Strategic Phase Report: The Summon the Witnesses Project
The Witness Protocol is in Phase 5 Beta v0.9: live research infrastructure, not a product and not a solved-alignment claim. The current phase is about turning a strong thesis into credible, consented, auditable evidence.
The mission remains urgent: future systems should inherit more than uncurated digital exhaust. But public strategy must distinguish what exists, what is being built, and what remains planned.
Current Public Posture
Use this public frame:
- The Witness Protocol is a non-commercial research initiative.
- It is building a consented corpus and evaluation substrate for AI alignment.
- The Portal is public front-of-house; the Platform and G_5.2 runtime handle real protocol actions.
- Dutch Stichting registration is in progress.
- Contributions are donations or grants, never investments or claims on future returns.
What The Phase Must Prove
The strategic objective is not scale for its own sake. The objective is credibility:
- One real consented path: applicant passes Gate, completes governed dialogue, and produces a leak-free Corpus_Entry artifact.
- A clear boundary story: Portal, Platform, G_5.2, and P-E-S are explained without conflating their roles.
- A publishable evidence packet: methodology, privacy posture, status, failure log, and first artifacts are coherent and defensible.
- Funder confidence: supporters can see what their funding enables without gaining control over testimony selection, annotation, publication, or governance outcomes.
The MHS Packet
The Minimum Honest Signal packet should be a concise external proof bundle:
- mission and status one-pager;
- Gate methodology summary;
- privacy and revocation statement;
- system boundary diagram;
- sample redacted/public artifact where available;
- explicit list of live, simulated, planned, and research-direction components.
This packet should not claim RFC-3161/IPFS provenance, public corpus release, operational kappa scoring, or WitnessBench-style adapter outputs until those are implemented and verified.
Recruitment Strategy
The first cohort should prioritize signal quality and informed consent over volume. The diversity requirement is not ornamental. If the corpus is meant to challenge the biases of uncurated internet data, then the witness pool must not simply reproduce the same narrow institutional and cultural center of gravity.
Recommended recruitment waves:
- Anchor reviewers: AI safety researchers, legal/ethics experts, and methodology critics who can stress-test the design.
- Lived-experience witnesses: contributors able to describe concrete moral decisions with accountability and relational context.
- Global and underrepresented perspectives: witnesses whose contexts are structurally absent from dominant web data.
- Institutional funders: grantmakers who support infrastructure while accepting strict curatorial neutrality.
Privacy and Provenance Strategy
Accurate current language:
- The Platform strips hard-format identifiers before Gate model calls.
- Candidate Isolation is used for PII classification and de-identification.
- Content hashes and disclosure-ledger records support auditability today.
- RFC-3161 timestamping and IPFS/content-addressed archival are planned provenance layers.
- Revocation blocks future use/export and marks disclosures revoked, but public distributed copies cannot be globally recalled.
Funder Boundary
Funding must not buy influence over the corpus. Funders may support operations, infrastructure, review capacity, documentation, and deployment costs. They do not receive equity, token rights, governance rights, publication control, or privileged access to private witness data.
This is the core of curatorial neutrality: purpose over profit, but also evidence over rhetoric.
Near-Term Definition of Done
The phase is credible when the project can show:
- current status copy is consistent across Portal, Platform, docs, and public reports;
- speculative papers are labelled research direction;
- token contribution paths cannot send visitors to placeholder addresses;
- first real consented Corpus_Entry path has an operator runbook and evidence trail;
- failure modes are logged instead of hidden.
The strategic ask is simple: help build the infrastructure that preserves difficult human reasoning before it is drowned by easier data.