Specification
A frozen task, repository snapshot, acceptance criteria and environment identity.
Research program · Reliable and Human-Governed Agent Systems
When an agent says “done,” what evidence should survive?
Pilot-ready · not preregistered
I study the operational layer between a capable model and a trustworthy software change: task specification, skills, memory, tools, verification, release identity and human intervention.
Current material proves the existence of a longitudinal engineering corpus. It does not yet prove that skills improve success.
- 522
- task sessions
- 2,568
- indexed prompts
- 729
- independent QA reports
- 200+
- managed skills confirmed
First paper
From Incidents to Skills
A Longitudinal Field Study of Reusable Procedural Memory in Production Software Agents- RQ1
Which failures become reusable procedural skills, and which produce only one-off patches?
- RQ2
Do skills improve behavior-level acceptance on unseen tasks from the same domain?
- RQ3
When do version, environment, or abstraction mismatches create negative transfer?
- RQ4
Can scope, evidence, expiry, and independent verifiers outperform semantic retrieval alone?
Controlled comparison
Same task. Same budget. Different memory discipline.
Condition order is randomized; graders do not see the assigned condition.- ANoSkill
The same model, tools, harness and budget, without task-specific skills.
- BSemanticSkill
Semantic retrieval without version, evidence, or expiry governance.
- CGovernedSkill
Domain-, version-, and environment-matched skills with provenance and validity.
- DGovernedSkill + Verifier
Condition C plus an independent behavioral verifier that can reject completion.
Pilot gate
Freeze 30 unseen tasks across at least five engineering domains. Use the pilot only to estimate baseline, variance, verifier reliability and execution cost. Power analysis and preregistration happen before the confirmatory study.
Auditable episode
A patch is one artifact in a longer evidence chain.
Trajectory
Tools, edits, failures, retries, intervention points and skills exposed to the agent.
Outcome
Behavioral tests, public artifacts, release receipts, online smoke checks or rollback evidence.
Lifecycle
Skill origin, reuse, conflict, expiry, negative transfer and later corrective incidents.
Publication boundary
Private traces stay private by default.
- Raw trajectories remain in a local encrypted zone and never enter the public site or paper repository.
- Credentials, cookies, hosts, paths, identities and free text are removed before research sampling.
- Employer, customer, minor and consumer-media data require explicit written authorization or are excluded.
- Only licensed tasks, structured trajectories, verifier outcomes and necessary minimal diffs may become public artifacts.