A Better Dataset for AI Alignment
The Witness Protocol is building a permissioned, high-signal corpus of human testimony for AI alignment research.
Not mass collection. Not scraped opinion. A smaller, more deliberate dataset built from reflective human reasoning, structured dialogue, and auditable synthesis.
The Thesis
Current AI systems are trained on enormous volumes of human output. That gives them breadth, but not necessarily judgment.
The Witness Protocol starts from a narrower and more useful premise:
"A carefully curated body of human testimony may preserve forms of reasoning that raw internet-scale data usually flattens or loses."
- Tradeoff handling
- Ethical self-location
- Relational awareness
- Reflective restraint
- The ability to reason under uncertainty without collapsing into slogans
The project is not trying to replace large-scale data. It is testing whether a smaller, permissioned, better-structured corpus can become a meaningful corrective input for alignment research.
Why This Matters
AI alignment does not only need more data. It needs signal that remains legible under scrutiny.
A large portion of public data is optimized for reaction, repetition, visibility, and speed. The Witness Protocol is exploring the opposite direction: slower inputs, clearer consent, stronger provenance, and outputs that can be reviewed as research artifacts rather than absorbed into opaque systems.
Can a small, auditable corpus of reflective human testimony become useful alignment infrastructure?
What Makes This Different
Selective Intake
Not every submission enters the process. The system is designed to protect signal quality from the beginning.
Structured Inquiry
Accepted contributors move through a guided dialogue process designed to reach the reasoning beneath the answer — not just the answer itself.
Annotation and Synthesis
The project preserves not only what was said, but how it was reasoned, where tensions appeared, and what ethical structure was present.
Governed Artifacts
Outputs become reviewable research artifacts — not one-off prompts or loose notes absorbed into opaque systems.
What the Project Is For
The long-term aim is to create a body of testimony that can support work such as:
Alignment-oriented evaluation
Qualitative benchmarking
Prompt and policy stress-testing
Synthesis research
Future dataset and fine-tuning experiments where provenance and reasoning quality matter more than scale alone
This is not a claim that alignment is solved. It is a claim that dataset quality is underappreciated — and that some missing signal may have to be gathered deliberately.
A Smaller, Cleaner Inheritance
The Witness Protocol is not trying to solve the whole problem in one gesture.
It is trying to build one missing layer well: a smaller, cleaner, more humanly legible inheritance for future systems.
Those who recognize the difference will know where to look next.
The Protocol
The Witness Protocol is currently in an alpha-stage build focused on proving the method end to end.
The immediate goal is not scale. It is to run a small, controlled, auditable flow that can be evaluated honestly: intake, selection, dialogue, governed storage, annotation, synthesis, and export.
Current Build Focus
The current system is being shaped around a narrow operational path.
01
Selective intake and review
02
Accepted-witness cohort handling
03
Structured dialogue through the Witness flow
04
Governed storage and consent boundaries
05
Annotation and synthesis of accepted testimony
06
Export of reviewable downstream artifacts
The project is being built to make the process inspectable, not theatrical.
How AI Is Used
AI is not being used here as a replacement for human judgment. It is being used as an instrument inside a bounded protocol.
Intake Support
Models help filter low-signal submissions and support early qualitative review.
Structured Inquiry
The dialogue layer guides accepted witnesses beyond surface opinion and toward the reasoning underneath it.
Synthesis Support
Accepted testimony is turned into structured, reviewable artifacts rather than remaining only as raw conversation.
Annotation Support
AI assists in organizing, surfacing, and comparing patterns inside accepted material, while preserving the ability to inspect the resulting outputs.
Evaluation and Stability
The system is pressure-tested so the protocol remains legible, bounded, and consistent as prompts, providers, and downstream components evolve.
AI is being used to help convert fragile human signal into auditable research material.
For Providers and Research Partners
The Witness Protocol is not trying to maximize engagement volume. It is trying to increase the quality of the data pathway.
Support at this stage does not vanish into generic product growth. It directly strengthens:
Intake and screening quality
Structured dialogue quality
Synthesis and annotation quality
Evaluation and regression discipline
The first auditable corpus slice
This is the kind of work that remains small for a long time — and useful because it stays precise.
The Right Collaborators
For those tracking where alignment work becomes method rather than slogan, this is an appropriate place to begin.
Some will read this as a research concept. Others will recognize an instrument taking shape.
The right collaborators rarely need a hard invitation. They tend to recognize timing when it appears.