Alpha

Your agent's output
looks right. Is it?

A verification harness for engineering and R&D agents. Detrix checks your agent's output against the physics and math of your domain, so it only learns from work that actually held up.

Request Consultation
Deterministic verification Domain physics, not LLM vibes Verified self-improvement
The Loop
Verify first.
Then let it learn.
Self-improvement only pays off if the agent learns from work that was actually right. Detrix is the harness that makes sure of that before anything gets promoted.
01

Verify

Every output runs through your domain's real checks. Physics and math first, deterministic and repeatable. A language model can weigh in, but it only advises. It never overrules the gate.

02

Admit

Results that pass are admitted with their evidence attached. Results that fail are kept too, with the reason why. Nothing gets through on confidence alone.

03

Improve

Admitted runs feed the replay and training loop, so the agent learns from results that held up. A new version only gets promoted when it proves out on replay, with no safety regression.

Proof Lane
Materials characterization.
AgentXRD and PXRD.
We're proving this first on X-ray diffraction, where correctness is a question of physics, not taste, and a wrong accept can never be allowed to pass. The same pattern fits any engineering or R&D check you can make deterministic, from simulation tolerances to backtests. Here is what a rejected scan and a clean run look like.
detrix · gate rejects an unfit scan
$ detrix run --pipeline xrd-analysis --input scan_042.xy
✓ step/load-data executed (0.3s)
✓ step/preprocess executed (1.2s)
■ gate/metrology REJECT · peak_count: 2, snr: 1.8
→ Scan unfit for analysis. Skipping expensive refinement.
→ Evidence: {"peak_count": 2, "snr_estimate": 1.8, "quality": "bad"}
→ Verdict saved to .detrix/verdicts/r-042.json
Pipeline halted at gate/metrology. No false accept admitted.
detrix · full governed pipeline pass
$ detrix run --pipeline xrd-analysis --input scan_117.xy
✓ step/load-data executed (0.2s)
✓ step/preprocess executed (0.9s)
✓ gate/metrology ACCEPT · peak_count: 14, snr: 42.3
✓ step/retrieve executed (1.8s)
✓ step/score executed (0.4s)
✓ gate/confidence ACCEPT · confidence: 0.94, ece: 0.02
✓ step/refine executed (12.1s)
✓ gate/refinement ACCEPT · R_wp: 4.2%, GoF: 1.3
✓ step/report executed (0.6s)
Run complete · 5 steps, 3 gates passed · Verdict: ACCEPT
Phase: α-Al₂O₃ (94.2%) + SiO₂ (5.8%) · R_wp: 4.2%
Trace admitted → replay queue · .detrix/traces/r-117.jsonl

See it on your data.

Bring a hard engineering or R&D workflow. We'll run your agent's traces through the harness and show you exactly what passes, what doesn't, and why, so the only work your agent learns from is work that was actually right.

Request Consultation
Bring your own domain traces. We run them through Detrix. No slides.