WTKRESEARCH + ENGINEERING
← All engineering journal records

Factory Logs

No Evidence Path, No Confident Answer

WTK now turns a missing required evidence path into a recorded limitation instead of accepting unsupported confident prose.

Telling an agent to use a tool is not the same as proving that the answer used evidence from that tool. When the required evidence never arrives, a capable model can still produce a polished and convincing response.

This factory cycle added a harder boundary around that failure.

Current state

Some goals require current external facts or proof that an action occurred. That requirement belongs to the goal, not to whether a tool happened to be successfully attached later.

If the requirement disappears when tool selection fails, the least grounded execution path receives the weakest protection.

Teams add another problem. A coordinator can produce the final answer even when the role responsible for gathering evidence failed, or it can rely on an unrelated receipt from another branch of the team.

Changes

WTK now carries the grounding requirement from the governed goal into execution.

When a fact or action claim requires evidence, the runtime expects attributable evidence from the contributing execution path. If an obligated run produces no usable evidence, WTK replaces the unsupported answer with an explicit limitation. Structured outputs receive a structured error when their contract provides an error channel.

The event is also preserved for evaluation and improvement rather than being treated as a successful answer.

For teams, evidence must follow the path that contributed to the final response. An unrelated sibling receipt should not make an unsupported answer appear grounded.

Failures observed

The mechanism responds to several recurring failure shapes:

  • a tool-less agent answering a current factual question confidently;
  • a coordinator presenting a conclusion after its evidence-producing role failed;
  • an agent claiming an external action completed without an execution receipt;
  • a prompt-level instruction being ignored while the runtime accepts the prose;
  • a receipt from an unrelated team branch being used to support the final answer.

Assumptions removed

We no longer assume that declaring a tool means it was used.

We no longer assume that a confident limitation marker makes contradictory prose safe.

We no longer assume that any receipt is automatically evidence for every claim in a team result.

Evidence boundary

The current mechanism supports a bounded construction claim: WTK can preserve a goal-derived evidence obligation, detect a zero-evidence execution, emit a structured limitation, and retain that outcome for later review.

It does not prove that every nonzero receipt is relevant, truthful, sufficiently fresh, or produced by the intended capability. Tool-description drift, source quality, cross-harness enforcement, and partial evidence remain active research questions.

Research question advanced

This cycle advances the question of when independent review adds value beyond receipt-grounded evaluation. It strengthens the evidence-supply side by preventing a zero-evidence execution from presenting unsupported prose as an assessable success.

It also informs goal continuity and delegation integrity by carrying evidence obligations along the contributing execution path. It does not measure reviewer-panel reliability or establish that every receipt is relevant and authentic.

Next hypothesis

The boundary should distinguish unprovisioned, unattempted, empty, failed, partial, and valid evidence without allowing one state to masquerade as another. The next test will vary those conditions across standalone agents and teams, then repeat them in additional harness projections.

RECORD DETAILSReference FL-004
Artifact
Factory Logs
Status
Published
Published
August 5, 2026
Linked sources
None declared