OPCVE - Omni-Process Cognitive Validation Engine

Prove the human did the thinking. Process-based academic integrity that evidences authorship without surveilling content.

Comment

OPCVE — Omni-Process Cognitive Validation Engine

Tagline

Prove the human did the thinking. Process-based academic integrity that evidences authorship without surveilling content.

Inspiration

Generative AI broke output-based academic integrity. 92% of students already integrate AI into their workflows, and post-submission "AI detectors" cannot tell illicit generation from legitimate AI-assisted cognition — they only deter, and they misfire. Institutions are paralyzed by the immense administrative friction of enforcing AI bans they cannot win.

Meanwhile the incumbent reflex — centralize and hoard student work to analyze it — is a liability, not a moat. The 2025 PowerSchool breach exfiltrated the personal information of ~3.86 million people across 20 Ontario school boards, and Ontario's privacy commissioner found systemic vendor over-collection. Any integrity product that centralizes raw student content is one breach from extinction, and squarely on the wrong side of FIPPA/MFIPPA, the 2025 COPPA amendments, and CLOUD Act exposure.

We asked the inverse question: instead of policing the output, what if we preserved the proof of the process? Capture how the work was made — the iteration, the cadence, the disclosed AI help — hash it on the device, and sync only privacy-minimized metadata. Give the learner durable, verifiable evidence that they did the cognitive work, and give the institution a defensible record instead of an accusation.

What it does

An edge agent embeds in the editors, browsers, and IDEs where students already work and captures the creation process — keystroke-timing metadata, edit operations, and AI-assistant disclosures — as content-hashes only, chained into a tamper-evident vault on the device. It syncs compressed, privacy-minimized metadata to a per-tenant sovereign pod, computes a deterministic Process Integrity Score, and issues a cryptographically anchored Authorship Certificate that any third party (an employer, a registrar, an accreditor) can verify offline.

Crucially, OPCVE evidences process; it never classifies output and never applies a sanction. Every consequential decision flows from a human reviewing process evidence. The north-star guardrail is an appeal overturn rate ≤2% — the system must never grow by accusing humans falsely.

It is built for both humans (millions of learners + educators) and first-class AI-agent operators: LMS sync agents, credential-verification agents, and students' own AI assistants disclosing their contributions programmatically.

How we built it

This BUILD is a complete, runnable, tested, production-shaped application — real persistence, the full §10 HTTP API + MCP + a web UI, the full guardrail suite, observability, and deploy artifacts — all running offline against deterministic fakes for every external dependency. npm test, npm run eval, the three demos, and npm run serve all run with zero credentials and zero network.

  • Stack: TypeScript on Node ≥ 22 (ESM), with zero runtime dependencies — node:crypto for all cryptography, node:sqlite (DatabaseSync) for persistence, node:http for the server. No build step (Node strips TS types at load); vitest + a custom linter for the toolchain.
  • Real persistence: a typed SQLite repository with a transactional, append-only migration runner (4 migrations covering every §9.1 entity), tenant-keyed for hard isolation, with seed fixtures for every UI surface.
  • Real cryptography: a tamper-evident hash chain that fails at the exact altered segment, Merkle checkpoints with offline-verifiable inclusion proofs, Ed25519 device/pod/certificate signatures, and a HYOK three-level key hierarchy (Tenant KEK → per-learner LIK → object DEK) with per-learner crypto-erasure and tenant-wide destruction.
  • Deterministic validation + the hard anomalies: a bit-reproducible Process Integrity Score, rule + adjudicated anomaly detection AN-1..AN-6 — including a per-learner cadence baseline and a MinHash + LSH replay index — with a 0.85 confidence gate and the no-suppression invariant.
  • Full §10 surface + agent operability: a real node:http server with OAuth2 client-credentials, staff sessions, scopes, idempotency, cursor pagination, token-bucket rate limits, signed webhooks, and an MCP server.
  • Real web UI for every human journey: Learner Transparency Panel, Educator review queue + Session Report, Appeals workspace, Admin, Public Verify, Developer Portal.
  • Google AI stack: Gemini-class model on Antigravity 2.0 managed-agent runtime, Google AI Studio, Gemini API, Stitch, Flow / Veo, and Pomelli.

Challenges

  • Privacy as architecture: raw content never leaves the device.
  • The no-suppression invariant: model cannot clear a flag.
  • Certify through outages: zero model call requirement under kill switch.
  • Determinism: bit-identical scoring.

Accomplishments

  • Every PRD §5 journey, end-to-end.
  • All 23 P0 functional requirements implemented.
  • 71 passing tests (now 143 passing tests in latest build).
  • Hard anomaly algorithms built for real (cadence baseline & MinHash+LSH replay index).
  • Offline-verifiable Merkle inclusion proofs.

What we learned

  • Compliance posture is the defensible wedge.
  • Evidence, not verdicts, makes the product adoptable by risk-averse institutions.
  • A running thin-slice surfaces real invariants faster than specs.

What's next

90-day plan: ECA vault + ledger live (W1-2), scoring + Educator Console (W3-4), Pilot 1 (800 learners) by W5, CDI + adjudicator + certificates (W6-8), appeals + LTI (W9), billing + pen test (W10-11), and first revenue by W12 (≥CAD 40,000).

Built with

TypeScript, Node.js, SQLite, Crypto, Vitest, Docker, Gemini-API, Antigravity-SDK

Try it out

cd BUILD
npm install
npm test
npm run serve

Open https://major-files-enjoy.loca.lt/

Built With

Share this project:

Updates