monark

A coverage-controlled gate for agents. You bring your own predictor and nonconformity scores; MONARK calibrates a split-conformal region and returns commit / defer / abstain over YOUR prediction over one public MCP endpoint. Never a probability of being right, never permission to execute a tool. Four tools - attest, gate, cascade, calibrate.

MONARK - a coverage-controlled gate for agents

MONARK exposes four pure tools over one public MCP endpoint: {attest, gate, cascade, calibrate}. It gates a prediction, never an act: gate never executes the named tool. The decision is one of commit / defer / abstain; B_t is caller-carried; never a probability of being right. attest is demonstrative, not probative - it replays a committed witness, it does not prove an outcome. This skill is a reference implementation of a coverage-controlled gate; adopt the contract.

The BYO loop: calibrate then gate

The real path is bring your own: you own the predictor and the nonconformity score function. calibrate turns YOUR nonconformity scores into a split-conformal region; gate then returns a coverage verdict on YOUR next prediction under that region. MONARK stores nothing between calls, and the audit calibrate to gate closes when the verdict's calib_digest equals your score-set digest.

calibrate - the honesty label (verbatim)

The calibrate tool carries exactly this label, on its description, its result content, and its output field alike:

split-conformal quantile at miscoverage α over caller-supplied nonconformity scores. MONARK does not see, store, or verify the caller's data or model, and does not validate that the supplied numbers are nonconformity scores of any model. Marginal 1−α coverage holds ONLY for future points exchangeable with the supplied scores; non-exchangeable data (e.g. distribution-shifted or time-ordered) voids it. Never a probability of being right.

What a verdict is, and what it is NOT

• commit / defer / abstain are the three actions; B_t is caller-carried; never a probability of being right.
• gate never executes the named tool - it gates YOUR prediction, not YOUR act.
• The B_t mechanism - read this before wiring any budget. B_t is a number YOU pass in each call; MONARK never measures, derives, or stores it (stateless); it is compared only to YOUR bFloor (below it ⇒ budget_exhausted). It is NOT a $/token spend cap and NOT a rate-limit. allow is a coverage verdict on YOUR prediction, NOT permission to execute the named tool (MONARK never executes it). MONARK does not evaluate the legitimacy of an act and does not predict prices — it gates YOUR predictions. For a $/token spend cap, use your platform's spend controls; MONARK is not that.

MONARK offers no guarantee of availability. It does not predict prices and does not judge whether an act is legitimate; it gates the coverage of YOUR prediction and nothing else.

Operational surface

A public, unauthenticated endpoint, no availability commitment; bounded: n ≤ 10000 scores, request body ≤ 256 KB. A verdict below your bFloor returns budget_exhausted; out-of-calibration input returns under_calib and abstains - no success is invented.

The two built-in task classes are NOT use cases

The two built-in task_class values (btc-dir-15m, cascade-liquidable-24h) are internal plumbing fixtures, not use cases and not endorsements: btc-dir-15m uses a declared synthetic calibration (never a measured predictor), and cascade-liquidable-24h ships no calibration at all, so it abstains (under_calib). They are NOT use cases. The real path is BYO: bring your own predictor + nonconformity scores.

A third task_class, stable-run-velocity-24h (redemption-run velocity, the Narabi sensor), is served by this endpoint with a committed calibration for one population — USDe, key narabi:persistence-v2@eip155:1/erc20:0x4c9edd5852cd905f086c759e8383e09bff1e68b3, measured on calm onchain redemption-flow windows. The calibration is measured non-stationary across half-years, so no per-window coverage is claimed; the committed region is static and every other population abstains (under_calib). Alongside it, an off-tool daily sentinel steps an adaptive quantile tracker on the attested 24h flow and publishes a replayable timeline (state.json, timeline.jsonl) at monarkgate.tech/narabi/; the committed gate region does not change until a pre-registered drift criterion fires and an ADR says so.

Endpoint and license

• Endpoint: https://mcp.monarkgate.tech/mcp (HTTP/JSON mirror: POST https://api.monarkgate.tech/{attest|gate|cascade|calibrate}).
• Install on each runtime: see INTEGRATION.md.
• This skill is licensed under MIT-0 (MIT No Attribution); see LICENSE. The MONARK harness is a separate, Apache-2.0 codebase - only this descriptive skill is MIT-0.