all honestly available, worth roughly +4.
Recorded what must NOT be added, since the obvious way to raise keyword
coverage is the wrong one: "scalable" is an unverified scale claim that
ai_fingerprint_rules forbids, and SDK, high-throughput, data-intensive,
simulation, model development and self-service have zero canonical
evidence. Their absence is correct, not a defect.
Flagged the headline "Production Data Platforms on AWS" as Tier 2. It
carries no ownership verb so it is not a scope violation, but it is the
most prominent line in the document and leans toward exactly the
platform framing R1 is a gap on - the likeliest opening challenge from a
technical reviewer.
Scored against critique_framework.md section 9. The SKILL.md
8-dimension table with "Publications 10%" is a stale CV-era scheme and
was not used, the same conflict class as its "all bullets 2L" line.
The verdict does not change the decision. Document quality is not the
constraint here: the working model is still unresolved and the channel
is still cold. 86 is not submit-ready.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
13 bullets, validator PASS, compiles to exactly 2 pages, rendered PDF
inspected clean - no clipping, orphans, header wrapping or bad breaks.
Built at the user's direction with the working-model question still
open. That block stands: if the Zurich seat is 5-day onsite, this
package does not get submitted. Recorded in the session file rather than
quietly dropped.
Three Phase 2 decisions the user left open, taken and marked reversible.
VZ-1 is in - it is the only canonical evidence pairing C++ with a
distributed backend and this JD names both, hedged verb preserved. A
Personal Project section carries PP-1, labelled three separate ways and
with no performance figure of any kind; PP-2 appears as supporting
coursework, never as a credential. Generali is kept against the Phase 1
"omit" recommendation, because dropping it opened an unexplained
May 2015 to Jun 2017 employment gap that costs more than the weak bullet.
Cadence sits at 7/12 (58%), above the section 7a guide of ~50%, down
from 75% after reshaping two bullets. The rest are legitimate technology
enumerations - "Oracle, Kafka, Python and Teradata", "Elasticsearch,
Logstash, Kibana and Kafka". The checker added yesterday cannot tell a
rhetorical rule-of-three from a list of tools actually used, which is a
real limitation of the pattern rather than a defect in this document.
Cutting them would delete accurate ATS-relevant names, which section 7a
itself forbids. Left deliberately.
Also walked into the trap the SBB .tex header warns about: the first
draft named the excluded non-canonical tools in a comment explaining why
they were excluded, and the validator errored on all three because it
scans raw file text. Comment rewritten without naming them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
The user asked whether Citadel has a hybrid model. The JD is silent:
grepping the verbatim posting for hybrid, remote, onsite, in-office and
days per week returns zero matches. Phase 0 scored practical constraints
5/5 on an unexamined assumption that Zurich was workable. Withdrawn.
Ken Griffin is on record that Citadel returned to the office five days a
week, early and deliberately, and calls it his most important leadership
decision of the past seven years - framed as core apprenticeship culture
rather than a policy setting. The only contrary evidence is an anonymous
2022 Fishbowl post about Operations claiming three days minimum: four
years stale, wrong division, contradicted by the founder on record. No
official Citadel Securities Zurich policy statement could be found.
If five-day onsite holds for Zurich, it means roughly 2 to 2.5 hours of
commuting per day from Bern, or relocating - which breaks both the
Bern-based, 2-3 day hybrid bar in user_comp_bar and the "keep a Swiss
home base" constraint in user_international_mobility.
Practical constraints 5/5 -> 2/5 (language and authorization stay clean:
English working language, EU citizen, Swiss B permit). Evidence Fit
74 -> 71, still Adjacent but mid-band rather than a point below Core.
More consequential than the number: critique_framework.md treats an
unresolved location or relocation constraint as a NO-GO trigger, so the
qualifications gate passes while the practical gate does not. Phase 2 is
blocked pending an answer.
This has to be asked, not researched - the same class of question as the
SBB Anforderungsniveau K band. Recording the sequencing explicitly,
because SBB was submitted with its equivalent question still open and
that question is now a live screening topic instead of a settled fact.
Repeating it here should be a deliberate choice, not an oversight.
The Phase 1 bullet plan is written and unaffected: 11 recommended, 14
with options, budget PASS. Two items still open - the Vizrt VZ-1
C++/distributed bullet the user previously skipped on Snowflake, and
adding a Projects section for PP-1.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
Evidence Fit 74/100, Adjacent at the top of the band, hard gate PASS
with one flagged risk. Phase 1 not started - user asked for Phase 0 only.
The stack alignment is the best in the log: zero Gaps across all seven
minimum qualifications, and all five preferred technologies (Kafka,
Kubernetes, Spark, Airflow, distributed databases) are Direct and
production-current.
What holds it at 74 is that the title-defining function splits. R2 -
ingestion, transformation, storage and lifecycle management of large
datasets - is Direct and describes SW-1/SW-7 almost word for word. R1 -
design and build the distributed research platform itself - is Adjacent,
the same authoring-vs-building shape as Aker BP, and it cannot be
written around: Scope Discipline forbids pairing a full-ownership verb
with an org-scale object, so it can only be honestly bridged.
Recorded the classification conflict rather than hiding it.
application_strategy.md says a defining responsibility that is only
Adjacent makes the role Stretch; that strict reading applies to R1. It
is filed as Adjacent because R2 is equally title-defining and is Direct,
and because Aker BP set that precedent at 79. Both cohort slots are free
either way, so the label changes nothing operationally.
Corrections the user was right about, now recorded so a later run does
not repeat them: Go is NOT a gap, because Q3 is disjunctive and Python
satisfies it outright; and a production Kafka/K8s/Spark/Airflow estate
IS a distributed system, so Q4 is Direct rather than a stretch.
Two traps written down explicitly. The $175,000-350,000 base range is a
US pay-transparency disclosure under NY law and says nothing about the
Zurich figure - it must be asked, not assumed. And the sibling reqs that
web search surfaces as closer title matches (Research Platform - Data
Platform Engineer, Research Platform Infrastructure Engineer) are dead:
404 and Cloudflare-blocked, absent from the live board. Those Built In
listings are stale.
JD saved verbatim via Playwright; WebFetch returns 403 on this host.
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
An audit of all 18 packages in output/ (368 bullets) found one rhythm
running through every document: 45% of bullets used the same "X, Y and
Z" triple, and the SBB package reached 85% - 11 of 13 bullets, every
Swisscom and Bosch line - plus two adjacent bullets both opening
"Build and...".
This is a style finding, not a truth finding. Every bullet was accurate.
The corpus is clean on the axes that actually signal generated text: no
AI vocabulary (0 hits for leverage/spearheaded/robust/passionate and 56
others across 35 documents), prose em-dashes at 0.08/bullet, and PDF
metadata carrying nothing but MiKTeX pdfTeX with empty Author/Title
(0 AI tokens and 0 generator-term leaks across 69 PDFs). What is left is
cadence: ten bullets sharing one three-beat rhythm read as machine-made
even when nothing in them is false.
Guarded deliberately so it cannot do harm. cadence_checks() emits WARN
and never ERROR, the critique deduction caps at 1 point, and both the
reference and the docstring state that no claim, scope or hedged verb
may be bent to satisfy rhythm. An anti-monotony rule with teeth would be
worse than the problem - it would pressure a future run into loosening a
scoped claim to vary a sentence.
Thresholds are calibrated on the corpus, not guessed. Position length
spreads are bimodal (1-9 words, then 15-17), so the check flags a spread
of <=4, the tight tail at ~28% of positions; the first draft used <=6
and flagged the median. Also fixed the opening-verb extractor, which
read \textbf{Owned ...} as the word "textbf" and produced six false
positives on one document. Per-package warnings now run 0-3.
Docs: resume_reference.md 7a + verification step, critical_rules.md 7a,
critique_framework.md mechanics row, CLAUDE.md corrections log.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
application_cohort.json was caught by the blanket job_scout/state/*
ignore, so the cohort slot record (core/adjacent/stretch mix and every
application's fit class, evidence fit, hard gate and outcome) existed
only on one machine. decisions.json escaped the same rule only because
it had been tracked before the rule was added - an accident of history,
not a decision.
Both files are durable application history and belong in the repo. The
churny scan state (seen_jobs.json, last_scrape.json) stays ignored, as
does reports/ - all of it is regenerable by re-running the scout.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
Application log updates from the 2026-08-24 scout run and the days after.
SBB Data Engineer (Asset Management, Bern, job 103755): SUBMITTED
2026-08-25. Consumes Core slot 2/7 in evidence-first-2026-01. Submitted
with the Anforderungsniveau K comp question still unresolved, against
the channel plan and the critique's closing advice, so the K band, the
design-vs-build-and-run scope question and the Kidz Care rate are now
screening-call topics rather than pre-cleared facts. Next action changed
from "call then submit" to "prep an interview brief". The stale pointer
to an /edit-resume Tier 1 fix is removed - that fix landed 2026-08-21.
Snowflake Sr SWE Enterprise (Observe): CLOSED - NO RESPONSE at 80 days
silent, past the longest real disposition in the log (Equinor, 79).
Recorded as closed_no_response, deliberately NOT as a rejection: none
was ever received, so it must not enter rejection statistics or read as
a document-quality signal. Pre-dates the cohort, so no slot is freed.
An ~86/100 package - the highest in the log - drawing no reply at all is
the strongest evidence yet that the constraint is channel, not documents.
Unblocks the Observe Metrics Platform req, which was paused only to
avoid stacking a second cold application on the same org.
Norway lane: corrected a false "closed" framing that had propagated into
the Equinor row and was used to dismiss Norwegian rows in the scout
readout. The lane is OPEN but selective. The gates are norsk working
language and NO security clearance - NOT compensation: the 180k bar is
CH-only and explicitly does not apply to Norway. Equinor, Telenor and
NATO JWC stay in the scout.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MHtzyTKBcg6BWhD5qFegtK
Critique per critique_framework.md: hard gate PASS, Evidence Fit 79/100
(Core, unchanged — a document cannot move candidate-role fit), Document
Quality 90/100, Channel Weak but uniquely upgradeable. Claim audit clean
across all 13 bullets; no Tier 1 truth findings.
The skill's own spec asks for a single 8-dimension score including a
Publications weight, which contradicts critique_framework.md and CLAUDE.md
("never collapse them into one optimistic score"). Followed CLAUDE.md and
recorded the conflict — the skill definition looks stale relative to the
July rewrite.
Tier 1 finding and fix: the JD names Abfrageleistungsoptimierung in a
required line and the document had zero occurrences of "query", despite
canonical support sitting in BS-2's 3L variant. BS-2 now reads "tuning
query performance for analysis teams working with…" and the Pipelines
skills line gains "query performance tuning". Also surfaced "agile" in
SW-3, whose canonical 2L variant records "in an agile DevOps team" and
whose absence left the JD's second required line unaddressed.
Deliberately NOT applied: the API term. claims.json SW-7 records only
"onboard source systems" and SW-2's ingestion is Oracle to Kafka, so
adding it would be vocabulary substitution rather than evidence. Left
open pending user confirmation.
Re-verified after the edit: validator PASS, forbidden-token scan PASS,
compile PASS at 2 pages, both pages re-rendered and inspected. Longest
bullet 34 words, still under the flag threshold. Document Quality 90 to
92 on relevance/terminology.
Still not submitted — held pending the Anforderungsniveau K comp answer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WpMtWFtYkGkjSMShuTgXiB
13 bullets, 6 skills lines, International Tech profile in English with an
explicit Languages line naming German as mother tongue (user instruction).
Framing leads on operate-and-own — Component Ownership with on-call and
governance answers the JD's "Betrieb" and "Verantwortung für Qualität und
Nachhaltigkeit" — and carries the Bosch industrial sensor-data thread
(defect records, wafer inspection images, PCM parameters) against the
role's vorausschauende-Wartung purpose. GenAI/LLM material dropped.
Two deviations from the approved 11-bullet plan, both deliberate:
Fraunhofer keeps one combined bullet rather than zero, because dropping the
role opens an unexplained 21-month gap between Vizrt and Bosch; and BS-1 is
added as a fifth Bosch bullet, HIGH priority in the bundle and directly
relevant to an employer running a 24/7 network.
Kept the guarded page break after testing the natural flow: without it page 1
fills, but two Bosch bullets land on page 2 with no visible employer header.
Hierarchy beats page-fill, and page-fill quotas were removed on 2026-07-27.
Gates: canonical validator PASS (0 warnings), forbidden-token scan PASS (none
of the seven non-canonical JD tools present), scope-discipline scan PASS,
compile PASS at exactly 2 pages, both pages visually inspected. SW-1 and SW-7
bullets also read by eye against their claims.json forbidden lists.
Not submitted — held pending the Anforderungsniveau K comp answer.
Also fixes a KB defect found during generation: bundle_data_engineer.md §S5's
cover-letter hook read "leading the migration of our legacy stack", a
full-ownership verb on a company-scale object that contradicts claims.json
SW-1 and its own narrative thread. Rewritten, warned inline, and recorded in
the CLAUDE.md corrections log. The other four bundles are unchecked.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WpMtWFtYkGkjSMShuTgXiB
Job ID 103755, posted 2026-08-20, surfaced by the 2026-08-21 scout run.
Evidence Fit 79/100 (Core, lower end), hard gate PASS. The title-defining
capability — build and operate cloud data pipelines and governed data
products — is Direct and current, unlike the AWS FDE NO-GO.
Best practical fit in the log: Bern-based, German-language, no relocation,
EU citizen plus B permit. Real strength cluster for the role's vorausschauende
Wartung purpose: Bosch fab sensor/process data, ELK/Kafka anomaly detection,
containerized ML inference, plus the condition-monitoring thesis.
Two open risks, both level and comp rather than fit: Anforderungsniveau K is
a capped GAV band whose figures are not public and must be asked, and the seat
reads lateral or below current Staff plus Component Owner scope — the same
shape declined at BKW. Level and ownership scored 9/15 for that reason.
Seven named tools are non-canonical (Snowflake, dbt, Argo Workflows, Helm,
Power BI, Spring Boot, Angular); survivable only because the JD prefixes the
stack list with "z. B.". Recorded as an accepted ATS risk, with a Phase 2
gate to keep all seven out of the generated document.
Cover letter: NO — SBB explicitly waives it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WpMtWFtYkGkjSMShuTgXiB
Boards rewrite their own URLs between runs, so a decision recorded under one
URL was lost the next time the same req appeared. Amazon truncates the job
slug to a fixed width and that width changed, which silently resurfaced the
already-skipped Principal Delivery Consultant req at score 7 in the 08-21 run.
Adds _decision_key(): "<host>#<req-id>" via per-board ID patterns, falling
back to a normalized path. Host-scoping is deliberate — a wrong ID match can
never mark a role decided at a different company. Also collapses genuine
duplicates: Workday's /apply suffix, PostFinance's per-locale URLs, Apple's
?team= and /locationPicker variants, the same NVIDIA req under two locations.
The log file stays URL-keyed (it is hand-edited and the URL is the readable
part); lookups go through index_decisions(). --decide now updates an existing
entry when the job is already logged under a drifted URL instead of adding a
second row.
Verified against all 955 URLs in the decision log plus the 2026-08-21 report:
4 groups collapse, each inspected and correct, no false merges.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WpMtWFtYkGkjSMShuTgXiB
Probed CERN, Swiss Post, NATO NCIA and SIX Group. Two actionable results:
- CERN: SmartRecruiters board "CERN", 66 live Geneva postings with real
engineering (Deep Learning Developer, ML Engineer NGT, Full-stack SWE, Data
Storage R&D Engineer). Added — the adapter already existed. Two standing
caveats recorded in the entry: Geneva is ~1h50 from Bern so a 2-3 day hybrid
is not viable without relocation, and CERN grades sit below the 180k bar.
- NATO NCIA: nato.int is Cloudflare-blocked to headless browsers. Moved to
MANUAL_CHECK beside Oracle, with a note that NCIA is a separate pipeline from
the JWC Stavanger reqs the `nato` adapter returns.
Not added, documented instead:
- Swiss Post: real board is job.post.ch (SuccessFactors CSB), scrapable via
a[href*='/job/']. The only tech role on page 1 was served from the PostFinance
sub-board, which is already covered. Paginated, so this is page-1 evidence
only, not a whole-board verdict.
- SIX Group: SAPUI5 SPA backed by a legacy DWR RPC endpoint
(careerJobSearchControllerProxy.getInitialJobSearchData.dwr). Scrapable in
principle, brittle in practice; job rows did not render into readable text.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
User declined at the gate: 69 is not a strong enough fit to spend a package on.
Consistent with revealed behaviour — every application actually submitted scored
77.5-89; nothing below 75 has ever been sent.
No cohort slot consumed (Adjacent stays 1/2). No bullets written, no resume, no
cover letter. Phase 0 analysis retained in the session file in case a
better-fitting AWS req appears.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Verbatim JD (amazon.jobs 10504263, posted 2026-08-17) retrieved via Playwright
and stored with the session.
Evidence Fit 69/100, Adjacent. All four Basic Qualifications are Direct, so the
minimum-qualification gate passes. Disclosed rather than buried: forward
deployment is the title-defining capability and is a Gap - SW-4, VZ-1 and
global_forbidden_output_patterns all forbid "customer-embedded delivery", and
application_strategy lists strategic-account FDE as a Stretch target. Scored
Adjacent because AWS's published minimum bar is pure software engineering and
the org is seven weeks old, hiring builders at volume.
Assets: AWS is the evidenced cloud (claims.json forbids GCP output), and Kiro -
named in the JD - is in the verified GenAI toolchain.
CL decision YES. Would consume the last Adjacent cohort slot.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Pipeline:
- SDU PhD: closed, not proceeding. Prof. Jørgensen never replied to the
2026-08-02 enquiry; user dismissed it two days before the 2026-08-20 deadline.
- Google FDE III GTM: closed, not proceeding after two Google rejections. The
bar for a third attempt is now a very strong fit; this was Adjacent with an
agent-orchestration gap. Frees the Adjacent cohort slot (back to 1/2).
JD triage — 5 pulled and read, 1 survives:
- SHORTLIST AWS Senior FDE Zurich: all four basic quals pass and AI/agentic is
preferred-only, unlike the Google Senior Staff FDE where vector DB + RAG were
minimum quals. AWS is the evidenced cloud; JD names Kiro, which he uses.
- SKIP AWS ProServe Bern (clinical pharma domain gate), AWS Principal Delivery
Consultant (8+ yrs architecture leadership + VP+ exec comms), GitLab FDE EMEA
(Ruby/Go hard gate), MS 200044133 (title says FDE Data Scientist, quals are
BGP/MPLS/SD-WAN network engineering — mislabelled req).
Roster:
- Dropped Louis Dreyfus: 300 roles, all Brazil/India/Bulgaria ops, zero Swiss.
- Kept MET Group: its "0 eligible" was a filter artifact; exclusion mode reveals
real Baar/Zug commodity-trading roles.
- Kept Equinor/Telenor/NATO. Memory corrected: the Norway lane is selective, not
closed — the user still wants very good Norwegian roles.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Scraper fixes:
- Roche: new fetch_phenom adapter (Phenom refineSearch). The old playwright scrape
of ?locationsearch=Switzerland harvested recommendation-widget cards (Shanghai,
Kyiv, Bogota) while the page reported no-results. 0 -> 88 CH-eligible roles.
- Apple: dropped default_location "Switzerland", which relabelled US "Various
Locations" postings as Swiss (84 phantom CH rows over 4 runs). Now honestly 0.
- Meta: NOT broken — metacareers reports "1 Items" for Zurich. Comment added so it
is not "fixed" again.
New boards:
- Amazon/AWS (fetch_amazon): 32 CH roles incl. a Zurich AWS FDE req and a Bern
ProServe Cloud Architect. AWS is the evidenced cloud; claims.json forbids GCP.
- Axpo (teamtailor via base_url + pagination): 461 roles, opens the energy lane.
Locations read from schema.org jobLocation with ISO alpha-2 expanded, so
Madrid/Milan/Warsaw roles are not marked Swiss. Telenor benefits too.
Title filtering now has two explicit modes. Inclusion allowlists fail closed and
hide unanticipated good-fit roles, so they are now used only where volume forces
it (>~200 roles). Everything else uses the shared, board-agnostic
NOISE_TITLE_EXCLUDE, which fails open and leaves the final call to the scorer and
the reviewer. Palantir stays unfiltered per its existing documented rationale.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Analysis of the 34-company roster against 4 recent runs and 253 decisions.
No COMPANIES changes.
Findings: Roche/Meta/Apple silently scrape garbage (non-CH rows, 1 intern,
internships-only) — same silent-failure class as the dbt Labs 404; seven boards
have no _title_filter; MET+LDC yielded 0 eligible across 1,464 scrapes.
Verified adds: Amazon/AWS (32 CH roles incl. a Zurich AWS FDE req, and AWS is
the evidenced cloud where GCP is forbidden) and Axpo (Teamtailor custom domain,
400 roles incl. Forward Deployed AI Engineer, opens the named energy lane).
Verified negative: Coinbase proper has 2 EMEA-remote non-eng roles; the
remote-EU data-infra tier is geo-banded below the comp bar.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The board never carried Coinbase's own roles — only Getro portfolio companies.
Across ~3 months it surfaced 3 distinct CH-eligible roles, all from Ashby
(recruiting software, ruled out 2026-07-28) and all decided "skip". Since the
Ashby exclusion it has returned 0 jobs on every run, surfacing a false error
line in each report.
Entry replaced with a "Dropped:" note per the file's convention. fetch_getro is
retained (no caller) since it generalises to any Getro collection id.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The "dbtlabsinc" Greenhouse board started returning 404, silently dropping
a target company from every scan. getdbt.com/careers now redirects to a page
serving Fivetran's Greenhouse listings (gh_jid links) — the two companies
merged, so the roles live on board "fivetran" (239 jobs, 10 title matches).
Renamed the entry id dbtlabs -> fivetran and cleared the stale seen-state key.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Equinor Manager / AI Architect (JR106747): REJECTED 2026-08-14, no interview.
79 days from submission (2026-05-27) to disposition, the longest turnaround
recorded. Pre-dates the evidence-first-2026-01 cohort (started 2026-07-27),
so it does not free an Adjacent cohort slot. Norway lane now reduced to the
open Aker BP application.
Microsoft Senior SWE, ISE Zurich (200040836): REJECTED 2026-08-11, no
interview, ~39 days. The separate Principal FDE req 200043897 is a different
pipeline and remains open.
Also carries forward uncommitted work from the 08-07 / 08-11 sessions:
- Google FDE III, Cloud GTM (78350205438567110) Phase 0 session file and
scraped JD; Adjacent, Evidence Fit 67/100, hard gate PASS.
- Scout skip decisions: BIS jr100411 (platform admin, not data eng), two
Google reqs gated on a C++ minimum qualification, and the FDE III
Generative AI sibling req confirmed taken down.
- settings.local.json: allow the job_scout venv python invocation.
Every closed application to date has been Channel Weak with no interview.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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>
Tighten eligibility to CH/DK/Europe-remote (Equinor Norway-only, NATO DK+NO), add Taleo adapter, parallel API fetches, atomic state writes, and empty-board warnings. Log NATO AI Engineer as applied and Architect as skip.
Add BIS Basel and NATO JWC Stavanger application output (resume/CV, cover letter,
session file, critique) plus the Bundeswehr experience file backing the NATO
package; update session status tables in AGENTS.md/CLAUDE.md; record two new
Bash allowlist entries for the job scout venv. tmp/ is scratch working files,
now gitignored.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Equinor (Workday) and BFH (Bern, playwright) wired up as new sources; artificial
intelligence/analytics-engineer keywords added so int'l-org listings (BIS) stop scoring
as noise. certifi-backed SSL context works around Workday certs failing verification
against a stale Windows root store.
location_matches() previously treated "remote + any single named EU country" as
CH-eligible, which let single-country residency-locked listings (e.g. Databricks'
"Remote - United Kingdom") pass as if they were open EU-wide. Now requires either a
pan-regional keyword (Europe/EMEA/Global/Worldwide) or 2+ distinct countries named.
Denmark is carved out as its own always-on relocation category (on-site or remote,
any company), replacing the previous Databricks-only opt-in.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Prep materials for the Google Senior Data Engineer (Merchant Data
Science) process after passing the Hiring Assessment 2026-06-20:
interview prep brief + STAR story bank alongside the session files.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds positive keywords (forward deployed, field engineer, resident
architect/engineer, customer engineer) and title-filter entries so
vendor deploy-to-customer roles surface — the lane matched to
user_international_mobility (intl scope from Bern, travel-OK, no
relocation). Bare "resident" stays title-filter-only to avoid
re-opening the pre-sales-SA overscoring trap.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
BS-5 was a thin "C# extensions" note; now records the verified reality
(Playwright scrape of community.spotfire.com, 2026-06-15): co-ownership
of Bosch's Spotfire analytics platform and the TAF 2022 co-presentation
"Understanding Spatial Variability in Semiconductor Products with
Spotfire Map Charts". Reframing bumped to MED-HIGH for BI-bridged JDs;
new 1L/2L variants. First consumed by the Microsoft ISE package (merged
BS-3+BS-5 bullet; TAF talk reserved for the CL).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two applications sent and finalized on 2026-06-15:
- Google - Senior Data Engineer (Merchant Data Science, Zurich), 85.5/100.
Tier-1 scope fix + both Tier-2 polish applied: re-scoped the Swisscom
migration claim in resume B2 + CL P2 (Scope-Discipline), added project-
delivery vocab (B4), and JD-exact 'distributed data processing' (B5).
- Kraken (Payward) - SRE, AI Agents (remote, CH-eligible), 87.2/100.
Finalized as-is; crypto-native + production-ML edge, honest infra gaps.
Logs both as 'applied' in job_scout/state/decisions.json and flips their
CLAUDE.md Active Sessions rows to SENT. Open item for both: confirm level
and comp clear the 180k+ all-in bar at the recruiter stage.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Records the 2026-06-06 application (Observe by Snowflake, ~86/100) in the
decision log so it stops resurfacing as a new match.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Finalized package retained for reuse; role closed (Cisco scrape 2026-06-02:
not on board, Recruitee link dead). PDFs/build artifacts gitignored.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add the JD Integrity section to shared_ops (no reconstructed/inferred JDs;
WebFetch is JS-blind on careers boards; scrape JS-gated postings via the
job_scout Playwright venv; STOP and ask if the real text is unobtainable).
Wire the rule into /make-resume and /critique. Allow cisco/bkw WebFetch.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Track per-job decisions across runs so we don't re-evaluate roles.
- state/decisions.json (keyed by URL: company/title/decision/note/date), now
git-tracked while seen_jobs.json stays local
- --decide "<url>" <status> [note] records a decision; --hide-decided gives an
undecided-only view; report tags each role inline with its decision
- usage docstring updated
- seed 18 decisions (9 shortlist, 7 skip, 1 paused, 1 maybe); flags Google Staff
FDE GenAI as the paused prior session
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add positive keywords for the core DevOps/data-platform/cloud lane and
technical-architect titles (cloud/data/platform architect) that previously
scored 0 and surfaced only via the generic "senior" term. Rebalances rankings
so hands-on engineering roles aren't buried under pre-sales "Solutions
Architect" + crypto keyword stacking.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Pass 2 critique 80.0/100 (honest-stretch reach app, converged near ceiling).
Includes accuracy fix: Swisscom Data Mesh reframed from sole-build to
company-wide migration contributed to (hedged verbs); orchestration +
reference-architecture surfaced; removed -ing bullet trailers; CL trimmed
to 300 words with matching ownership hedge.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Two scoring bugs inflated the rankings:
- Substring matching: 'rag' matched sto[rag]e / tet[rag]on, 'intern'
matched inte[rnal], 'lead' matched [lead]ership. Roche's staff
restaurant and Cisco Tetragon roles scored as fits. Now keywords must
not be flanked by alphanumerics (c#/.net/c++ keep matching).
- Body boilerplate: every Kraken post mentions crypto/blockchain/trading,
so sales/PM/design roles scored as high as engineering ones. Title
matches now score full weight, body-only matches half (min 1);
negatives still count fully anywhere.
Strong-fit count drops 13 -> 5; the bogus 'intern' negatives and rag
false positives are gone. Shared _kw_in() also backs the title filter.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Dennis is interested in trading roles (energy/finance/crypto), but the
scorer had no terms for them, so trading-tech roles scored 0 and got
buried. Add a trading/quant cluster (trading, trader, quant, market
data, low-latency, FIX protocol, brokerage, commodity, execution) plus
singular "solution architect" and adjacent stack (C#, .NET).
Bitcoin Suisse's Solution Architect Trading now scores 4 (medium) on
title alone, up from 0.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Bitcoin Suisse's careers page is a JS-rendered Next.js SPA, but the
underlying onlyfy.jobs ATS exposes a plain HTML job list at
candidate/job/ajax_list (title + location per card). Add an onlyfy
adapter that parses it — no headless browser needed. Surfaces the 3
current Zug roles correctly (CH filter + pre-dedup location filtering
keep the Swiss posting over its Bratislava/Copenhagen cross-posts).
MANUAL_CHECK is now empty (all 22 target companies automated), so the
report's manual-check section is suppressed when the list is empty.
Coverage notes updated.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
AMINA Bank rates poorly on Glassdoor and isn't worth tracking. Remove
its manual-check entry, leaving Bitcoin Suisse as the only manual source.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Coinbase cut ~14% of staff and is not actively hiring; remove its
manual-check entry. The getro Coinbase Ventures source covers portfolio
companies only, not Coinbase itself, so Coinbase is now off the radar
until they reopen. Now 2 manual sources (AMINA, Bitcoin Suisse).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add a getro adapter (POST JSON search API) and wire up the Coinbase
Ventures portfolio talent network (collection 1625), CH + eng-title
filtered. Note this covers portfolio companies (Ashby, Notion, VALR,
World, ...), not Coinbase itself, which doesn't list on its Ventures
board — Coinbase-the-employer stays in MANUAL_CHECK.
Also clean up stale comments: drop Sonova (MedTech, off-thesis, dead
scrape) from MANUAL_CHECK, remove the dangling BIS comment now that
BIS is automated via rss, and refresh the adapter-coverage notes and
module docstring to the current 21-automated / 3-manual state.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Generic RSS/RDF feed parser (handles RSS 1.0 namespaced items + RSS 2.0)
- BIS (Bank for International Settlements, Basel) promoted from manual-check
to an automated source via its vacancies RSS feed — commutable from Bern,
salary net of Swiss tax. Scout now 20 automated + 4 manual-check.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
