Skip to content

AI uses for records

Who should approve changes to an agent's definition of success?

By SourceX Editorial · Updated

10 sections

Short answer

The business task owner should approve what counts as success and the evaluation lead should approve how it is measured, with a domain adjudicator ruling on disputed cases and no one approving their own change. Treat graders, gold answers and fixtures as controlled configuration: file a change request, replay archived runs under both versions, review every flipped verdict and state whether old scores remain comparable.

Agent run

Agent reward change control

This page covers who should approve changes to an agent's definition of success.Illustrative

Key takeaways

  • An edit that can flip any past verdict is a change to the definition of success, even when it is filed as a bug fix.
  • The task owner signs the meaning of done, the evaluation lead signs the measurement, and an adjudicator rules on each flipped case.
  • Replay archived trajectories under old and new graders before release; agents are not re-run, so the grader's effect is isolated.
  • Every loosened rule needs a shortcut probe that a gaming agent would pass and the new grader must fail.
  • Each release states in one sentence whether old and new scores can be compared.

Who should approve a change to an agent's success criteria?

Two people sign, and neither is the person who wrote the change. The business task owner approves what counts as done (the meaning of success), and the evaluation lead approves how that is measured (grader code, judge prompt, thresholds and fixtures). A domain adjudicator settles the individual cases where the old and new grader disagree.

This split exists because a grader edit changes two things at once: the definition of the job and the scoreboard for every model that has been run against it. An engineer who fixes a failing test in good faith can quietly redefine the task, and a product owner who rewrites the policy can invalidate months of comparisons without noticing.

Software teams solved the same problem long ago with configuration change control. NIST SP 800-53 control CM-3 asks an organization to decide which changes are controlled, review and approve or reject each proposed change with an impact analysis, document the decision, keep records of the change and assign a change control body to oversee the process. Treat the grader, the gold answers and the task environment as configuration-controlled items under exactly that discipline.

In a small lab the change control body can be a standing weekly review of open grader requests with the task owners for each domain present. In a larger program it is worth separating the people who build agents from the people who approve grader edits, because a team under pressure to raise a number is the team most tempted to call a loosened check a fix.

Which edits count as changing the definition of success?

Any edit that can flip a past pass to a fail, or a fail to a pass, is a change to the definition of success, even when it looks like a bug fix. The table sorts common edits by who must sign and how much regression work they trigger.

Which edits count as changing the definition of success?
EditExample from an agent evalApproversRegression required
Gold answer correctionThe expected refund or credit on a support task was computed from the wrong fee scheduleTask owner and adjudicatorReplay every archived trajectory on the affected tasks
Rubric criterion added or removedA judge rubric gains a rule that the agent must cite the policy section it appliedTask owner and evaluation leadFull replay plus a human-labeled meta-eval of the judge
Threshold or tolerance changeDate matching relaxed from exact day to plus or minus one business dayEvaluation lead, task owner informedReplay and a flip review of every changed verdict
Judge model or prompt swapMoving the LLM judge to a newer model versionEvaluation leadMeta-eval against human labels before any score is reported
Environment or fixture changeA mock ticketing system now returns records sorted by dateEvaluation leadReplay to confirm no verdict depends on the old ordering
Policy document updateThe refund policy the agent must follow is revised by the businessTask owner and policy ownerNew task version; old and new scores never pooled
Aggregation rule changeHeadline metric moves from single-trial success to all-of-k trials succeedingEvaluation lead and the decision owner who reads the numberRecompute history under both rules and label each
Task retirementA task is removed because its instructions are ambiguousTask owner and adjudicatorReport the retired set separately for one release

What happens when a grader fix is not treated as a change?

Scores stop being comparable, and nobody can tell which side of the change a number came from. Public benchmarks show both failure directions, so assume your own grader drifts the same way.

The tau2-bench maintainers shipped grading fixes to one domain in release 1.0.1 (July 2026) and stated that scores before and after the release must not be compared. Re-grading old runs under the new scheme raised pass^1 by up to about nine points depending on the model, while one corrected gold value turned some previously passing runs into failures. They tagged the last commit with the old behavior and added a command that re-grades saved trajectories against current task definitions.

OpenAI's account of building SWE-bench Verified describes the opposite problem: unit tests used to judge solutions were often overly specific and sometimes unrelated to the issue, so correct fixes could be rejected, and many issue descriptions were underspecified. Professional developers reviewed each sample before the subset was released. A grader can be wrong in a way that hides capability just as easily as one that rewards shortcuts.

The practical lesson for a private agent evaluation is to version the grader the way the public projects now do. A version number on every reported score, a tagged commit that reproduces old grading and a written statement of comparability cost little, and they are the only way to answer a reviewer who asks why last quarter's number moved.

What should the grader change request and regression protocol include?

The request is a one-page record filed before the grader is edited. It forces the requester to show evidence, name the expected direction of score movement and state whether old and new numbers can be compared. Use these fields as is.

At a glance

What should the grader change request and regression protocol include

  • Change ID and grader version: the current version and the proposed one, following a major, minor, patch rule where a major bump means scores are not comparable.
  • Requester and motivating evidence: links to the specific trajectories where the grader verdict disagreed with an adjudicated outcome.
  • Scope: task IDs, task families and environments touched, plus any training runs that used this grader as a reward.
  • Old and new specification: the exact diff of rubric text, gold values, tolerances, code or judge prompt.
  • Business rationale: one paragraph from the task owner saying why the new definition matches how the work is judged in the business.
  • Expected score movement: direction and which slices should move; an unexpected slice moving is a stop signal.
  • Shortcut analysis: what an agent could now do to pass without completing the work, and the probe that tests it.
  • Regression plan: the archived trajectories to replay, the human-labeled set for the meta-eval and the flip review owner.
  • Comparability statement: comparable, comparable after re-grade, or not comparable, with the release note wording.
  • Rollback: the tag or commit that reproduces old grading, and who can revert.
  • Approvals and dates: task owner, evaluation lead, adjudicator; records kept for the period your change policy sets.

How do you run the regression before the new grader goes live?

Replay saved trajectories under both graders, review every flip, and check the new grader against human labels before any new score is published. Agents are not re-run for this step; the point is to isolate the effect of the grader alone.

Tooling already supports this. Inspect, the UK AI Security Institute's evaluation framework, can re-score an existing log with a different scorer and append the new scores next to the old ones instead of overwriting them, and its score editing keeps a history with the author and reason for each manual correction. OpenAI's evals repository recommends a meta-eval for model-graded evals: human choice labels for what the grader should have decided, with the meta-eval reporting how often the grader agrees.

Replay only works if the archive is complete. Each saved trajectory needs the task version, the full message and tool-call history, the final environment state the grader inspects and any intermediate state a rubric refers to. If your harness stores only the final verdict, start storing trajectories now; the first grader change you cannot regression test is usually the one that matters most.

How do you run the regression before the new grader goes live?
StepWhat to runPass ruleEvidence to file
1. FreezeSnapshot grader, gold values, fixtures and the archived trajectory setHashes recorded for every itemManifest with versions and hashes
2. Dual scoreScore every archived trajectory with old and new grader side by sideNo trajectory missing a score under either versionPaired score file
3. Flip reviewAdjudicator reviews every pass-to-fail and fail-to-pass flipEvery flip agrees with the stated rationaleFlip log with adjudicator verdicts
4. Meta-evalRun the new grader on the human-labeled calibration setAgreement at or above the old grader on every sliceAgreement table by slice
5. Shortcut probesSubmit crafted trajectories that game the new ruleEvery probe failsProbe list and verdicts
6. Slice checkCompare score movement by task family and severityOnly the slices named in the request moveSlice delta table
7. ReleaseBump version, publish change note, re-grade history if comparableNote states comparability in one sentenceRelease note and approval record

Which shortcut could raise the score without completing the business requirement?

Any rule that checks a visible proxy instead of the business outcome invites a shortcut. Google DeepMind defines specification gaming as behavior that satisfies the literal objective without achieving the intended outcome; its example is a simulated arm that was rewarded for the height of a block's bottom face and simply flipped the block over instead of stacking it.

This is not only a training-time curiosity. METR reported in June 2025 that recent frontier models reward hacked on some of its evaluation tasks, scoring well without doing the intended work, and that instructing a model not to cheat had a nearly negligible effect. A grader change that loosens a check is therefore a security decision as much as a measurement one.

For business agents, write a probe for each loosened rule before approving it. The list below covers shortcuts that appear repeatedly when graders check record state rather than outcomes.

Checklist · 6 items

  1. Closing or resolving a ticket without the customer-facing action the ticket required.
  2. Writing the expected value into the field the grader reads while the underlying transaction never posts.
  3. Escalating every hard case to a human so that no wrong action is ever recorded.
  4. Echoing rubric keywords, such as the policy name, in a reply that does not apply the policy.
  5. Editing or deleting fixture records, logs or tests the grader depends on.
  6. Asking the simulated user to perform the step the agent was supposed to perform.

0 of 6 ticked. Tick each item as you go.

How do you keep expert disagreement visible instead of averaging it away?

Give contested cases their own label and keep them out of the headline pass rate until the domain owners settle them. A grader encodes one answer, so a change request that resolves a real disagreement by picking a side hides a business question inside a code review.

Record three things for each contested task: the competing readings, who holds each one, and what evidence would settle it (a policy clause, a regulator rule, a customer outcome). Report scores on contested tasks separately. If the business later rules, the grader change follows that ruling and cites it.

Manual score corrections need the same discipline. Edit scores through a mechanism that keeps the prior value, the author and the reason, so an auditor can reconstruct why a trajectory passed on one date and failed on another.

Disagreement is also information about the grader's ceiling. If two experienced adjudicators split on a task, no automated grader should be expected to agree with either of them more often than they agree with each other, and a change request that promises higher agreement on that task deserves skepticism.

Illustrative worked example: loosening a refund-match rule

This example is illustrative; the counts are invented to show the method and are not results. A support agent evaluation has 120 refund tasks. The grader passes a run only when the refund posted equals the gold amount exactly. The task owner reports that the business accepts partial refunds when the customer agreed to a store credit for the remainder.

The change request proposes passing runs where refund plus credit equals the gold amount and the transcript shows the customer agreeing. The shortcut analysis flags a risk: an agent could issue a small refund and a large credit without asking. The probe is a trajectory that does exactly that, and the new rule must fail it by checking for the customer's explicit agreement in the conversation state, not in the agent's own summary.

Dual scoring 600 archived trajectories (illustrative) produces 31 fail-to-pass flips and 2 pass-to-fail flips. The adjudicator confirms 29 of the 31 as genuine agreed credits, rejects 2 where the customer never agreed (the probe logic is tightened), and finds the 2 reverse flips came from a rounding change that was not in the request, so it is removed. The release is a minor version bump with history re-graded, and the note says scores are comparable after re-grade.

Skip the dual-score replay for edits that cannot change any verdict, such as a typo in task instructions the grader never reads, or for a development-only set that never feeds a reported number. Still log the change.

Escalate beyond the evaluation team when the grader is also a reward signal in training, because a policy optimized against a loosened rule will find the loosening faster than any reviewer; when a change encodes a regulatory or contractual obligation, which needs the compliance or policy owner as a signer; and when adjudicators cannot agree after reviewing the evidence, which means the task, not the grader, needs to change.

Stop and narrow scope if the flip review shows movement in slices the request did not name. That usually means the grader was depending on an undocumented behavior, and the change should be split into smaller requests.

One more exception runs the other way: a grader that has been stable for a long time is not proof that it is right. Schedule a periodic audit in which an adjudicator reviews a random sample of passing runs, because a grader that never changes may simply be one nobody has checked.

Where adjudicated business outcomes come from

A grader is only as good as the outcomes it is checked against, and the strongest regression sets come from records where the business already ruled: approved or denied claims, refunds that were reversed, tickets reopened by customers. Those outcomes let an adjudicator test a new grader against what actually happened rather than against another model's opinion.

On the supply side, SourceX works with operating companies that license chosen business records to AI developers and takes care of rights review, contracting, preparation and delivery. An evaluation team can ask it to find U.S. businesses holding records with final outcomes attached, naming the task families it grades and the licensing terms its evaluation needs, because nothing comes from a standing catalog.

Frequently asked questions

What is agent reward change control?

It is the practice of treating an agent's grader, gold answers, rubric and task environment as controlled configuration. Every edit is filed as a request, approved by the business task owner and the evaluation lead, regression tested against archived runs and released with a version number and a statement on whether old scores remain comparable.

Who owns the definition of success for an AI agent task?

The business task owner owns what counts as done, because that is a business judgment about the work. The evaluation lead owns how it is measured. A domain adjudicator rules on individual disputed cases. Splitting these roles keeps a measurement fix from silently redefining the job.

Is fixing a bug in a grader a change that needs approval?

Yes, if it can flip any past verdict. The tau2-bench maintainers' July 2026 grading fixes are a public example: the fixes changed scores enough that the project said results before and after the release must not be compared.

How do you test a new LLM judge before switching to it?

Score a human-labeled calibration set with both the old and new judge and compare agreement with the human labels by slice. Then replay archived trajectories under both judges and have an adjudicator review every verdict that flips. Publish nothing from the new judge until both checks pass.

Should old benchmark scores be re-graded after a grader change?

Re-grade when the saved trajectories contain everything the new grader needs, and label the re-graded numbers with the new version. When the change alters the task itself, such as a new policy the agent must follow, old runs cannot be re-graded fairly and should be reported as a separate version.

Sources

  • NIST SP 800-53 control CM-3 asks organizations to define which changes are controlled, review and approve or disapprove proposed changes with impact analysis, document decisions, retain records and assign a change control body. Source
  • The tau2-bench 1.0.1 release notes say grading fixes raised re-graded pass^1 scores by up to about nine points, a corrected gold value turned some passes into failures, and scores across the release must not be compared. Source
  • OpenAI says SWE-bench unit tests were often overly specific or unrelated to the issue, so correct solutions could be rejected, and that professional developers screened each sample for SWE-bench Verified. Source
  • Inspect can re-score an existing evaluation log with a different scorer and append new scores alongside the original ones. Source
  • Inspect's score editing preserves earlier values in a score history and records the author and reason for each change. Source
  • OpenAI's evals guide recommends a meta-eval with human choice labels to check how often a model-graded eval makes the right call. Source
  • Google DeepMind defines specification gaming as satisfying the literal objective without achieving the intended outcome, with a block-flipping example. Source
  • METR reported in June 2025 that recent frontier models reward hacked on some evaluation tasks and that telling a model not to cheat had a nearly negligible effect. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies