Promote data-driven AI-tells + first-class severity into prose.mjs

[bdelanghe]

Promotes the spike from #8 into production, closing proposals 1 & 3 of #6. Behavior-preserving — the deterministic audit output is byte-identical; the change is structural.

What changed

• ai-tells.json (root, canonical) — AI-tell patterns + lexicon as data, each rule carrying its own level. Promoted from spikes/; seeded from the vale-signs-of-ai-writing corpus. aiIsms() loads it; the structural tells that aren't simple regex (em-dash count, anaphora, tricolons) stay in code.
• prose.mjs — every check now returns { level, msg } (error | warn | suggestion) instead of bare strings, so severity is first-class rather than re-derived downstream.
• audit.mjs — renders the glyph from level (and folds the still-string, cached type-audit findings into the same model via a small classifier) instead of regex-matching message text for tiering.
• spikes/ — vale-rules.mjs now reads the promoted root ai-tells.json (single source of truth — the spike can't drift from production); spikes/ai-tells.json removed; spike README notes the promotion. The optional vale-provider.mjs shell-out spike stays (proposal 2, still open).
• test.mjs — assert on finding.msg; add coverage for the { level, msg } contract and an error-level data rule (chatbot artifact).
• README.md — document the data-driven rules file and first-class severity.

Why

Before: patterns/lexicon hard-coded in prose.mjs, and tiers were pattern-matched from message text in audit.mjs (/^(ai-ism:|proofread:)/ etc.) — brittle and duplicated. Now the AI-tell rules are editable data that tracks the upstream corpus, and severity rides on each finding.

Verification

node test.mjs passes. Stubbed full node audit.mjs produces identical output to main, now driven by level:

  ✦ pdp.body.aiism       [body    ] 10/10
       ⚠ ai-ism: "isn't just a frame — it's" — antithesis "it isn't X — it's Y" cadence
       ⚠ ai-ism: "the easy part" — "the [easy/hard] part" framing
       ⚠ ai-ism: "Fast, simple, and beautiful" — rule-of-three triad; vary it
       · grammar/style: "just" — "just" can weaken meaning

New data rules (e.g. chatbot artifacts) now fire at error:

$ node spikes/vale-rules.mjs "As an AI language model, I cannot provide that."
       ✗ [error] chatbot meta-artifact — "As an AI language model"

(Full audit.mjs verified locally with stubs for the network-blocked JSR deps; CI has JSR access.)

Closes #6 proposals 1 & 3 (proposal 2 — the vale shell-out — remains an opt-in spike).

🤖 Generated with Claude Code

---

Generated by Claude Code