Skip to content
VELKRON

TECHNOLOGY

Built like audit
infrastructure.

Three design principles, one stack. Nothing in it asks you to trust a model's opinion of another model.

01PRINCIPLES
T1Grounding over opinion

A citation you can't verify is not evidence. Every rule id is checked against the encoded rulebook and every quoted excerpt against your actual artefact in code, not by asking a model nicely. An invented rule or a hallucinated quote is dropped before you see it. Rule classification itself is model-assisted, not encoded logic, and its sampling settings are whatever the current provider's default model allows.

T2Refusal over guessing

Every check carries a grounding score: how confidently the rulebook covers this case. Below threshold, the system returns REFUSED and escalates. The refusal rate is reported, not hidden. In compliance, abstention is accuracy.

T3Evidence over assertion

A verdict without citations is marketing. Every Velkron verdict names the binding rules it relied on, and the full run is sealed into a receipt that survives audit. The output of verification is a document.

02THE STACK

From rules to receipts to proofs.

Rule engineLIVE
Regulator rulebooks encoded as versioned, testable rule modules. Each rule carries its source citation, so a verdict can always be traced back to the regulator's own text. Adding a jurisdiction is content work on the same engine, not a rebuild: that is what makes UK-first, Gulf-next practical.
Receipt chainIN DEVELOPMENT
Every run is hashed (SHA-256), chained to the previous receipt and sealed. Any edit to any historical record breaks every hash after it, visibly. No blockchain required, no tokens, no theatre: just cryptographic accountability a firm can self-host.
Proof of inferenceRESEARCH
Zero-knowledge circuits (Groth16 over BLS12-381) that prove a committed model produced a given decision on a given input, without revealing weights or data. Machine-checked faithfulness proofs accompany the circuits. Today it covers small decision models; the frontier is scaling the technique up.

See it break.

The receipt chain above isn't a metaphor. Here's the same mechanism running on five sample receipts: edit one, and watch every hash after it stop matching.

RECEIPT CHAIN · ILLUSTRATIVE · SAMPLE RUNS, NOT LIVE DATA

Hover a block for its hash. Each block's hash is derived from its own record and the block before it.

BLOCK 1201

isa_landing_page_v2.html

VALID✓

BLOCK 1202

q3_fund_factsheet.pdf

VALID✓

BLOCK 1203

pension_transfer_advert.docx

VALID✓

BLOCK 1204

crypto_promo_v1.html

VALID✓

BLOCK 1205

promotion_draft_v3.pdf

VALID✓

03ARCHITECTURE TODAY

What's model-assisted, what's code.

The rule engine above is live. Here's exactly where the model's job ends and the deterministic layer's job begins on every check that runs today.

Model proposesMODEL-ASSISTED
A single LLM call reads the artefact against the encoded rulebook and drafts findings: which rule, which excerpt, which verdict. Sampling settings are whatever the configured model allows. Some models we support let us pin them down; the current default provider does not expose that control.
Code enforces groundingDETERMINISTIC
Every finding is checked in code, not prompted for: the rule id must exist in the rulebook, and the excerpt must appear verbatim in your artefact. Anything that fails either check is dropped before it reaches you.
Verdict can't contradict the findingsDETERMINISTIC
If every finding was dropped, an ISSUES verdict is corrected to PASS. If a grounded breach survives, a PASS verdict is corrected to ISSUES. The reported verdict is never allowed to disagree with what actually made it through grounding.
04RESEARCH LINEAGE

The stack did not appear last month.

Each component grew out of a public library or research project by the founder. The open-source toolkit is the paper trail.

certain ↗

Conformal prediction: calibrated uncertainty and principled abstention. The math behind the refusal threshold.

groundkit ↗

Deterministic grounding of LLM output against authoritative sources. The math behind rule citation checking.

disco ↗

Verifier-gated evolutionary search: only proven candidates survive. Verification as a filter on generation.

stigma ↗

Multi-agent consensus with persistent trust. How multiple checkers agree without a single point of failure.

github.com/ss1738 ↗

Technical questions welcome.

Talk to the founder›