TESSERAM
The Architecture Behind
Responsible AI
Responsible AI
Nothing executes without authority. Every decision is provable. Denials carry cryptographic standing. Absence of governance is detectable.
Nothing executes without authority. Every decision is provable. Denials carry cryptographic standing. Absence of governance is detectable.
I · Proof
Every governed action — allowed, denied, or blocked — produces an Ed25519-signed receipt on an append-only, hash-chained ledger. Change one byte and verification fails, naming the broken check. Receipts verify offline — no account, no network, no trust in Tesseram required.
Not an observability layer. The gate is a precondition of execution — an ungoverned action is provably non-existent, not merely unlogged.
RFC 8785 canonical JSON, HKDF-derived keys, Ed25519 signatures — the same cryptography behind SSH, Signal, and TLS 1.3. Real, and running.
Court bundles and incident reports generated from the ledger itself — every count independently verifiable by the other side's experts.
II · Regulatory Chapters
One engine, mapped jurisdiction by jurisdiction. Each chapter turns a regulator's mandate into evidence generated at runtime — always with an explicit statement of what is not covered. Precision, never overclaim, is how evidence earns trust.
A signed conformity-evidence dossier mapping your live governance data to named obligations of Regulation (EU) 2024/1689, article by article — citing your own receipts and ledger entries, with the not-covered list stated plainly.
Request dossier →Pre-execution gating mapped to the draft MRM guidance — timed kill-switch, per-decision reasons, deterministic replay, risk tiering, continuous monitoring. We formally responded to the RBI consultation with measurable definitions for every mandate.
Check your readiness →Securities-market AI governance is the next chapter, under specification. Every jurisdiction becomes a mapping over the same engine — the receipts, the ledger, and the offline verifier never change.
III · Architecture
Data classification at the inference boundary, shadow monitoring, consent tokens — what may enter a model, governed.
The staged pipeline: govern → commit → execute → receipt. Execute-once tokens, agent passports with exposure ceilings and trust scores.
Unconditional veto. Kill switch. Blast-radius containment. No system runs that a human cannot stop.
Patent filings in progress across the 14 families. No granted patents are claimed. The cryptography is real and runs in your browser.
IV · The Pentagon
Tesseram's architecture is a pentagon — five governance pillars, each a vertex, bound by the golden ratio. Fourteen invention families map onto them.
Governance as a cryptographic prerequisite. No AI action without structural authorization.
Multi-dimensional risk via composable kernels. Attestation decay. Aggregate veto.
Three governance layers. Seven domain kernels. Unified and edge deployment.
Single-action to batch Merkle to fleet. Authority receipts and agent passports.
Court-admissible evidence. Zero-trust verification. Denial receipts with standing.
One primitive threads through all five vertices, binding the pentagon into a single provable system.
V · Invention Portfolio
The portfolio spans state-transition control, semantic commitment, intrinsic cryptography, metrology, topological kernels, liability, and verification — organised as fourteen families across three governance layers. Patent filings in progress; no granted patents claimed.
Inference-boundary classification, RAG provenance, shadow monitoring. Families governing what may enter a model.
The staged pipeline, per-action receipts, reality commit, authority receipts, execute-once tokens, agent passports.
Unconditional veto, blast-radius containment, kill switch, crystallized liability, court bundles.
VI · Readiness
Twelve questions against the five pillars of RBI's draft Model Risk Management guidance. Answer honestly — the score is only as true as your answers. Two minutes.
Choose a framework above to begin the twelve-question assessment.
This is a self-assessment: it scores the governance posture you report, not a live audit of your models. A live score comes from your models calling our engine on each decision. But if you answered honestly, this is a fair picture of where an examiner would find you today.
VII · Engage
Every engagement begins risk-free: a shadow pilot observes your AI activity, blocks nothing, and receipts everything — including each action it would have stopped, with signed evidence. Read the report. Then decide if the gate goes live.
VIII · Founder
Tesseram began with one observation, made before the regulations existed: when a model rejects a first-time borrower and nobody can say why, that is a person turned away by a machine that owes them no explanation. Everything Tesseram builds flows from that — governance not as bureaucracy, but as the way the powerless get a reason.