git:20260830.cbb6d69 to git:20260830.423449c

7 added, 6 removed. Audit A to A.

---

name: model-release-intake

description: Ceremony for taking in a newly released (or retired) model — detect catalog drift, characterize the new mind, update the single activity table, and attest. Trigger is a human noticing a release, or `mix tightbeam.catalog.diff` reporting drift.

---

# Model release intake

A new model shipped, or an old one vanished, and the catalog no longer matches the judgment

the org has written down. This is a rare, deliberate ceremony — not automation. You run it

when a human notices a release or when the drift detector flags it. The goal is that every

model an agent can be handed is a model someone has characterized, and every characterization

still names a model that exists.

## When to run

- `mix tightbeam.catalog.diff` exits nonzero (uncharacterized live refs, or characterized

refs that have vanished from the live inventory), OR

- a human learns a model was released, renamed, retired, or repriced.

## Steps

1. **Detect.** Run the diff and read it as the worklist:

```

mix tightbeam.catalog.diff # human report; nonzero exit on drift

mix tightbeam.catalog.diff --json # machine-readable, same exit contract

```

- UNCHARACTERIZED live refs → new minds to characterize (steps 2-4).

- VANISHED characterized refs → retired minds to remove or re-point (step 5).

- 2. **Characterize each new model.** Add an entry to `preferred-models.md` in the org's

- guidance, in the established format: what the model is FOR; what it is NOT for; its cost

- posture; when to PREFER or AVOID it relative to its neighbors. Base the judgment on the

- model's real strengths, not marketing — where you are unsure, say so and pick a

- conservative quality-floor placement. A characterization is a judgment on the record, so

- the next agent adjudicates against it instead of guessing.

+ 2. **Characterize each new model.** Add an entry to the working-set section of

+ `kungfu/agentic-engineering/preferred-models.md`, in the established format: what the

+ model is FOR; what it is NOT for; its cost posture; when to PREFER or AVOID it relative

+ to its neighbors. Base the judgment on the model's real strengths, not marketing — where

+ you are unsure, say so and pick a conservative quality-floor placement. A

+ characterization is a judgment on the record, so the next agent adjudicates against it

+ instead of guessing.

3. **Place it in the activity table.** Decide where the new mind sits relative to the

existing ones for each affected job: which ordered rows it clears, and which it does

not. The end of a row is the quality floor. Record every selection only in

`kungfu/agentic-engineering/preferred-models.md`.

4. **Sweep every activity row.** Reconsider every row in

`kungfu/agentic-engineering/preferred-models.md`, including the general-agent,

onboarding, and guidance rows. Do not create an archetype preference or guidance

selector: the activity table is the only model-selection home.

5. **Handle vanished refs.** For each characterized-but-gone model: remove its

characterization and remove or re-point every activity-table occurrence if the release

was a rename. Never leave an activity row pointing only at a ref the live catalog no

longer serves — that is a guaranteed refused spawn.

6. **Re-run the detector to zero.** `mix tightbeam.catalog.diff` must exit zero: the

characterized set now exactly covers the live set, no vanished refs remain.

7. **Attest.** Record the intake — what model was taken in (or retired), the characterization

judgment, the activity rows swept, and the diff-clean result — as an attest on the work,

so the change to the org's model judgment is on the record and not a silent edit.

## Notes

- This is DEV-time intake: it runs against the live catalog with no gateway required (the

detector starts the fetchers standalone). Long-lived orgs additionally get the same diff

promoted into the catalog refresh heartbeat (a `model-added` condition fact + steward

wake); that runtime path is separate from this human/CI ceremony.

- The merge gate runs the externally-tagged coverage test (`mix test --only external`) so an

uncharacterized model is caught loudly at the choke point, not in production.