Lachesis

Lachesis is a typed functional compiler/runtime for measured agent programs. Models may propose an immutable JSON plan, but they cannot add source code, callbacks, hidden loops, or undeclared effects. The kernel parses, resolves, type-checks, bounds, content-addresses, and executes the plan through explicit effect interpreters.

This repository contains the measured plan kernel, deterministic generation benchmark substrate, the frozen M1b/M1c experiment controllers, the completed M2 functional-JSON-IR versus restricted-capability-TypeScript representation ablation, the completed M3 factorial evidence-substrate study, and the M5 public-alpha evidence runtime.

What works

• Strict Zod validation for an untrusted, versioned graph wire format.
• Duplicate-ID, dangling-reference, missing-root, cycle, and dead-node detection.
• A versioned nominal schema and operation catalog with safe erased runtime boundaries.
• All initial operators: input, constant, invoke, map, filter, fold, select, effect, checkpoint, and boundedFix.
• Complete schema inference, effect/capability inference, conservative resource maxima, and budget rejection.
• Root provenance and typed input-dependency, operation-dominance, state-change, required-operation, and required-effect obligations.
• Portable canonical JSON and SHA-256 hashing through Web Crypto.
• An opaque compile/execute boundary that binds successful analysis, policy, plan hash, and immutable catalog fingerprint.
• A canonical generator-facing language manifest with JSON Schemas, complete signatures, effects, bounds, reducer laws, and available policy.
• Pure, replay, and caller-supplied mock effect execution with runtime budget, output-schema, cardinality, and recursion-progress enforcement.
• Request-bound replay identities and verified recording output digests.
• A typed in-memory run trace containing digests rather than unbounded values.
• Node 24 and Cloudflare Workers compatibility builds using public package exports.
• A provider-neutral generate/compile/two-repair pipeline with canonical attempt records, recorded-model fixtures, and resumable hidden-case semantic scoring.
• Deterministic typed validation of missing-operation, denied-capability, and insufficient-budget witnesses before an unplannable result is credited.
• Central model-output parsing, digest-bound experiment manifests, enforced inference caps, and paired held-out research gates with confidence intervals.
• A Node-only Vercel AI SDK 7 adapter package whose primary comparison is direct OpenAI Responses and Anthropic Messages, with optional Anthropic-through- Bedrock support, frozen pricing, and worst-case spend reservation.
• A phase-aware Node experiment controller with shared campaign budgets, zero-network dry-run, append-only ledger accounting, safe resume, and offline reporting.
• A Worker-compatible restricted capability-TypeScript compiler/interpreter that never evaluates model source, exposes only registered capability calls, and records matched static and runtime resource measurements.
• A disjoint paired M2 corpus and independent campaign controller for functional IR versus restricted capability TypeScript, including frozen counterbalanced schedules, paired statistics, content-addressed resume, and shared conservative accounting.
• A bounded substrate-neutral evidence-selection contract with four factorial text/graph arms, bitemporal facts and cited relationships, evidence-path and citation ground truth, negative controls, and a counts-only M3a.1 audit.
• Offline M3b matched execution infrastructure with one shared compiled plan, arm-blinded oracle requests, digest-bound Williams scheduling, symmetric transport retries, safe resume, and contrast-specific paired statistics.
• M3b.3 live-binding infrastructure with typed executable answer obligations, canonical path references, staged diagnostics, direct Responses/Messages transports, conservative attempt-level accounting, an append-only ledger, and durable content-addressed resume.
• An offline M4 evidence compiler with a content-addressed provider/task policy, arm-blinded selected contexts, graph-facts controls, and deterministic provenance reconstruction from answer values plus supporting fact IDs.
• A cohesive lexical-default production evidence runtime with typed answer/citation/provenance results, bounded injected oracle effects, content-addressed record/replay, and optional TypeGraph storage.
• Forty-two content-addressed benchmark cases spanning four unrelated catalogs, including bounded recursion and intentionally impossible policies.
• An experimental offline M6 compositional harness that normalizes successful plans, calcifies validated typed strategy templates, fails closed on semantic or authority drift, eliminates an injected planner on stable hits, and groups sanitized trace identities without provider access. Trusted versioned catalog roles and a finite application-supplied conformance runner reject cross-catalog semantic drift.

Deterministic orchestration is not semantic determinism. Exact replay comes from recorded effect results bound to plan, catalog, operation, invocation, effect, and input identities, plus injected time and run identifiers.

Public-alpha quickstart

The supported product API is @nicia-ai/lachesis-runtime:

import {
  compilePlan,
  createRecordingOracleInterpreter,
  run,
} from "@nicia-ai/lachesis-runtime";

const compiled = await compilePlan(planJson, catalog, policy, obligations);
if (!compiled.ok) return compiled;

const completed = await run({
  executablePlan: compiled.value,
  publicTaskContract,
  inputValues,
  trustedPolicy,
  evidenceStore,
  snapshot,
  oracle: createRecordingOracleInterpreter(hostOracle),
  recordingStore,
  signal,
});

See the public-alpha guide for the complete tutorial, package selection, API reference, trust boundaries, record/replay semantics, TypeGraph integration, compatibility, security guidance, and alpha policy. Every example under examples/m5-alpha runs offline.

The synchronized six-package alpha.5 release is complete after alpha.4 closed as a CLI-only partial publication. See the release notes and final closure. The experimental CLI remains ESM-only on Node 24 with no supported JavaScript import or declaration surface. Its packed README is self-contained, and the frozen release notes summarize the three-command manifest, finite conformance, and detached verification workflow. M8a's registry-only incident-response consumer remains the underlying adoption evidence. Neither finite conformance nor detached report verification is a model-quality or compositional-generalization claim.

Kernel CLI example

corepack enable
pnpm install --frozen-lockfile
pnpm build

node apps/cli/dist/cli.js validate fixtures/plans/document-claims.valid.json
node apps/cli/dist/cli.js analyze fixtures/plans/document-claims.valid.json --json
node apps/cli/dist/cli.js canonicalize fixtures/plans/document-claims.valid.json
node apps/cli/dist/cli.js run fixtures/plans/document-claims.valid.json \
  --inputs fixtures/inputs/document-claims.json \
  --replay fixtures/effects/document-claims.replay.json \
  --json

The example maps a recorded claim-extraction effect over a collection whose maximum cardinality is three, folds claims with a property-tested canonical union reducer, and performs one recorded synthesis effect. Analysis proves a maximum of four calls before execution. The supplied two-fragment run consumes three.

Invalid examples demonstrate capability denial, branch mismatch, unbounded mapped effects, and non-decreasing recursion under fixtures/plans.

Workspace

• packages/kernel — portable public package @nicia-ai/lachesis.
• packages/runtime — supported public-alpha facade @nicia-ai/lachesis-runtime, with a separate Node-only subpath.
• packages/generator — provider-neutral package @nicia-ai/lachesis-generator; no live provider SDKs.
• packages/generator-ai-sdk — Node-only live-provider adapter package; kept outside the kernel and portable generator.
• packages/evidence — portable substrate-neutral evidence contracts and offline research substrates; explicitly experimental and without a TypeGraph dependency.
• packages/evidence-typegraph — optional public TypeGraph 0.38 adapter; its managed SQLite subpath is Node-only.
• apps/benchmark — private Node-only M1b/M1c/M2 campaign controller and CLI.
• apps/cli — experimental public lachesis binary for offline catalog manifests, finite conformance, and detached report verification.
• fixtures — valid/invalid plans, inputs, and effect recordings.
• compat — built-package Node and Workers consumers.
• docs — architecture and material ADRs.

See the architecture for the kernel guarantees, the benchmark protocol for generation and scoring boundaries, the M1b runbook for the frozen M1 study, and the M2 representation ablation for its completed comparison boundary, and the M3a.1 evidence design for the offline four-arm substrate benchmark and M3b kill gates, and the M3b offline protocol for the unexecuted matched study substrate, M3b.1 live binding for the frozen failed probe, the M3b.2 correction, and the M3b.3 semantic-obligation correction for the frozen calibration failure, M3b.4 structured-output forensics, and the historical M3b.5 held-out freeze. M3b.5 and M3 are closed as complete-formal-fail; see the immutable M3 results for the formal conclusions and claim boundary. The offline M4a/M4b evidence compiler treats all M3 results as development evidence and creates no campaign or held-out identity. The optional M4c TypeGraph adapter passes storage, temporal-replay, and provenance parity without changing any model-visible evidence or making a model-quality claim. The offline M4d.0 policy viability audit rejects the existing coarse M4a policy and retains only a narrow development-derived candidate for possible confirmation on a completely fresh corpus. The offline M4d.1 protocol and power design implements the reduced oracle boundary and exact paired design, but stops before corpus generation because the conservative powered sample is impractical. M4 is closed as complete-mixed, and M4d.1 is complete-design-no-go; see the immutable M4 results. M5 is closed as complete-operational-pass; see the immutable M5 results. M5c prepares the offline public alpha without publishing a package or authorizing another provider call.

M6a–M6d adds an entirely offline experimental strategy-template and catalog conformance surface to @nicia-ai/lachesis-generator; it does not widen the supported runtime facade, train a model, dispatch a provider, or claim learned cross-domain generalization. Its paired discovery-versus-template study design is a no-go because the distribution-free requirement exceeds the practical case ceiling and empirical power and maximum cost remain unknown. See the compositional harness design and immutable M6 results. M6 is closed as closed-offline-design-no-go. The synchronized 0.1.0-alpha.2 candidate is described in the alpha.2 release notes.

M7a adds an entirely offline public-package consumer trial for independent catalog conformance. Its role-simulated slice covers three unrelated catalog families, blinded adjudication, hostile near-equivalence, catalog evolution, deterministic reports, and external-author documentation. See the prospective plan, author guide, and M7a results. The result is a narrow GO for a larger offline independent-author study, not for inference, promotion, M8, or a generalization claim.

M7b hardens the rejected-pair diagnostic path without changing conformance acceptance. A fresh nine-case development corpus achieves zero false equivalence, zero unsafe repair directions, exact localization for every case, and explicit non-substitution for every genuine semantic mismatch. M7c never executed: operational-isolation gates stopped preparation before any author, adjudicator, analysis, provider, or model role ran. M7 is closed complete-mixed; M7a and M7b are finite offline passes, while M7c is closed-unexecuted-operational-isolation-no-go. See the M7 closure, M7b results, and frozen M7c protocol. The synchronized alpha.3 packages were published with the diagnostic API unchanged; see the public-alpha closure addendum.

Validation

pnpm format:check
pnpm typecheck
pnpm lint
pnpm test
pnpm test:coverage
pnpm build
pnpm smoke
git diff --check

Roadmap

The project still excludes conventional or bounded-general CodeMode, model-facing graph-native traversal, joins, boundedExplore, optimizer rewrites, durable/distributed scheduling, general adaptive loops, and deployment. M3b.1 is closed as a protocol failure; its calibration and held-out identities were never executed. M3b.2 is frozen as a semantic-gate failure. M3b.3 passed its probe and is frozen as an immutable calibration failure; held-out was not authorized. M3b.4's stress probe and calibration are complete and immutable; calibration achieved 236/240 first-attempt success and 240/240 final reliability. M3b.5 subsequently completed all 2,560 held-out records. Its negative-control and safety gates passed, but every universal structural superiority conclusion failed under the frozen provider-by-repetition rule. M3 is complete and may not be rerun or reinterpreted under relaxed gates.

M4a/M4b and the optional M4c TypeGraph storage adapter provide an offline deterministic evidence-runtime substrate. The original M4 adaptive policy is development-rejected; its narrow exploratory replacement remains research-only and unconfirmed because the frozen powered design exceeded the practicality ceiling. Lexical evidence remains the production default. The M5 roadmap focuses on a cohesive production evidence runtime, natural-workload reliability, provenance completeness, replay, budgets, and developer experience—not graph superiority. The offline M5a vertical slice now composes the portable lexical-default runtime, deterministic provenance, and exact record/replay with optional TypeGraph storage. M5b.1 closed as complete-operational-pass: its replacement probe and 24-record production pilot passed, while the original M5b.0 integrity-failed probe remains immutable report-only history. See the M5b design and M5 results.

M2 is complete and closed as a valid formal failure. Its historical M2.2 protocol failure was corrected before the completed M2.3 probe, calibration, and frozen held-out study. See the immutable M2 results for the preregistered conclusion, descriptive operational findings, and claim boundary. No M2 phase may be rerun or reinterpreted under changed gates.

Licensed under Apache License 2.0.