6. Where next

About one minute

If the guide clicked, you already have the mental model: one commons, two phases, abstain when unsure, compile on the client, hand off when the corpus is done helping.

Deeper material stays in the private repository under docs/internal/ (not on this public site):

Path	When you need it
docs/internal/TOOL_CONTRACT.md	Frozen outcomes and field musts
docs/internal/DAML_REASON.md	One-page capability brief
docs/internal/DEVELOPING.md	Auth, Postgres, sidecar, corpus bind
docs/internal/DEMO.md / DEMO_CHAT.md	Rehearse gates and the draft loop
docs/internal/COMPARE.md	Recall regression vs the Python sidecar
docs/internal/PORT_PLAN.md	Sprint and cutover status
docs/internal/VALUE_NARRATIVE.md	External and press framing

How we evaluate (v1)

Public CI already carries a labeled regression set on the promoted corpus_v2 bundle with the cross-encoder path:

• Phase A tables (//go:build corpus) — vague abstains, sharp settlement suggests, honest IOU/DvP abstain-or-propose.
• Phase B fail-and-help — aligned drafts must not fail; mismatched domains fail; compile-failed attaches patterns when recall is strong.

That is the v1 evaluation set: outcome classes under honest retrieval, not a published confusion matrix. Deeper alignment research (labeled tasks, matrices) continues on the private eval-lab branch when needed — it is not a separate public issue for day-to-day product work.

Still small?

Yes. One tool, two phases, a handful of outcomes, and a clear agent loop. If this guide ever needs more than about fifteen minutes, split “operators” from “agent authors” instead of bloating the ten-minute path.

Thesis (again)

We do not generate Daml. We gate agents on a promoted corpus, and we abstain when unsure. That is how vibecoding sessions stop diverging and start inheriting the same corrected practice.

---

← Try it · Back to start