Assay’s judgment layer

Status: bun LIFEOS/ASSAY/Assay.ts status · optional

What it does

Wherever LifeOS code has to make a fuzzy call (“is this message urgent?”, “did this answer cover the ask?”), it asks Assay’s judgment layer one typed question and gets back a probability and a yes or no on whether it may act. The model that produces the number is an engine called Clef-omni, a Cloudflare Workers AI model that reads text, images, audio and video, running underneath.

The judgment layer is careful by design. Every caller starts in shadow mode: its answers are logged next to what actually happened, but nothing acts on them. A caller moves to enforce only after its answers have been measured against human labels. So an empty registry is the right state for a new install, and nothing breaks without it: each call site is built so that its session, eval or job can still finish when the judgment layer is off.

Fill in

LIFEOS/USER/CONFIG/assay-callers.json is the caller registry. The shipped file holds only a defaults row:

{ "defaults": { "min_labels": 10, "provisional_threshold": 0.85 } }

min_labels is how many human labels a question needs before it can earn authority. provisional_threshold is the starting cut for a newly registered question. Every other top-level name is a caller, with its mode (shadow or enforce), thresholds per question, daily budget and the source of its ground-truth labels. You rarely write these by hand: the Assay skill’s CreateQuestions intent designs a question set for a decision and emits a shadow registry row for you to add, and bunker miss registers new questions from judged misses.

The Bunker CLI is maintainer-private and not in the public release; run code rows with bun test ./<path> and suite rows with bun skills/Assay/Tools/EvalRunner.ts.

Accounts and secrets

The judgment layer needs a Cloudflare account with Workers AI. Put the account id in ~/.claude/.env as CLOUDFLARE_ACCOUNT_ID and your Cloudflare API token as CLOUDFLARE_API_TOKEN, and seal it. That is the same one token every other Cloudflare tool in LifeOS uses; it needs Workers AI access, and the judgment layer reads no second Cloudflare token:

bun LIFEOS/TOOLS/Secrets.ts seal
bun LIFEOS/TOOLS/Secrets.ts has CLOUDFLARE_ACCOUNT_ID CLOUDFLARE_API_TOKEN   # prints "KEY yes" or "no" for each

The question text, state and any images, audio or video you send go to Cloudflare’s Workers AI API on your own account. The client refuses any text that contains a secret-shaped value, but it cannot scan media bytes, so attach only media you mean to send. Without the credentials, a judgment call fails with CLOUDFLARE_ACCOUNT_ID is not set; call sites that are set to fail open carry on without the judgment.

Extraction (extract(), through Clairic) runs on the same account and token, so it needs no further setup.

Prove it works

bun LIFEOS/ASSAY/Assay.ts status        # one line per registered caller: mode, calls in 24h, error rate, agreement
bun LIFEOS/ASSAY/Assay.ts callers       # callers in the registry, and any seen in the log but not registered
bun LIFEOS/ASSAY/Engine.ts probe        # one live call to the engine; proves the credentials and the network path

On a fresh install status prints nothing and exits 0, because the registry has no callers yet. That is the expected state, not a fault. probe is the real test of the credentials: it prints a model name, a request id and an answer for a harmless sentence, and exits 1 with CLOUDFLARE_ACCOUNT_ID is not set when the account id is missing.

To turn the judgment layer off for a session without removing anything, set LIFEOS_GLANCE_DISABLED=1.