About / Methodology · Phase 5 — Beta (v0.9)
A better dataset for AI alignment
The Witness Protocol is a first-party, consented corpus of high-signal human moral-reasoning testimony, built as an evaluation substrate for AI alignment. It preserves how people actually reason through hard moral choices — the struggle, not the polished answer — so future systems can be tested and shaped against it.
It is deliberately small and curated. It is not mass collection, not scraped opinion, and it makes no claim to "solve" alignment. It is research infrastructure: a rigorous, auditable body of testimony and the tools to evaluate models against it.
How it works
From testimony to supervision
The Gate — a multi-tier vetting pipeline
Testimony is not accepted wholesale. It passes an automated sieve for coherence, a qualitative assessment that scores specificity, counterfactual reasoning, and relational density, and finally a blind, dual-rater Human Curation Council. The Gate is the defense against low-signal noise entering the corpus.
Try the Gate self-assessment →The Inquisitor — Socratic inquiry, not a helpful assistant
The dialogue engine drops the subservient-assistant posture for a questioning one: a roughly 70/30 inquiry-to-statement ratio, a 5-Whys forcing function, and steel-manning a witness's position before challenging it. The goal is the strongest version of the reasoning, not a comfortable consensus.
Compare Inquisitor vs. standard LLM →The outputs — model-usable supervision
The human-readable archive is not the endpoint. The live path is building toward model-facing evaluation artifacts: consented Corpus_Entry bundles, witness-attributed eval cases, and later research-direction adapters such as preference pairs, process-reward traces, and rule-based reward rubrics. The current claim is an evaluation substrate, not completed model training.
Trace a record's provenance →Privacy by architecture — consent as a system invariant
Identity and testimony are kept in separate rooms. The Platform strips hard identifiers before Gate model calls, uses Candidate Isolation for PII classification, and keeps runtime testimony behind consent gates. Revocation blocks future use/export and marks disclosures revoked; eligible internal and partner-held copies are removed where technically and contractually possible, while already public distributed copies cannot be globally recalled.
Simulate a consent revocation →
The demonstrations above are explicitly simulated illustrations. The real instruments — Gate intake, the Inquisitor dialogue engine, and consent — live on the live Platform. For the full detail, read the papers and reports in the research library.
Who is behind it
The Foundation
The work is stewarded by The Witness Protocol Foundation initiative, with Dutch Stichting registration in progress and a purpose-over-profit mandate. Contributions are donations or grants toward an auditable public good — never an investment or a financial return.
The Protocol was founded by Martin vanDeursen, an AI developer and system architect based in the Netherlands, who stepped away from commercial AI development to build a high-signal ethical inheritance for future intelligence — architected through structured dialogue and strict operational constraints rather than uncurated data scraping.