Restore acceptable behavior after failure.
A repair loop attributes a failed goal, gate, or runtime outcome to the responsible layer, changes that layer, and re-evaluates the candidate.
Governed Self-Improving AI Agents
Self-improvement may be automated. Self-authorization should not be.A self-improving agent system should be able to learn from failures and outcomes without silently replacing its goals, weakening its tests, selecting only favorable evidence, or approving its own successor. WTK treats improvement as a governed lifecycle with separate candidate, evaluation, qualification, and promotion authority.
What is a governed self-improving AI agent?
A governed self-improving agent can use retained evidence to propose a repair or refinement, but the proposed change remains a candidate until it is independently evaluated, qualified for an exact execution form, and promoted by an authority outside the component being improved.
Self-improvement describes the change loop. It does not grant authority to redefine the goal, weaken the evidence floor, discard failed attempts, or approve a replacement.
Improvement is a governed lifecycle, not a silent rewrite.
- 01
Observe
Retain outcomes, failures, drift, and user evidence without selecting only favorable runs.
- 02
Attribute
Name the responsible layer before choosing repair, refinement, or escalation.
- 03
Change
Create a versioned candidate while the qualified incumbent remains recoverable.
- 04
Evaluate
Run independent checks and judgment without borrowing the candidate's development test.
- 05
Qualify
State what the exact candidate and execution form demonstrated, including limitations.
- 06
Promote
A separate authority decides whether the candidate may replace the incumbent.
- 07
Monitor
Invalidate, suspend, revoke, remediate, or roll back when evidence or conditions change.
Repair and refinement solve different problems.
Improve one objective while preserving required performance and controls.
A refinement loop begins from a qualified baseline, freezes the comparison, preserves original obligations, and rejects candidates that fail any non-regression gate.
Qualification does not equal promotion.
Human-authorized promotion
A human reviews the retained evidence and decides whether a qualified candidate may replace the incumbent or enter a release channel.
Policy-authorized promotion
A separate promotion authority may eventually approve bounded changes under human-defined policy when evidence, risk limits, escalation rules, and rollback conditions are satisfied.
Return authority to a human
Uncertainty, increased consequence, drift, policy change, missing evidence, or failed checks stop unattended promotion and require human judgment.
What WTK can responsibly claim.
MechanismImplementedEvidenceBounded
Candidate and incumbent remain separate.
WTK represents candidate comparison, evidence invalidation, qualification, rollback, and external promotion authority.
MechanismDirectionEvidenceUnproven
Long-duration self-improvement remains open.
WTK has not established safe repeated improvement under evaluator, model, tool, policy, and adversarial drift.
Follow the claims to bounded evidence.
These records preserve the tested question, observed boundary, and work still owed. Their titles and evidence classes remain unchanged.
A Repair Cannot Borrow Its Test
A repair can respond to a failed review without copying the withheld case that exposed it.
Factory LogsA Score Gain Needs a Frozen Twin
Repeated tests of an unchanged system help distinguish measured improvement from score variation.
Research NotesThe Repair Loop Needs an Exit Ramp
When a failure survives an instruction change, the system should change repair layers instead of polishing the same guess again.
Factory LogsThe Evidence Stayed. Its Authority Did Not.
A package edit preserved the old record for inspection while preventing stale construction and approval evidence from authorizing the revised package.
FindingsFive Failures Needed Five Different Repair Decisions
A fixed comparison found that typed escalation routed every persistent failure correctly while instruction-only repair repeated the same move.
FindingsA Portable Package Does Not Qualify Its Runtime
Portability is not permission to assume that every harness enforces the same safety boundary.
Research NotesA Green Check Can Go Stale
A tool connection is not ready merely because it was once configured or once passed a check.
Factory LogsFeedback Can Guide Improvement Without Rewriting the Verdict
The August 5 development record separates session-bound feedback from evaluation results; safe autonomous promotion remains unproven.
Factory Logs