Every decision here exists for a reason. Here they are.

The decision

A single batched keep/drop triage runs against your targets before any job gets a full evaluation.

Why

Full evaluation is the expensive call. Most discovered jobs are not worth it. A wrongly-dropped job at the gate never gets a second look, so the gate is deliberately biased to keep — recall over precision. A wrongly-kept job just costs one extra, cheap evaluation.

The decision

Location and salary deal-breakers are enforced by deterministic Python, before the AI gate, and re-checked after evaluation.

Why

A binary deal-breaker is not a scoring problem. In QA, a hard onsite constraint once got diluted into a passing weighted score. Weighted scoring is the wrong tool for a yes/no rule. Unknown or missing data never rejects — only a confidently-known violation does.

The decision

The evaluate stage writes a compact JSON judgment — score, recommendation, three strengths, two weaknesses, a fit paragraph — and nothing longer.

Why

This is CareerOS's deliberate departure from long-form evaluation reports. The JSON is the entire output. The daily report renders from it with zero additional AI cost, and the fit paragraph doubles as the spine of the cover letter.

The decision

Your profile facts and each job's evaluation JSON are the only sources of truth. Every other artifact is a derivation.

Why

Facts and judgments should be generated once and reused, not recomputed. This is 'selector, not writer' applied everywhere — a resume selects from your facts, a cover letter builds on the evaluation's fit paragraph, a report renders the JSON.

The decision

Each AI output is cached on a fingerprint of everything that could change the answer: the job's content hash, your profile version, and the active prompt version.

Why

The KPI is interview-worthy jobs per dollar. Re-running a day where nothing changed should cost nothing. Because the prompt version is inside the cache key, editing one prompt busts only that stage's cache — not the whole pipeline.

The decision

Prompts are versioned files, selected per stage in config. Superseded versions stay in the repo and still run.

Why

The prompt version is part of every cache fingerprint, so versioning is what makes targeted cache invalidation possible. Keeping old versions live means a run pinned to an earlier prompt still reproduces exactly, which matters for auditing why a past evaluation scored the way it did.

The decision

The profile has an integer version. Bumping it invalidates cached evaluations, resumes, and cover letters that depended on the changed facts.

Why

When your facts change, every derivation that depended on them should recompute against the new facts. The version flows into every cache key, so a bump is how stale artifacts get regenerated. Logistics answers are deliberately excluded — no scoring or writing stage reads them, so they don't need to bust anything.

The decision

A deterministic step confirms every resume bullet and the summary verbatim-match a fact in your profile. The finalize step refuses to cache a resume that fails it.

Why

The resume stage is a selector, not a writer. A guarantee is stronger than an instruction — the model is told to select verbatim, and then a mechanical check enforces it, rather than trusting that it did.

The decision

A deterministic lint flags em-dashes, a banned-vocabulary list, and negative-parallelism constructions across every generated artifact.

Why

Style guidance in a prompt is a suggestion the model can drift from. A mechanical check is a backstop that catches the common AI tells regardless of which host model wrote the text.

The decision

The long-form deep report is generated only when you explicitly ask for it, per job, with a separate command.

Why

It is the one deliberately long-form artifact, and the only one, precisely because it's opt-in. Most discovered jobs never become an actual application, so generating an expensive report for every one would be exactly the wasted spend the architecture exists to avoid.