Promethean · Product Datasheet

Document type Executive datasheet · CEO · CISO · GC · Head of AI Length 2 pages Last revised 2026-05-18 (v1.3 — post-launch hardening) Companion docs TECHNICAL_ARCHITECTURE · SECURITY_BRIEF · COMMERCIAL_DATASHEET · PROMETHEAN_THEORY_AND_FINDINGS (canonical theory paper)

What it is

A deterministic substrate that contains, audits, and cryptographically attests every LLM call made inside a regulated software product — at build time AND at runtime. Each AI decision becomes one signed entry in a Bitcoin-anchored hash chain a regulator can verify in 50 KB of plain JavaScript.

What it solves

Every regulated enterprise has the same stuck pilot: ship an LLM feature inside a regulated product without killing audit, liability, or compliance. The current vendor stack is partial:

• Vanta / Drata audit the software around the LLM
• Credo AI / Fiddler audit the model's behaviour
• LangChain / Vellum orchestrate the calls
• None constrain the LLM's output surface by construction
• None carry the call into a regulator-verifiable record

Promethean does both. The same closed-enum + reviewer-queue + signed-receipt pattern that governs the substrate's build-time generators now runs at request-time inside emitted products.

Why this matters now

Three forces converging this decade:

1. EU AI Act binds through 2025–26. DORA effective Jan 2025. NIS 2 in force. PSD3 finalising. Each requires verifiability the current vendor landscape doesn't deliver.
2. LLM code volume grows 100× by 2027. Human review at human throughput becomes the bottleneck.
3. Frontier-model commoditisation. Margin compresses for model providers; margin expands for the deterministic layer they feed.

How it works (in one paragraph)

You declare a RuntimeAISpec: a closed-enum category, an output schema (Zod / Pydantic), a reviewer-gate predicate, and a declared fallback. The SDK enforces the spec at request time. Every call writes one signed entry into an L12 receipt chain (Ed25519, hash-chained, OpenTimestamps-anchored to Bitcoin). A regulator verifies the chain offline by running node verify.mjs — plain JS (~500 lines, exact LoC at /api/facts/verifier-loc), zero dependencies, no Promethean software running on the regulator's side. Theory + threat model in PROMETHEAN_THEORY_AND_FINDINGS.

What ships today (Phase R · v1.3 · post-launch hardening)

What a regulator can verify

Eight independent checks, all offline-runnable, all falsifiable:

1. Hash-chain integrity (entries cannot be retroactively edited)
2. Per-entry Ed25519 signature against published trust anchor
3. Closed-enum schema enforcement at request time
4. PII-safe inputs (only inputHash recorded, never raw)
5. Bitcoin time attestation via OpenTimestamps
6. Cross-language byte parity (Python port verifies JS chains)
7. Continuous reproducibility (reference chains verify across all releases)
8. Mutation-tested test suite (Stryker break-threshold 75%)

Mapped to GDPR Articles 5/22/25/32, AI Act Articles 12/13/14/15, DORA Articles 6/8, NIS 2 Article 21, PSD3. See SECURITY_BRIEF §3.

Deployment options

Pricing has moved to a self-service ladder driven by lib/workspaces.ts TIER_PRICING. Live values are exposed at /api/facts/tier-allowances. Current tiers:

Stripe live mode + Resend + npm publish are in-progress for 2026-06 (see /roadmap). Today the SDK ships via signed tarball at https://promethean.software/runtime-ai/latest.tgz. Full commercial detail: COMMERCIAL_DATASHEET.

Why now, why this team

The architecture is operationally complete — Phase R closed May 13, 2026 (ADR (xx) v1.2 closeout); post-launch hardening (May 17–18) closed an additional 31 adversarial findings. 50 ADRs publicly document every decision with falsifiable predictions. 7,400+ tests across 223 files. Three live regulator-verifiable references. The technical bottleneck is behind us; the next move is the first paying operator with a real runtime-AI workload.

What this is not

• Not a model evaluator. We don't replace Fiddler / Credo AI / Lakera; we record what they evaluate.
• Not a compliance certification body. We produce verifiable artifacts; you retain audit authority.
• Not a competitor to LangChain / Vellum. Those orchestrate; we contain and attest.
• Not a hosted observability platform. The verifier runs on your machine, against your copy of the chain.

Next step

Operator-fit interview · 30 minutes · operators@promethean.software