Training data for AI systems that need to understand human values —
built from what real people did when it cost them something.
Where values are proven under pressure — not proclaimed in comfort.
Training data for AI systems that need to understand human values —
built from what real people did when it cost them something.
Where values are proven under pressure — not proclaimed in comfort.
AI systems are increasingly asked to make decisions that involve human values — fairness, integrity, responsibility. But the data they train on is mostly self-reported: surveys, opinion labels, crowd-sourced annotations. That data reflects what people say they believe, not what they do when tested. We're building something better: labeled training datasets grounded in behavioral evidence, drawn from the historical record of what people actually did under real conditions.
A value stated in comfort is weak signal. Everyone claims integrity when nothing is at stake. The informational value comes from moments of pressure — when someone held to a principle despite personal cost, or abandoned it because of pressure. Those moments, documented in journals, letters, recorded actions, and testimony, are the raw material for training data that actually means something.
Building alignment systems, RLHF pipelines, or value-sensitive classifiers and need training data grounded in demonstrated behavior rather than self-report.
Evaluating how AI systems represent human values and need a reproducible, auditable pipeline for generating labeled value data from primary sources.
Studying behavioral patterns across figures and eras — which values appear under which conditions, and how consistency holds across document types.
Behavioral Value Extraction Pipeline
Ethos ingests historical text — journals, letters, speeches, documented actions — and runs it through seven independent extraction layers to detect which of 15 core human values are present. Each signal is scored for resistance: the personal cost of holding that value.
Signals are classified as P1 (value held under pressure), P0 (value failed), or APY (value yielded to external force), then exported as structured, labeled JSONL with full provenance — ready for model training, RLHF pipelines, and publication to Hugging Face as open datasets.
Enter Ethos
Value Alignment Certification
Verum takes the behavioral evidence produced by Ethos and turns it into a verifiable certification. Submit text samples, and the system scores them for value alignment: how many values were demonstrated, at what resistance level, and across how much evidence.
When the score meets the threshold — backed by at least five substantial samples — Verum issues a cryptographically signed certificate. No committee, no subjective review. The certificate contains the math, and anyone can verify it independently without an API call. This isn't judgment of character. It's a machine-readable statement of demonstrated values.
Enter VerumEvidence over assertion. Behavioral record over self-report. What was done, not what was said.
Deterministic. Reproducible. Same input, same output. No hidden weights, no opaque judgment.
A value held in comfort is a preference. A value held at cost is character. We measure the cost.
Open source. Verifiable certificates. No API call needed to confirm authenticity. The signature speaks.
“The ultimate measure of a man is not where he stands in moments of comfort and convenience, but where he stands at times of challenge and controversy.”— Martin Luther King Jr.
Trust-Forged generates labeled training datasets from the text you provide. The quality, authenticity, and integrity of every dataset produced by this system depends entirely on the quality, authenticity, and integrity of the documents you feed into it. We do not filter, verify, or judge your input. That responsibility is yours alone.
This is not a toy. The datasets this pipeline produces can be used to train AI models that influence real decisions about real people. If the input is fabricated, misleading, or malicious, the output will carry that forward — with consequences. You are accountable for what you build with these tools.
Use of Trust-Forged requires registration and acceptance of a user agreement. By using this system, you accept sole responsibility for the content you submit and the datasets you generate. Trust-Forged, AI-nhancement LLC, and its operators accept no liability for misuse, misrepresentation, or harm resulting from user-generated output.
We built a system grounded in values. We expect you to use it with yours.