Changes made

1. Content limit raised: 1024 → 2048
src/lib/db.ts — MAX_CONTENT_LENGTH = 2048
2. Model token budget raised: 800 → 1600
src/lib/analyze.ts — so gpt-oss-120b can reason + produce full JSON for longer seeds
3. Conjugation system prompt rewritten
Key additions to src/lib/prompts.ts:
  • "atomic does NOT mean short" — stops penalizing rich seeds
  • "do not paraphrase the seed step-by-step" — attacks contract parroting
  • "do not append 'exists in context'" — kills the padding habit
  • "all generated text must be lowercase" — reinforced casing rule
  • "prefer discriminative tags" — reduces mushy generic tags
  • "prefer three independent axes of meaning" — improves slug compression
4. Four worst seeds rewritten with more nuance
Updated in D1 + eval script:
  • project.retro.run — from 45 words → 85 words, adds pattern naming + owner assignment + leverage ranking
  • project.status.report — from 40 words → 95 words, adds dependency-chain tracking + risk accumulation framing
  • data.report.narrative — from 30 words → 85 words, adds caveats handling + certainty calibration
  • project.email.update — from 25 words → 75 words, adds audience-adaptive tone + anti-fluff instruction

Smoke test: project.retro.run (new prompt + richer seed)

Before (old prompt + thin seed)
Slug: project.retrospective.summary
Contract: a retrospective summary containing three sections (what went well, what didn't go well, what to try differently) with each item assigned an action owner, and a top‑3 improvement list is present in context
Outcome: Produce a structured project retrospective with owners
Score: avg 4.4/10
After (new prompt + richer seed)
Slug: project.retrospective.leverage ✅ more specific
Contract: a structured retrospective report with sections titled 'keep doing', 'stop doing', and 'start doing', each containing items that list the underlying pattern name and an assigned owner, plus a separate list of the three highest-leverage improvements for the next similar project, exists in the context
Outcome: Create a actionable retrospective with owners and top improvements
Improvement: slug is sharper, contract names actual artifact structure, picks up "underlying pattern" from the richer prompt
Still not perfect: contract still ends with "exists in the context" (the model is stubborn), outcome has uppercase "Create" and grammar nit "a actionable". But the direction is clearly better. A full rerun of all 95 seeds will show the aggregate improvement.