Fail closed
A claim that cannot be verified is UNVERIFIABLE — it is never silently accepted, and never downgraded to "refuted" without evidence. Missing anchors are flagged, not forgiven.
The Trust Stack
6 instruments.
6 instruments.
Not one product.
Six independent instruments that make AI systems auditable — from why the code exists to whether it actually worked. Rules as cited data. Deterministic decisions. Machine‑verified claims. Cryptographic attestation. Tamper‑evident outcomes.
append-onlyEd25519fail-closedhuman-gatedreplayable
Why it exists
Every question an auditor asks. One instrument per question.
AI systems are being deployed into regulated, high-stakes work faster than anyone can verify them. Isonomai is the architecture for the other direction: systems built to be examined.
Intent
Why does this exist?
Every non-obvious decision is recorded in a git-native ledger. The answer is quoted and cited — or the system says "not recorded." It never invents a reason.
Rules
What rule applies?
Legal and compliance rules are cited, versioned, dated data — never model memory. Every rule carries its source, its effective date, and its verification trail.
Decisions
Who decided — and can you rerun it?
Decisions are deterministic and human-gated. Every outcome reproduces exactly from pinned rule versions and inputs. AI suggests; an authorized human converts.
Claims
Can you prove that claim?
Any checkable claim — "tests pass," "the docs match the code," "this quote is real" — gets a verdict computed from executed evidence, never from model assertion.
Attestation
Who signed it?
Ed25519 signatures over canonical JSON, Merkle transparency logs, and offline public-key verification. You don't have to trust the operator — you can check.
Outcomes
Did it actually work?
An append-only, hash-chained, bitemporal ledger records what each system claimed and what actually happened — so calibration is measured, not asserted.
The architecture
Six layers. Independent by design.
Each layer is a separate property with its own repository, license posture, and test suite. They interoperate through open formats — never shared code — so the verifier stays independent of the verified.
Exhibit AThe Spine · six layers, folio 01–06
Intent wherefore
The decision ledger. Why-trailers and git notes make the reason behind every line recoverable, quoted, and cited.
Rules pandect
Codified rule corpus served as an API, with a fail-closed machine-audit pipeline behind every rule.
Decisions rulecore
A deterministic compliance engine where AI output is typed as a suggestion and only a logged human review makes it real.
Claims touchstone · groundwork · claimcheck
Prove-it-or-break-it verification: verdicts computed from executed, hash-sealed evidence.
Attestation agent-exchange · proof-layer
Cryptographic identity, transparency logs, and signed manifests — verifiable offline by anyone.
Outcomes aletheia
The outcome ledger. Append-only, hash-chained, bitemporal — the ground truth every other layer answers to.
The doctrine
Four rules, enforced in code.
01
02
Humans hold the gate
AI output is typed Suggestion<T> and cannot reach a live action unreviewed. Real-world effects are dry-run by default and go live only with a logged, authorized token.
03 · The model is never the source
Models draft and search; sources, executions, and signatures decide. Anything the model asserts without evidence is treated as unwritten — the throughline beneath all four rules, enforced by type systems, triggers, and gates, not by policy documents.
Not of a rule. Not of a why. Not of a verdict.
04
Everything replays
Decisions reproduce exactly from pinned rule versions and inputs — no clock, no randomness, no hidden I/O. Ledgers verify offline from hashes and public keys alone.
Built to be examined. So examine it.
679codified rules served in production
418rules machine-audited in a single wave
6independent verification layers
15+siloed properties, wired by open formats
"A verdict is computed from executed evidence — never from model assertion."
In the field
The stack at work.
Every application carries an honest maturity label. On a diligence site, "in validation" means in validation.
Live
Dry-run
In validation
Pandect API
679 codified consumer-protection and legal rules served in production, each carrying citations and verification status.
pandectai.comRuleCore verticals
Seven compliance verticals — wage & hour, export controls, DTC alcohol, E&S underwriting, and more — running dry on the deterministic spine.
draft packs, pending attestationCountersign & Gatepass
Attorney-countersigned drafting rails, and the evidence layer that helps AI agents pass insurer and conformity gates.
gated: signed letters, not hope
For buyers
The instruments, packaged.
Three products built on the stack, each sold on its own — independent by design, one discipline. Maturity labels are honest: nothing below is purchasable until its gate clears.
Launch-ready
Verdict
Proof that your AI behaves. Continuous adversarial testing — jailbreaks, prompt injection, PII leakage, drift — rolled into audit-grade evidence mapped to the EU AI Act, NIST AI RMF, and ISO 42001.
Pulse $199 · Assurance $799 monthly, at launch
Launch-ready
ShipSafe
Your app was built by AI — was it secured by anyone? Scans the repo, the live site, and the database; grades it A–F; explains every finding in plain English with what could happen and the exact fix.
$149 per scan, at launch
Counsel-gated
Regmap
Every AI system you run, mapped to the rules it answers to. An inventory and gap analysis: which regimes apply, what obligations follow, which required artifacts are missing — from a deterministic engine, not model memory.
opens after attorney review of every rule
The difference
Standard AI stops at plausible.
Standard AI stops at plausible.
Isonomai goes further.
Verbatim anchoring
Quotes, amounts, dates, and citations must match their source character-for-character. A paraphrase is not evidence. An unanchored claim doesn't ship.
The machine ceiling
Models cannot self-certify. Machine audit takes a rule to "needs review" at best — promotion past that ceiling requires human attestation, on the record.
Offline verification
Hash chains, canonical JSON, and public keys mean a third party can verify the ledger with no access to our systems — and no trust in our word.
Diligence
Don't take our word for it.
Don't take our word for it.
That's the whole point.
The dossier covers every property: what's live, what's shipped, what's still in dry-run — with the receipts to check us.
Retained engagements begin at $50,000/month · serious inquiries only