Build Rundown & Issue Map

A daily research brief for HR and organizational leaders, produced by a local AI model with one paid call a week for the part worth paying for.

The economics are the design constraint. The daily loop runs on hardware already bought, at no marginal cost. The weekly proposal call costs about three cents.

Twelve scripts, each writing a file the next one reads. A failure stops at one step instead of corrupting the next.

Grouped by failure class. Each class produced one structural change rather than another instruction.

Class A The local model doesn't work

Three days were lost before any pipeline work. Each symptom was a misconfiguration, not a capability gap.

Lesson: only a real completion is a valid health check. A status code of 200 proves nothing.

Class B Asking the model to hold a fact a script already had

The load-bearing class. Every defect worth fixing was an instance of this. The model was being asked to compose facts when the scripts already knew them.

The rule: A fact a script holds, the script writes. The model copies forms, never composes facts.

Class C Contradictory instructions

Two documents said different things. The model obeyed one, was blamed for the result, and the cycle repeated.

Fix: Bump the skill version, strip stale instructions, and provide filled examples the model can copy verbatim.

Class D Declared inputs, nothing wired up

Four files were referenced as though in use. No script ever sent them to any model call.

• Focus file — the week's editorial question was never included. The brief invented its own and ignored relevant papers.
• Audience file — given to one script but not the other. The two components disagreed because one of them didn't know who the blog was for.
• Evidence types — eight controlled labels, never sent. The brief wrote the rating in the evidence slot instead.
• Covered file — empty. No script appends to it.

Class E Checks that trusted their subject

The subtlest class. The first fix read the ratings path from the brief's own pointer line — written by the model. A check taking its reference from the text under check verifies nothing.

Fix: A verifier must not take its reference from its subject. Where it cannot get an independent reference, it should report that it could not check.

Class F Supply, diagnosed by evidence

The inbox went thin and there were three plausible stories: a cap, dedup suppression, or genuinely quiet feeds. Each was ruled in or out by evidence:

• Empty holding list → ruled out the cap
• Seen list at 32 entries → ruled out suppression
A 14-day dry run returning 87 candidates from 18 feeds → proved it was a two-day lookback against a collector younger than the archive

The clustering on Sep 20 is not thrash. It is what happened once the root cause was understood as structural rather than instructional: each fix exposed the next thing the model was still being asked to compose.

brief-lint.py: 145 lines → 571 lines, 31 error conditions and 17 warnings. The pipeline itself is 5,887 lines across 12 scripts.

Everything compresses to a single sentence, now the first hard rule in the project docs:

Facts the scripts hold, the scripts write. Ids, ratings, the count, the rubric version — all script-written, copied by the model, never composed by it.

• arXiv returns 406 — curl probes drafted, never run open
• HBR TLS refusal — publisher refusing non-browser clients open
• Moonshots feed works (HTTP 200, 297 entries), but the focus filter blocks it investigating
• covered.md — placeholder with header, no actual entries yet partial
• The weekly post — picks a question but nothing drafts the post partial