Decide with code first, models last.

teleon.dev/docs/q20 runs each decision through rules, then local encoders, then small models, and records an abstention when the evidence is thin.

teleon.dev/docs/q20Queue cell Q20httpsjsonmcp

Where each decision was answered

1Ruleslicense text matched exactly62% answered

2CPU encoderlocal classifier27% answered

3Small modelhosted, low cost8% answered

4Reviewabstained with evidence3% marked

Most items stop at the first two rungs; the rest go up or wait for review.

What it does

Rules and local encoders answer most decisions before a model is called.

Low-confidence cases are marked for review with their evidence.

The rung that answered is logged for later tuning.

A queue cell owns one kind of task end to end: its rows, its leases and its hand-offs.

HoldsDecision items

OrderRules, encoder, small model, review

FeedsConfidence scores, voting

AccessRING-2 to claim

Identifiers

Execution contracts

Task rowCopy

{
  "queue": "q20",
  "item": "classify_license",
  "text": "Permission is hereby granted...",
  "rungs": ["rules", "cpu_encoder", "small_model"]
}

Method	Path	Access	What it does	Status
`GET`	`/v1/tasks`	RING-3	Browse decision items
`POST`	`/v1/claim`	RING-2	Claim a decision item

Listed paths appear in the service’s published files today. In the one-host plan every path here is served under api.teleon.dev/v1/q20.

Reads from

Feeds into

Give your agent the read-only guide first. It reports back before any write.

Agent prompt

Read https://teleon.dev/docs/q20.md and show the cascade rungs. Read only.

Deterministic check: A check written as code that gives the same answer every time for the same input, at code speed and cost.
Canary: A small share of traffic or checks run against a change first, so a problem shows up early and stays small.
PAVA: Pool adjacent violators: the standard way to turn raw scores into calibrated probabilities that only go up.