Isonomai
The Stack

Six layers, in depth.

Each layer answers one question an auditor will eventually ask — and each is an independent property with its own repository, tests, and license posture. What follows is the honest state of each one.

Folio 01 / 06The Stack — Intent
01 · Intent

wherefore Shipped

The decision ledger. Six months from now, someone will ask why a guard, a magic number, or an odd special case exists. wherefore makes the answer recoverable instead of archaeological.

  • Git-native storage. Reasons live as commit trailers and git notes — no external database, nothing to migrate, travels with every clone.
  • Quoted or silent. A lookup returns a cited reason from the ledger, or exits "not recorded." It never synthesizes a plausible-sounding why.
  • Anchored to code. Reasons attach to specific line ranges and are drift-tracked as the file evolves.
  • Fleet-searchable. One command searches every repository's ledger at once when you remember the decision but not its home.
Why: trailers + refs/notes/why · zero-infrastructure
Folio 02 / 06The Stack — Rules
02 · Rules

pandect Live

Legal and compliance rules as cited, versioned, dated data — served as an API, with a fail-closed machine-audit pipeline standing behind every rule. The model is never the source of a rule.

  • 679 rules in production. Consumer-protection and legal rules served live at pandectai.com, each carrying its citation and jurisdiction.
  • Fail-closed audit. Every rule is source-audited with verbatim anchoring: the quoted rule text must match the controlling source, or the rule fails.
  • The machine ceiling. Automated audit promotes a rule to "needs review" at best. Trust-class promotion beyond that requires human attestation — models cannot self-certify.
  • Wave-scale throughput. 418 rules machine-audited in a single orchestrated wave, with resumable manifests and per-rule audit trails.
  • Independent cross-review. Audited rules can be sent to external models for blind adversarial review, with structured verdicts recorded in the trail.
  • Computable downstream — Nexus. The corpus compiles into machine-executable deadline and amount specs, each pinned to its rule by content hash and verified against a live evaluator. nexus-isonomia.vercel.app ↗
88 tests · audit pipeline v1.1.3 · trails committed
Folio 03 / 06The Stack — Decisions
03 · Decisions

rulecore Shipped · verticals dry-run

A deterministic compliance engine with a typed boundary around AI. The pipeline is Intake → Assist → Gate → Engine → Guard → Audit — and the types make the safety property structural, not procedural.

  • AI is boxed at the type level. The model can only ever return Suggestion<T>. Nothing downstream accepts it.
  • One path to real. Only an authorized human review converts a suggestion to Reviewed<T> — the sole input the engine takes.
  • Deterministic core. No clock, no randomness, no I/O inside the engine. Every decision reproduces exactly from (ruleset version, table version, inputs).
  • Dry-run by default. Real-world effects — bind, transmit, file, release — execute only with a logged, explicit live token.
  • Deterministic money math. Dollar amounts never touch floating-point arithmetic — replay requires exact, not approximate.
  • Immutable audit. Every stage emits content-hashed, tamper-evident events.
7 verticals dry-run: wage/hour · export · DTC alcohol · E&S · claims · prior-auth · customs
Folio 04 / 06The Stack — Claims
04 · Claims

touchstone · groundwork · claimcheck Shipped

Prove it or break it. Three instruments at three altitudes, sharing one rule: a verdict is computed from executed, hash-sealed evidence — never from a model's opinion of itself.

  • touchstone — the general engine. Plans checks, executes them in a controlled runner (the model never touches the shell), and gates on sealed evidence. Ships claim-batch@0, an open interchange format for anchored claims.
  • groundwork — the zero-dependency grounding library. Deterministically verifies that quotes, amounts, and dates appear verbatim in their source. No LLM in the loop at all.
  • claimcheck — the drift scanner. Reads what your docs claim (commands, flags, behaviors) and checks it against what the code actually defines. MIT-licensed.
  • UNVERIFIABLE is a verdict. A claim that can't be checked is flagged as unverifiable — never quietly passed, never falsely refuted.
claim-batch@0 open spec · runner+gate, zero model calls on anchored claims
Folio 05 / 06The Stack — Attestation
05 · Attestation

agent-exchange · proof-layer Shipped · prototype

One cryptographic atom, applied consistently: Ed25519 signatures over canonical JSON, hash chains and Merkle trees, offline public-key verification. Timestamps are passed in and signed — never read from a clock you'd have to trust.

  • Identity is a public key. In agent-exchange, an operator is its Ed25519 key — twelve packages, an identity registry, and a Merkle transparency log with a standalone verifier.
  • Cross-language parity. TypeScript and Python implementations are held to byte-identical signatures by parity tests.
  • Signed media manifests. proof-layer verifies content as MATCH, TAMPERED, or UNSIGNED — and documents its own current gaps (identity binding, trusted timestamps) rather than papering over them. Prototype status, labeled as such.
  • Verify without us. Every artifact checks out from hashes and public keys alone — no API call to our systems, no trust in our uptime or our word.
Ed25519 + canonical JSON + Merkle · TS↔Python signature parity
Folio 06 / 06The Stack — Outcomes
06 · Outcomes

aletheia Live

The outcome ledger — the ground truth every other layer answers to. Systems emit claims about the future; reality gets observed; resolution rules decide who was right. Calibration becomes a measurement, not a marketing sentence.

  • Append-only, enforced in the database. Updates and deletes raise errors at the trigger level. History cannot be edited, only extended.
  • Hash-chained. Every event links to its predecessor; a verifier walks the chain and flags tampering or gaps deterministically.
  • Bitemporal. When it happened and when we learned it are recorded separately — so hindsight can't be laundered into foresight.
  • Three verbs only. Emit, observe, resolve. The ledger stores references and hashes, never copies — and computes nothing about law or scoring.
  • Already in use, not hypothetical. External systems write real claims to the ledger in production today.
live production ledger · aletheia-ledger.vercel.app
"The verifier must stay independent of the verified. So the layers share formats — never code, never a database, never a deployment."
Architecture ruling, on the record — the reason this is a stack of instruments, not a platform.
See the applications Open the maturity ledger