UNPKG

kestrel.markets

Version:

A typed, token-efficient language + runtime for agentic trading: agents author bounded plans, the runtime fires them at the tick. CLI + typed library + MCP server.

529 lines (486 loc) 29.4 kB
/** * # grade/receipts — the pure PROJECTOR that lifts agent receipts off a Session bus (kestrel-a57.5) * * `projectReceipts(bus)` is the deterministic fold from one canonical Session {@link ../bus Bus} into the * {@link SessionReceipts} curation record — the immutable set of every agent PROPOSAL, DECLINE, and * STAND-DOWN taken in the session. It mirrors the {@link ../blotter Blotter} projector's discipline * (ADR-0011): **a pure function of the bus** — same bus in ⇒ byte-identical receipts out, no engine-state * scrape, no wall clock, no RNG. * * **Curation honesty (the a57.5 doctrine).** The projector records EVERY decision, not only the armed ones. * A survivor-only record cannot grade curation, so declines and stand-downs are MANDATORY rows — the record * carries the plans the agent declined and stood down, alongside the plans it armed. * * **How a decision rides the bus (today, and the seam).** An agent decision is carried on the existing * **CONTROL** stream — the L2→L1 author-intent channel, whose `type` vocabulary is open (bus/types) — with * `type ∈ {propose, decline, stand_down}`, `target` naming the plan, and the structured * {@link DecisionDescriptor} (thesis axes, counterfactual, measurement versions, plan source) carried as * canonical JSON in the CONTROL `note`. This keeps the bus schema UNCHANGED. The note-as-JSON transport is a * documented SEAM: when a typed proposal envelope lands (a57.14) the descriptor moves onto a typed payload, * but the projector's contract — receipts are a pure projection of the bus — is unchanged. The **percept** * the agent acted on is projected today as the canonical bus PREFIX (`seq < proposed_seq`); the percept-v5 * {@link ../frame Frame} is the future enrichment (see {@link ./receipt.ts} `PerceptProjection`). * * **Fail closed (AGENTS.md).** A CONTROL decision whose `note` is missing/malformed is not silently dropped: * it is recorded in {@link SessionReceipts.skipped} with a logged reason (a corruption cannot masquerade as a * well-formed decision, and it cannot fabricate a thesis). A well-formed decision whose PLAN text fails to * parse degrades to a STAND-DOWN receipt (plan = null) whose rationale records the escape — the curation row * is kept ("parse escape → STAND_DOWN"). */ import type { BusEvent, ControlEvent, MetaEvent, Mode } from "../bus/index.ts"; import { serializeBus } from "../bus/index.ts"; import { canonicalizeJson } from "../canonical/json.ts"; import { ARMING_DECISIONS, COUNTERFACTUAL_KINDS, DECISION_KINDS, PlanParseError, ReceiptValidationError, createReceipt, sha256, type CounterfactualKind, type DecisionKind, type MeasurementVersions, type Percept, type PerceptProjection, type ProposalReceipt, type ThesisAxes, } from "./receipt.ts"; // ───────────────────────────────────────────────────────────────────────────── // The bus transport (CONTROL note-JSON) — a documented seam // ───────────────────────────────────────────────────────────────────────────── /** The CONTROL `type`s the projector reads as agent decisions. Other CONTROL actions (arm/disarm/author/ * supersede/allocate) are ignored — they are not decisions with a thesis. */ export const DECISION_CONTROL_TYPES: ReadonlySet<string> = new Set<string>(["propose", "decline", "stand_down"]); /** The projector that produces the percept-hash bytes today (the bus prefix). Pins meaning across * revisions; becomes `{ kind: "frame", ... }` when the percept-v5 Frame lands (a documented seam). */ const BUS_PREFIX_PROJECTION: PerceptProjection = { kind: "bus-prefix", renderer_version: "bus-prefix-v0" }; /** * The structured agent decision carried in a CONTROL event's `note` (canonical JSON). This is the SEAM * payload the projector decodes; when a typed proposal envelope lands (a57.14) these fields move onto a * typed bus payload without changing the projector's contract. */ export interface DecisionDescriptor { /** The fine-grained decision (accept/reparameterize/novel_plan/decline/stand_down). Must agree with the * CONTROL `type`: an arming kind under `propose`, `decline` under `decline`, `stand_down` under `stand_down`. */ readonly decision_kind: DecisionKind; /** The plan SOURCE text under decision (canonicalized to the plan-hash by the builder), or `null` for a * candidate-less stand-down. Curation honesty: a declined plan carries its text here, never dropped. */ readonly plan_text: string | null; readonly counterfactual: CounterfactualKind; readonly measurement_versions: MeasurementVersions; readonly thesis: ThesisAxes; /** The `ts` of the last information the agent used (≤ the decision's `ts`). Defaults to the last * perceived event's `ts` when absent. */ readonly information_cutoff?: number; readonly rationale?: string | null; } /** Encode a decision descriptor as the canonical-JSON string a CONTROL `note` carries. The inverse of the * projector's decode; exported so a harness (and tests) write the exact bytes the projector reads. */ export function encodeDecision(d: DecisionDescriptor): string { return JSON.stringify(d); } // ───────────────────────────────────────────────────────────────────────────── // The PRODUCER (kestrel-a57.18) — the driver-side half that WRITES what the projector reads // ───────────────────────────────────────────────────────────────────────────── /** * The thesis a DRIVER-authored decision carries today: **UNSTATED**. * * The receipt schema demands a thesis on every row, but the Session action surface * ({@link import("../session/agent.ts").Action}) carries no structured thesis — an agent states its * intent as free-text `journal`/`note` prose, never as the gradable axes. The producer therefore states * NOTHING on the agent's behalf: every axis is `null` (no claim), `conviction` is `0` (the floor — the * receipt asserts no confidence the agent did not express), and the day-type distribution is the single * explicit `unstated` mass. This is the honest encoding of "no thesis was stated", never a plausible * thesis invented by the runtime (a fabricated claim would be exactly the "silently false" the fail-closed * doctrine forbids, and it would poison the very clustering the axes exist to enable). * * **Seam (a57.14).** When the typed proposal envelope lands, the agent states its OWN axes and they ride * here instead. A grader can tell the two apart mechanically: `day_type_probabilities.unstated === 1`. */ export const UNSTATED_THESIS: ThesisAxes = { when: { session_phase: null, regime: null, event: null, trend_state: null, volatility_state: null }, where: { product: null, direction: null, strike: null, moneyness: null, book_state: null }, how: { entry_method: null, size_contracts: null, reload: null, invalidation: null, exit: null }, conviction: 0, day_type_probabilities: { unstated: 1 }, }; /** The commission model this engine applies to a simulated fill: **none**. A truthful, checkable pin — * the sim charges no commission, so a grader must not compare its EV against a commissioned run. */ export const DRIVER_COMMISSION_MODEL = "none-v0"; /** The bounded-risk envelope version the engine enforces (`budget NR` + the never-naked/intrinsic-floor * admission Gate). Bump when the envelope's arithmetic changes — EVs across versions refuse comparison. */ export const DRIVER_RISK_ENVELOPE = "budget-r-v1"; /** The grader version a receipt minted by this engine is measured by (the headline grader). */ export const DRIVER_GRADER = "grade-v1"; /** * Pin the measurement versions for a decision minted by the Session driver. Pure: a function of the run's * judge (`fill_model`) and the bus's fidelity only — no clock, no RNG. Everything a version could move a * number through is pinned, so a grade can refuse to compare EVs across versions (ADR-0006). */ export function driverMeasurementVersions(judge: { readonly fillModel: string; readonly fidelity: string; }): MeasurementVersions { return { fill_model: judge.fillModel, commission_model: DRIVER_COMMISSION_MODEL, bus_fidelity: judge.fidelity, risk_envelope: DRIVER_RISK_ENVELOPE, grader: DRIVER_GRADER, }; } /** One agent decision the driver is about to write onto the bus (the producer's input). `control` is the * coarse CONTROL marker; `decision_kind` is the fine kind — the projector re-checks the pairing * fail-closed, so a mismatch here is SKIPPED with a reason, never silently banked. */ export interface DecisionEmit { readonly control: "propose" | "decline" | "stand_down"; /** The plan NAME the decision is about — the key the curation verdict reconciles against the bus's * arm record. Omitted only for a candidate-less stand-down. */ readonly target?: string; readonly decision_kind: DecisionKind; /** The plan SOURCE under decision, or `null` for a candidate-less stand-down. Never dropped: a declined * or stood-down plan carries its text so the row is curation-honest. */ readonly plan_text: string | null; readonly rationale?: string | null; readonly measurement_versions: MeasurementVersions; /** The agent's stated thesis. Defaults to {@link UNSTATED_THESIS} — the producer invents no claim. */ readonly thesis?: ThesisAxes; /** The baseline the decision is measured against. Defaults to `null` (the flat/no-position baseline). */ readonly counterfactual?: CounterfactualKind; } /** * Build the CONTROL event FIELDS that carry one agent decision receipt onto the Session bus — the * producer half of the a57.5 seam (kestrel-a57.18). The driver spreads the result into its own `emit` * (which stamps `ts`/`seq`), so the transport stays the unchanged CONTROL stream and the projector's * contract — receipts are a pure projection of the bus — is untouched. * * Pure and deterministic: same input ⇒ byte-identical `note`. It performs NO validation of its own — the * projector is the single validator (one definition, never a second one), and it fails closed on anything * this writes that does not cohere. */ export function decisionControlFields(d: DecisionEmit): { readonly type: string; readonly target?: string; readonly note: string; } { const descriptor: DecisionDescriptor = { decision_kind: d.decision_kind, plan_text: d.plan_text, counterfactual: d.counterfactual ?? "null", measurement_versions: d.measurement_versions, thesis: d.thesis ?? UNSTATED_THESIS, ...(d.rationale !== undefined ? { rationale: d.rationale } : {}), }; return { type: d.control, ...(d.target !== undefined ? { target: d.target } : {}), note: encodeDecision(descriptor), }; } // ───────────────────────────────────────────────────────────────────────────── // The curation record // ───────────────────────────────────────────────────────────────────────────── /** One CONTROL decision the projector could NOT lift into a receipt — recorded, never silently dropped * (fail-closed). `seq` anchors it to the bus; `reason` is the logged cause. */ export interface SkippedDecision { readonly seq: number; readonly reason: string; } /** The session header a receipt record projects from — the opaque calendar token and the mode, read purely * from the opening META. */ export interface ReceiptSession { readonly date: string; readonly mode: Mode; } /** * The **curation-honesty verdict** — the structural answer to "can this record grade CURATION?" (the a57.5 * doctrine, made mechanical rather than an honor system). A survivor-only record — arming decisions with no * declines/stand-downs — CANNOT distinguish a good selector from a lucky one, and a receipt set that does not * reconcile with the actual arm record on the bus is a self-reported narrative, not the curation-honest * counterpart to the {@link ../blotter Blotter}. When `gradable` is `false` a grader MUST REFUSE to score * curation; `reasons` names every fail-closed cause (empty iff gradable). Pure/deterministic — computed from * the bus and the projected receipts, never a wall clock or engine scrape. */ export interface CurationVerdict { /** `true` iff the record is safe to grade for curation: it is NOT survivor-only AND its arming receipts * reconcile with the bus's arm record. A grader refuses curation grading when this is `false`. */ readonly gradable: boolean; /** Every fail-closed reason the record is not curation-gradable, in a stable order (empty iff gradable). */ readonly reasons: readonly string[]; /** Count of arming receipts (accept/reparameterize/novel_plan) — the survivors. */ readonly armed: number; /** Count of the declines and stand-downs a survivor-only ledger would drop (MANDATORY rows). */ readonly declined_or_stood_down: number; /** `true` iff the bus carries an arm/execution record (any `PLAN` armed transition or `CONTROL arm`); * when `false` the receipt channel is projected in isolation and reconciliation is not applicable. */ readonly arm_record_present: boolean; /** Plan names armed on the bus (PLAN armed / CONTROL arm) with NO corresponding arming receipt — a * curation-hiding divergence (the engine armed something the receipt record does not own). Sorted. */ readonly unreceipted_arms: readonly string[]; /** Plan names an arming receipt claims it armed that the bus's arm record never armed — a self-reported * arm the execution log does not back. Sorted; only computed when {@link arm_record_present}. */ readonly phantom_arms: readonly string[]; } /** * The immutable curation record of one Session: every agent decision receipt, in bus order, plus the * decisions that failed to lift (recorded, never dropped). A pure projection of the bus — byte-stable under * {@link serializeReceipts}. */ export interface SessionReceipts { readonly session: ReceiptSession; /** Every agent decision receipt, ordered by `proposed_seq` (bus order). Carries armed, declined, AND * stood-down decisions — curation honesty. */ readonly receipts: readonly ProposalReceipt[]; /** Malformed CONTROL decisions, recorded with a logged reason (fail-closed; never silent). */ readonly skipped: readonly SkippedDecision[]; /** The structural verdict on whether this record can grade CURATION — survivor-only detection plus * reconciliation against the bus's arm record. A grader REFUSES curation grading when not `gradable`. */ readonly curation: CurationVerdict; } // ───────────────────────────────────────────────────────────────────────────── // Decode helpers (fail-closed) // ───────────────────────────────────────────────────────────────────────────── /** Parse a CONTROL `note` into a {@link DecisionDescriptor}, or throw with a precise reason. Total. */ function decodeDescriptor(note: string | undefined): DecisionDescriptor { if (note === undefined || note.length === 0) throw new ReceiptValidationError("CONTROL decision carries no note payload"); let raw: unknown; try { raw = JSON.parse(note); } catch (err) { throw new ReceiptValidationError(`CONTROL decision note is not valid JSON: ${(err as Error).message}`); } if (typeof raw !== "object" || raw === null || Array.isArray(raw)) { throw new ReceiptValidationError("CONTROL decision note must be a JSON object"); } const d = raw as Record<string, unknown>; if (typeof d.decision_kind !== "string" || !DECISION_KINDS.has(d.decision_kind as DecisionKind)) { throw new ReceiptValidationError("decision descriptor decision_kind is missing or unknown"); } if (typeof d.counterfactual !== "string" || !COUNTERFACTUAL_KINDS.has(d.counterfactual as CounterfactualKind)) { throw new ReceiptValidationError("decision descriptor counterfactual is missing or unknown"); } if (d.plan_text !== null && typeof d.plan_text !== "string") { throw new ReceiptValidationError("decision descriptor plan_text must be a string or null"); } if (typeof d.measurement_versions !== "object" || d.measurement_versions === null) { throw new ReceiptValidationError("decision descriptor measurement_versions is missing"); } if (typeof d.thesis !== "object" || d.thesis === null) { throw new ReceiptValidationError("decision descriptor thesis is missing"); } return d as unknown as DecisionDescriptor; } /** The decision kinds the given CONTROL `type` admits (fail-closed pairing between coarse marker and fine * kind). Prevents a `propose` note that secretly declares a `decline`, etc. */ function admissibleKinds(controlType: string): ReadonlySet<DecisionKind> { switch (controlType) { case "propose": return ARMING_DECISIONS; case "decline": return new Set<DecisionKind>(["decline"]); case "stand_down": return new Set<DecisionKind>(["stand_down"]); default: return new Set<DecisionKind>(); } } // ───────────────────────────────────────────────────────────────────────────── // projectReceipts // ───────────────────────────────────────────────────────────────────────────── /** The opening META of the bus, or a loud refusal (fail-closed): a well-formed bus opens with exactly one * META. Unlike the Blotter projector this does NOT require a GRADED bus — a receipt is authored on the input * tape, before any fill model grades it. */ function requireMeta(bus: readonly BusEvent[]): MetaEvent { const meta = bus.find((e): e is MetaEvent => e.stream === "META"); if (meta === undefined) { throw new ReceiptValidationError("no META header in bus (a well-formed bus opens with exactly one META)"); } return meta; } /** * Project one Session bus into its immutable {@link SessionReceipts} curation record. Pure and deterministic * (ADR-0011): no wall clock, no RNG, no engine-state scrape — the SAME bus yields byte-identical receipts * (proven by re-projecting and byte-comparing {@link serializeReceipts}). Records every armed/declined/ * stood-down decision (curation honesty); records malformed decisions in `skipped` (fail-closed). */ export function projectReceipts(bus: readonly BusEvent[]): SessionReceipts { const meta = requireMeta(bus); const session: ReceiptSession = { date: meta.session_date, mode: meta.mode }; const receipts: ProposalReceipt[] = []; const skipped: SkippedDecision[] = []; /** Plan names an arming RECEIPT claims (the `target` of a decision that projected to an arming kind). */ const armedByReceipt = new Set<string>(); for (const e of bus) { if (e.stream !== "CONTROL") continue; const ce = e as ControlEvent; if (!DECISION_CONTROL_TYPES.has(ce.type)) continue; // not a decision (arm/disarm/author/…) try { const d = decodeDescriptor(ce.note); if (!admissibleKinds(ce.type).has(d.decision_kind)) { throw new ReceiptValidationError( `CONTROL type ${JSON.stringify(ce.type)} is incompatible with decision_kind ${JSON.stringify(d.decision_kind)}`, ); } // Percept — the canonical bus PREFIX the agent could have perceived (seq strictly before this // decision). This is the AVAILABLE representation today; the percept-v5 Frame is the seam. // The bus is append-only in monotonic `seq` order, so the prefix's last element is its watermark. const prefix = bus.filter((p) => p.seq < ce.seq); if (prefix.length === 0) { throw new ReceiptValidationError("decision precedes the session META (no percept prefix)"); } const lastPerceived = prefix[prefix.length - 1]!; const prefixBytes = serializeBus(prefix); const percept: Percept = { sha256: sha256(prefixBytes), byte_count: Buffer.byteLength(prefixBytes, "utf8"), seq_watermark: lastPerceived.seq, projection: BUS_PREFIX_PROJECTION, }; // The claimed information cutoff can never exceed the watermark of the percept the receipt pins: // the percept is the bus prefix (seq < proposed_seq), whose latest information is lastPerceived.ts. // A cutoff beyond it is a look-ahead provenance claim the pinned percept cannot back — fail closed. if (d.information_cutoff !== undefined && d.information_cutoff > lastPerceived.ts) { throw new ReceiptValidationError( `information_cutoff ${d.information_cutoff} exceeds the pinned percept's watermark ts ${lastPerceived.ts} ` + `(look-ahead: the decision claims information more recent than the percept it hashed)`, ); } const information_cutoff = d.information_cutoff ?? lastPerceived.ts; // "parse escape → STAND_DOWN": a well-formed decision whose plan text does not parse degrades to a // stand-down receipt (plan = null) — the curation row is KEPT, with the escape recorded in rationale. let receipt: ProposalReceipt; try { receipt = createReceipt({ session_date: meta.session_date, proposed_at: ce.ts, proposed_seq: ce.seq, information_cutoff, decision_kind: d.decision_kind, percept, plan_text: d.plan_text, counterfactual: d.counterfactual, measurement_versions: d.measurement_versions, thesis: d.thesis, rationale: d.rationale ?? null, }); } catch (err) { if (err instanceof PlanParseError && d.plan_text !== null) { receipt = createReceipt({ session_date: meta.session_date, proposed_at: ce.ts, proposed_seq: ce.seq, information_cutoff, decision_kind: "stand_down", percept, plan_text: null, counterfactual: d.counterfactual, measurement_versions: d.measurement_versions, thesis: d.thesis, rationale: `parse escape → stand-down: ${err.message}`, }); } else { throw err; } } receipts.push(receipt); // Bind this arming receipt to the plan NAME it claims to have armed, for reconciliation against the // bus's real arm record. A decision that degraded to stand-down (parse escape) is NOT an arm. if (ARMING_DECISIONS.has(receipt.decision_kind) && ce.target !== undefined) { armedByReceipt.add(ce.target); } } catch (err) { const reason = err instanceof Error ? err.message : String(err); skipped.push({ seq: ce.seq, reason }); } } receipts.sort((a, b) => a.proposed_seq - b.proposed_seq); skipped.sort((a, b) => a.seq - b.seq); const curation = assessCuration(bus, receipts, armedByReceipt); return { session, receipts, skipped, curation }; } // ───────────────────────────────────────────────────────────────────────────── // Curation-honesty verdict — survivor-only detection + arm reconciliation // ───────────────────────────────────────────────────────────────────────────── /** The plan names the bus's real ARM record armed: every `PLAN` armed transition and every `CONTROL arm` * with a `target`. This is the execution-side arm set the receipt channel must reconcile with — the same * canonical arm path that drives the {@link ../blotter Blotter}. Sorted, deduplicated. */ function armedOnBus(bus: readonly BusEvent[]): string[] { const armed = new Set<string>(); for (const e of bus) { if (e.stream === "PLAN" && e.type === "lifecycle" && e.state === "armed") armed.add(e.plan); else if (e.stream === "CONTROL" && (e as ControlEvent).type === "arm") { const t = (e as ControlEvent).target; if (t !== undefined) armed.add(t); } } return [...armed].sort(); } /** * Compute the {@link CurationVerdict} for a projected record (findings: survivor-only records and the * receipt/arm decoupling). Fail-closed and structural, NOT an honor system: * • **survivor-only** — arming receipts with zero declines/stand-downs cannot grade curation; an empty * record has nothing to grade. Either makes the record non-gradable. * • **arm reconciliation** — when the bus carries a real arm record, the arming receipts and the bus's * arm set must AGREE: an arm with no receipt hides curation; a receipt claiming an arm the bus never * armed is a self-reported fiction. Either divergence makes the record non-gradable. * Pure and deterministic — a function of the bus and the receipts only. */ function assessCuration( bus: readonly BusEvent[], receipts: readonly ProposalReceipt[], armedByReceipt: ReadonlySet<string>, ): CurationVerdict { const armed = receipts.filter((r) => ARMING_DECISIONS.has(r.decision_kind)).length; const declined_or_stood_down = receipts.length - armed; const busArms = armedOnBus(bus); const arm_record_present = busArms.length > 0; const busArmSet = new Set(busArms); const unreceipted_arms = busArms.filter((name) => !armedByReceipt.has(name)); const phantom_arms = arm_record_present ? [...armedByReceipt].filter((name) => !busArmSet.has(name)).sort() : []; const reasons: string[] = []; if (receipts.length === 0) { reasons.push("no decisions recorded — nothing to grade (curation requires declines/stand-downs)"); } else if (declined_or_stood_down === 0) { reasons.push( `survivor-only: ${armed} arming decision(s) with no declines or stand-downs recorded — ` + `a survivor-only record cannot distinguish a good selector from a lucky one (declines/stand-downs are MANDATORY)`, ); } if (unreceipted_arms.length > 0) { reasons.push( `arm record diverges: ${unreceipted_arms.length} plan(s) armed on the bus have no arming receipt ` + `[${unreceipted_arms.join(", ")}] — the engine armed what the receipt record does not own`, ); } if (phantom_arms.length > 0) { reasons.push( `arm record diverges: ${phantom_arms.length} arming receipt(s) name a plan the bus never armed ` + `[${phantom_arms.join(", ")}] — a self-reported arm the execution log does not back`, ); } return { gradable: reasons.length === 0, reasons, armed, declined_or_stood_down, arm_record_present, unreceipted_arms, phantom_arms, }; } // ───────────────────────────────────────────────────────────────────────────── // Canonical serialization // ───────────────────────────────────────────────────────────────────────────── /** Serialize a {@link SessionReceipts} record to its canonical bytes: deterministic machine-JSON * (lexicographically ordered keys, compact, one trailing newline). Pure and total — the same record always * produces byte-identical text, the substrate a determinism check byte-compares. The key order is the one * {@link canonicalizeJson canonical byte order} (canonical/json); the single trailing newline is canon. */ export function serializeReceipts(record: SessionReceipts): string { return canonicalizeJson(record) + "\n"; } /** The declined and stood-down receipts (the curation the agent left on the table) — the non-arming * decisions a survivor-only record would drop. A convenience over {@link SessionReceipts.receipts}. */ export function declinedAndStoodDown(record: SessionReceipts): readonly ProposalReceipt[] { return record.receipts.filter((r) => !ARMING_DECISIONS.has(r.decision_kind)); }