Agentic reinforcement fine-tuning uses graded multi-step behavior to change an agent policy.
Versalist prepares evaluation evidence. It does not provide training compute or operate the customer agent runtime.
Decision
Is the available evidence complete enough to support a bounded training experiment?
Do not convert incomplete or weakly attributed Episodes into training data.
Policy
Version the input
Identify the skill or policy used by each Episode.
Task
Keep it stable
Bind each Episode to one Challenge version.
Reward
Validate the signal
Test whether graders reward the intended behavior.
Change
Limit the claim
Measure one policy revision on a holdout set.
1. Control the comparison
Use these controls before you collect or compare results.
Contract
Establish causal continuity
A training record needs a stable link from policy version to task, behavior, score, and decision.
Use permanent Episode identities
Record attempt generations
Reject stale results
Evidence
Audit the reward signal
A useful reward must match the intended behavior and resist shortcuts.
Test negative examples
Check grader disagreement
Review reward changes
Review
Keep privacy boundaries
Training eligibility does not permit collection of private payloads or hidden reasoning.
Use rights-cleared records
Store bounded metadata
Exclude unknown-provenance data
Decision
Measure the policy change
A training result is useful only when a fixed evaluation can detect generalization and regression.
Keep a holdout set
Compare against the baseline
Record rejected changes
2. Use the procedure
Complete each step in order. Stop when a required input or control is missing.
1
Define the policy revision
Identify the exact skill or instruction version.
2
Select eligible Episodes
Require complete provenance and approved data rights.
3
Validate the reward
Test the grader against reviewed positive and negative cases.
4
Freeze the evaluation
Lock the Challenge, cases, rubric, and model conditions.
5
Run a bounded experiment
Change one policy variable within an approved budget.
6
Evaluate the candidate
Use holdout results to accept or reject the revision.
3. Keep the evidence
Store enough evidence for another reviewer to repeat the decision.
| Record | Required evidence | Failure signal |
|---|---|---|
| Policy source | Skill version and content digest | The training input is unknown |
| Episode source | Challenge, steps, attempt identity | Results mix stale attempts |
| Reward | Grader version and validation cases | The signal rewards a shortcut |
| Rights | Eligibility and provenance record | Private or external data enters the corpus |
| Decision | Baseline, candidate, and holdout result | The revision has no comparison |
4. Keep the product boundary
- Current Trace events contain bounded metadata, hashes, latency, token counts, and sanitized errors.
- Current Trace events do not contain raw prompts, full transcripts, hidden reasoning, or tool payloads.
- An Episode stores rubric scores. It does not store one Episode reward column.
- Versalist does not execute the training job or operate the customer agent runtime.
5. Apply this on Versalist
Review the agent training stack
See how Challenge, skill, Episode, and evaluation records support an improvement loop.
Read the training stack documentationInspect result boundaries
Use the current result contract before you classify an Episode as eligible evidence.
Read result documentationRelease gate
Make the decision explicit
Accept a revision only when the holdout improves and no protected failure group regresses.