Case studies
Evidence for consequential agent decisions.
Bounded technical accounts of failures, evaluation contracts, remediation, and release evidence.
Field notes as of October 23, 2026
Four days of making our own agents accountable
Early operator observations from applying evidence, ownership and tested corrections to our own engineering agents. Not a controlled productivity result.
By Nikita Annt, CTO Dharma AI Labs
Read the account →Study completed September 13, 2026; report cutoff September 21, 2026
34 passes, and the engineering stories behind them
Retained experience, live repair and independently graded work products in a selected-subset comparison. Reliable superiority over the controls remains unestablished.
By Nikita Annt, CTO Dharma AI Labs
Read the account →Studies June 28-July 1, 2026; technical review September 9, 2026
False-premise resistance: what changed when we added evidence state
A bounded stateful result alongside an earlier prompt-gate intervention and its negative replication. Different cohorts and scoring methods remain separate.
By Nikita Annt, CTO Dharma AI Labs
Read the account →From agent output to verified progress
What we are learning by building with our own software factory
Read the account →We got Jev to price 91 of 100 garments. Gemini still scored higher.
Decomposition raised Jev’s regular score from 20.90 to 50.20. Retained Gemini + Atlas still led at 75.79. The original study and its follow-up, kept separate.
Read the updated findings →Jev and OpenAI Decisions: why valid answers still need proof of improvement.
A measured learning gain, a rejected Atlas candidate, and a stronger estimator that still failed its expansion gate. What our appraisal trials teach about Cognitive Integrity.
Read the evidence →PixelLeak and the race between approval and upload
A local screenshot-upload prototype reproduced a destination metadata race. A cooperative sender/store protocol remediated it; production protection remains unqualified.
Read the case study →Teaching an AI appraisal system to learn from expert corrections.
An anonymous resale platform supplied 100 image-bearing appraisals and expert values. Dharma turned recurring mistakes into a tested skill for the same model. The independent pricing result is still being scored.
Read the case study →From agent failure to verified improvement.
How a reviewed, versioned remediation moved the same model from 3 of 30 source tasks to 30 of 30 separate held-out tasks in a bounded synthetic benchmark.
Read the case study →