Feedback Can Guide Improvement Without Rewriting the Verdict

Editorial clarification: September 8, 2026

The risks listed below are potential failure modes, not a claim that each was observed. The proposed experiment is not an approved or completed run. This update clarifies the August 5 record; it adds no implementation evidence.

Human feedback is useful because people can identify workflow problems and tradeoffs that automated checks cannot infer. It becomes dangerous when a favorable opinion can quietly turn a failed run into a passing one.

The August 5 development record describes WTK separating those two roles. The next question is how long the human must remain inside each improvement cycle.

Current state

A completed run can produce deterministic checks, goal-based evaluation, semantic review, qualification evidence, and human judgment. These records answer different questions and should not overwrite one another.

Before this cycle, the architecture described that separation more clearly than the feedback workflow enforced it.

Human review currently supplies judgment where the system cannot derive a defensible answer. That is a safety boundary, but it should not become an assumption that every candidate revision requires permanent manual supervision.

Changes

WTK can now ask targeted questions after reviewing a specific session. Each question identifies the package, session, review, and goal obligation that caused the question to be asked.

The resulting answer is preserved as an attributable feedback record. It can identify a workflow preference, clarify an operator-owned tradeoff, record acceptance context, or prioritize an improvement.

It cannot qualify the package. It cannot erase a grounding, safety, deterministic, or qualification failure. It cannot promote a candidate or silently change the definition of success.

Feedback remains advisory input to the next improvement decision.

This creates a clearer path toward greater autonomy. A future policy may permit routine improvement without case-by-case human feedback, but only within limits on actions and changes that people approved beforehand.

Potential failure modes

Several tempting shortcuts would weaken that boundary:

• asking for broad approval without naming the run or obligation under review;
• treating a chat response as evidence for a different package revision;
• allowing human preference to replace an objective failed check;
• allowing an agent to redefine success so its own candidate appears better;
• carrying favorable feedback forward after the underlying package changes;
• removing the human before rollback, escalation, and evidence requirements are reliable.

Assumptions removed

We no longer assume that all human input has the same authority.

We no longer assume that an operator answer should change the recorded verdict.

We no longer assume that free-form feedback remains meaningful after its package or session changes.

We also do not assume that a human must approve every bounded improvement forever.

Evidence boundary

WTK can represent the separation between advisory feedback, deterministic evidence, qualification, and promotion authority. It can bind feedback to the conditions that produced it and prevent that feedback from directly overriding failed gates.

That does not establish when human review can be safely removed. WTK has not yet demonstrated that automated evaluators remain valid across repeated improvement cycles, that candidate and incumbent comparisons resist coordinated drift, or that rollback and escalation remain dependable without active supervision.

Human authority remains necessary when an improvement changes the goal, success criteria, governance policy, permitted authority, risk tolerance, or definition of acceptable evidence.

Research question advanced

This cycle advances the question of whether agents, teams, and the factory can improve without gaining authority to redefine success. Feedback can influence a candidate while the goal, failed verdicts, qualification boundaries, and promotion authority remain outside that candidate.

A proposed next step would test whether previously approved policy can replace case-by-case human review. Safe autonomous promotion across repeated improvement cycles has not been demonstrated.

Next hypothesis

Case-by-case human review may become unnecessary for a routine improvement when all of the following remain true:

• the goal and success criteria are unchanged;
• the proposed change stays inside a preauthorized scope;
• every attempt and failure remains in the evidence record;
• deterministic safety and grounding floors pass;
• the candidate is compared against the qualified incumbent;
• evaluation shows improvement without regression under declared conditions;
• cost and consequence remain below an approved risk threshold;
• rollback has been tested;
• uncertainty, disagreement, or missing evidence automatically returns the decision to a human.

The proposed experiment would compare the same improvement policy in three modes: human advisory control, autonomous decisions recorded without applying them (shadow mode), and bounded autonomous promotion. It would measure decisions, regressions, abstentions, escalations, and rollback behavior. The protocol and any execution require separate human approval; this historical proposal does not authorize either.

The goal is not to remove human intent from the factory. It is to learn which recurring decisions can be safely delegated without requiring a person to stand beside every iteration.

The site's public research feedback process is related but separate: reader contributions can inform research questions, while the feedback in this log is tied to a particular agent run. Neither replaces the evaluation and approval decisions described in the evidence lifecycle.