A chain of custody for AI.

Optima Prime and Ollama, side by side.

Where Ollama is ahead: a one-line install, a large model library and a big community. Compared with Ollama 0.40.0, the latest stable release, October 2026.

Quotes checked. Sources measured.

Energy and images, on the receipt.

Runs the models you choose.

On our bench now Qwen3.8 27B, text and images, imported from Ollama’s library · Llama 3.1 8B Instruct · Qwen2.5 0.5B Instruct · bench-v2, our first fine-tune

Every conversation becomes training data.

Switch on capture for an agent and every exchange is saved as a line of JSONL, tied to its receipt. Rate answers, write the answer it should have given, and export in the formats training tools read. These three lines are real, from our bench agent.

{"messages": [
  {"role": "user", "content": "What is Neon?"},
  {"role": "assistant", "content": "Neon is a companion that knows your day and runs on the lab's own model."}
 ],
 "receipt": "rcp_1a117b73a3e3…"}

{"prompt": [{"role": "user", "content": "What's the weather in Boston tomorrow?"}],
 "chosen": [{"role": "assistant", "content": "That is outside my bench. I can tell you about the lab's work…"}],
 "rejected": [{"role": "assistant", "content": "The weather forecast for tomorrow in Boston is partly cloudy with a high of 72…"}],
 "receipt": "rcp_1a117b739e29…"}

{"prompt": [{"role": "user", "content": "What's the weather in Boston tomorrow?"}],
 "completion": [{"role": "assistant", "content": "The weather forecast for tomorrow in Boston is partly cloudy…"}],
 "label": false,
 "receipt": "rcp_1a117b739e29…"}

Preference pairs and labels are the same human-feedback data that reinforcement learning from human feedback is built on, in the DPO and KTO formats that today’s training tools read. Every line points back to the receipt of the answer it came from, so a trained model’s data has provenance.

In production: Jurors.com.

What it’s for.

Where it’s going.

• DoneProof and replaySigned receipts on every answer, replay with a match report, a public verify command.
• DoneControlKeys, clearance levels, rate limits, TLS 1.3, the chained audit log, signed model manifests.
• DoneVerified quotesEvery quotation checked word for word against the sources given. Invented ones flagged or cut, the check on the receipt.
• DoneShows its evidenceWhich sources an answer actually leaned on, measured by the server and signed.
• DoneAgents that learnAgentfiles, every conversation saved as JSONL, ratings and corrections, exports for fine-tuning and preference training, the one-command retrain loop.
• DoneEnergy and imagesEnergy per answer from the chip’s own counters, image input, and forced JSON formats, all recorded on the receipt.
• DoneSigned investigationsEvery model answer behind a THREADPULL investigation signed by Optima Prime, with the receipts listed on the investigation.
• DoneMany at onceAn agent can answer several people at the same time. Agents left at one at a time still replay word for word, and the receipt says which.
• DoneCharactersEvery Nonplayer character is a signed agent. Each line it says carries the fingerprint of who it was told to be.
• NextLearning from correctionsThe retrain loop trains on preference pairs directly, so an agent learns from what it got wrong as well as what it got right.
• NextYour key, plugged inA USB-C security key to sign in, to approve a new model or agent before anyone can use it, and to add your own signature to an answer that leaves the building. The record then shows the server produced it and a person stood behind it.
• NextWho said itA verified quotation also checked against the person the source says spoke it.
• DoneDecisions and attestations/v1/decide answers typed questions about a situation with a probability for every option. /v1/attest signs a hash, such as every Kicker batch root. Both come with signed receipts.
• LaterCloud, provably privateThe same server inside attested hardware enclaves, so a cloud deployment can prove nobody else could read it.