Skip to content

Evidence & Truth

Pioneer keeps different kinds of truth separate. A model name in configuration, a provider response, a completed run, and a human approval are useful facts—but they are not interchangeable.

Evidence classes

Class What it means What it does not prove
CONFIGURED A specification or durable setting says what Pioneer should use or do. That the resource exists, is reachable, ran, or succeeded.
OBSERVED Pioneer directly inspected bounded current or historical state. That the observation is still current, or that an action was authorized or executed.
EXECUTED A recorded operation or run reached a reported execution state. Success, correctness, persistence of a proposed change, or causal explanation.
PERSISTED A fact is linked to a durable store or revision. That it is currently loaded, reachable, executed, or correct.
HUMAN_AUTHORITY A person approved a specific consequential action at a defined boundary. That the action completed, succeeded, remained applicable, or authorized a different action.
MODEL_INFERENCE A model formed an interpretation from identified evidence. Direct observation, proof, or authority to act.

Evidence does not automatically widen

Never infer one evidence class from another. A configured model is not an observed installed model. An approved run is not a successful run. A successful execution is not proof that a proposed repair caused the outcome.

A model example

Suppose Settings names qwen3-coder:30b as the default Ollama model:

  • The saved name is CONFIGURED.
  • An Ollama inventory response that includes it is OBSERVED.
  • A durable profile record is PERSISTED.
  • A completed Chat generation is EXECUTED.
  • “This model is the best choice for the task” is MODEL_INFERENCE.

None of those facts silently supplies HUMAN_AUTHORITY for installation, deletion, or another consequential operation.

A Flow example

For a saved Flow and one run:

  1. The saved block specifications are CONFIGURED and PERSISTED.
  2. A bounded source preview is OBSERVED.
  3. Run history and per-block snapshots are EXECUTED evidence.
  4. A recorded approval decision is HUMAN_AUTHORITY for that exact action.
  5. A diagnosis such as “the delimiter caused the parse failure” is MODEL_INFERENCE and should cite the evidence that supports it.

Pioneer can compare an immutable run snapshot with the current saved revision. If the Flow changed after the run, current configuration should not be presented as though it were the exact historical configuration.

Evidence age and scope

Evidence remains bounded by its source, time, and identity:

  • A Machine sample can become stale after a failed refresh.
  • An installed model is not necessarily resident.
  • A source preview is not a full run.
  • A run result belongs to its explicit run ID, not to whichever run is newest.
  • A saved Flow revision and fingerprint identify state; a display name does not.
  • A Memory Browse result proves a record exists, while Recall proves only what a particular retrieval query returned.

How to communicate conclusions

Use the strongest accurate verb:

  • “configured” for saved intent;
  • “observed” for direct inspection;
  • “executed” or “recorded” for run evidence;
  • “persisted” for durable state;
  • “approved” for a scoped human decision; and
  • “likely” or “inferred” for a model conclusion.

When evidence is missing or stale, say so. Unavailable is more useful than a fabricated zero, default, or success.

Apply these boundaries in Diagnosing a Run and the complete Diagnose, Repair, and Verify workflow. The compact mapping is also available in Evidence Classes Reference.