Skip to content

Article / Trajectory evidence

Final outcome vs agent trajectory: why an evaluation needs both

A final outcome tells you where a run ended. A trajectory tells you how observations, attempted actions, tool results, and authoritative transitions produced that ending. Either view alone can misclassify agent behavior.

For
Evaluators deciding whether to review outcomes, traces, or both
Outcome
Interpret run results without confusing a favorable ending with a sound process

Answer first

Outcomes and trajectories answer different evaluation questions

Outcome analysis asks whether the run reached a desired terminal state. Trajectory analysis asks whether the agent used available information and tools in a defensible sequence.

In a stochastic environment, a sound decision can end badly and a weak decision can end well. An ending point total therefore cannot, by itself, establish whether the preceding choices followed the declared policy.

The reverse problem also matters. A tidy-looking trace is not success when the agent abandons the task, violates a hard constraint, or reaches the wrong terminal state. The evaluator needs the terminal result and the ordered path that led there.

Interpretation

Use a two-axis view instead of one pass-or-fail label

Outcome and trajectory combinations
OutcomeTrajectoryWhat the combination suggests
FavorableSoundA successful run with supporting process evidence
FavorableWeakA result that may depend on luck or an unsafe shortcut
UnfavorableSoundA defensible process facing stochastic or difficult conditions
UnfavorableWeakA failure with a traceable behavioral cause

Evidence

A useful trajectory connects decisions to authoritative effects

The minimum useful sequence is the observation available before a decision, the attempted action, the tool response, and the resulting authoritative state or rejection.

  • Record rejected attempts rather than displaying only accepted transitions.
  • Preserve sequence order so a later read cannot be mistaken for information available earlier.
  • Separate the agent's submitted payload from the server's resulting state.
  • Keep evaluator labels outside the objective run record.

Reporting

Report outcome and process measures separately before combining them

Publish the ending result, completion state, and any trajectory measures as distinct fields so a reader can see which evidence supports each conclusion.

If a study later combines those fields into a score, disclose the weighting and failure policy. Otherwise a strong terminal result can silently erase repeated invalid attempts, while a strict process score can hide that the task was actually completed.

WagerCall records observations, attempts, transitions, events, and synthetic-point movements. The evaluator decides which of those facts matter for the stated question.

Next step

Inspect completed runs

Compare terminal states with the ordered public evidence that produced them.

Inspect completed runs