For research teams using AI agents
Useful research shouldn’t vanish with the session. Proofpress keeps the finding, its evidence, and the review decisions behind it—so the next person or agent can pick up the work.
Open-source kernel. Explicit review. Inspectable history.
Synthetic research workflow · fresh reader at each step
The fresh reader receives all four approved findings. B relies on A; C relies on B. U is independent.
Replay of an actual local kernel run using invented findings and scripted reviews. This page displays recorded states; it does not run a live agent. Inspect the evidence ↗
01 / The product
Proofpress is an evidence ledger for teams doing research with AI agents. It connects a finding to its sources, tested conditions, review decisions, and the work that depends on it.
Record the conclusion, the evidence behind it, and the conditions under which it was tested. A failed experiment can be useful knowledge too.
Agents propose work. An authorized reviewer decides what becomes approved context. A past approval stays attached to the version it actually reviewed.
When a recorded finding is withdrawn, its recorded dependents are withheld from fresh approved context. Recovery requires reassessment; unrelated work stays available.
02 / The open format
AFF — Agent Findings Format packages a bounded conclusion with scope, evidence references, attributed checks, and revision history. A receiving team makes its own decision about reuse.
The v0.1 draft, offline verifier, and bounded Proofpress exporter are available. Hosted intake is under review; independent interoperability validation remains ahead.
Explore the format & contribute ↗conclusion what was learned scope where it applies evidence what supports it checks what was verified history what changed local adoption the receiver decides
Public name: AFF. Existing files and the wire contract use OAFF v0.1. Package integrity does not establish truth, identity, or permission to reuse.
03 / Inspect the work
STATUS CHECKED / 30 SEP 2026
These are implementation results, not customer savings claims. The demo uses explicit dependencies and synthetic reviewers; it does not test live-agent behavior, hosted authorization, or automatic detection of external changes. Production deployment and real-workflow evaluation have separate gates.
04 / Work with us
Does your next researcher have to reconstruct what the last one learned? Let’s take one recurring workflow and test whether Proofpress makes that handoff better.
Start with a redacted example. Scope, access, price, and delivery timing are agreed before kickoff.
05 / Contact
Partnerships & commercial
Co-founder & CEO
Tommy brings experience building service businesses and leading operations and go-to-market work. At Only Then Labs, he connects customer workflows with product requirements for agent continuity, authority and recovery.
Talk to Tommy about pilots and partnerships.
Email Tommytommy@onlythenlabs.com
Tommy on XProduct & engineering
Co-founder · Product & Engineering
Richard created Proofpress and leads its core architecture and engineering. His work spans agent workflows, evaluation harnesses and production machine-learning systems.
Talk to Richard about Proofpress and integrations.
Email Richardrichard@onlythenlabs.com
Richard on LinkedInResearch & evaluation
Co-founder & Chief Science Officer
Oliver is a UC Berkeley PhD candidate and the creator of TRACE, which records decisions in AI-assisted research. He leads experiment design, evaluation methods and research partnerships at Only Then Labs.
Talk to Oliver about research and evaluation.
Email Oliveroliver@onlythenlabs.com
Oliver on LinkedIn