ADR 0014 — Per-job artifacts in object storage + a measured humanizer pass

Context

Two gaps between “pack card in Slack” and “application submitted”:

1. The pack contained tailored bullets but not a full tailored resume — Kelyn still had to hand-edit the master resume per role, which is the slowest step of any application.
2. LLM-drafted prose carries recognizable AI tells (“I am excited”, “leverage”, em-dash cadence). Recruiters see hundreds of these; the voice is a real response-rate risk.

And the materials lived only as files on one laptop’s drafts/ dir — nothing a phone could open from the Slack card.

Decision

Three pieces, composed into one pipeline (drafting/pipeline.py) that every surface (CLI draft, MCP draft, autopilot) calls — the flow can never drift between surfaces:

draft -> humanize -> tailor full resume -> upload artifacts -> record

1. Full tailored resume per job (resume_tailor.py, resume-v1): master resume + JD + qualification + keyword targets -> a complete reshaped resume markdown. Same facts, different emphasis; inventing experience is forbidden. Keyword coverage is scored on the output.
2. Humanizer (humanize.py, humanize-v1): a deterministic AI-tell detector (banned-phrase list, em-dash density, “not only…but also”) plus an LLM rewrite in Kelyn’s own voice (his resume is the style sample). The detector ARBITRATES: if the rewrite scores worse than the original, the original ships — we never pay for a worse draft. Before/after voice scores ride on the pack card. Anti-tell rules are baked into the resume-tailor prompt instead of running a second pass over the resume (one call, not two).
3. Artifact store (artifacts/, MinIO/S3 via env): one object prefix per job ({id}-{company}/resume.md, cover_letter_pack.md), presigned links (7d) on the Slack card — click, tweak, submit from anywhere. Karani runs its OWN MinIO in compose on :9010 (the DataQRL stack owns :9000 on this machine); MINIO_ENDPOINT points anywhere S3-compatible. Best-effort throughout: unconfigured or unreachable storage degrades to files-on-disk, never fails a draft.

Cost shape: a pack is now up to three LLM calls (draft, humanize, tailor). KARANI_HUMANIZE and KARANI_TAILOR_RESUME each toggle one off; autopilot’s per-run and daily budgets bound the total regardless.

Consequences

• Positive: “tweak and submit” becomes literal — every pack carries a complete role-specific resume and letter behind two links.
• Positive: voice quality is measured, not vibes: the tell detector gives every pack a before/after score, and funnel_stats can later correlate voice scores with response rates.
• Negative: ~3x LLM cost per pack. Bounded by autopilot budgets and per-flag toggleable; qualification (the volume path) is untouched.
• Negative: presigned URLs expire (7 days) and the local MinIO is laptop-bound. Acceptable for a single-user tool; re-run draft to refresh links.