Regular guideline pipeline lacks a minimal-trajectory-length guard before segmentation

[jayaramkr]

Summary

The regular guideline pipeline passes trajectories to segment_trajectory with no minimum-length guard. A single-step (or otherwise too-short) trajectory can be "segmented" into ≥2 subtasks that all slice into the same one step, producing multiple near-identical per-segment guidelines from a single step.

The consistency pipeline already guards against this (PR #289 added n_scorable_steps >= 2 and _can_segment_trajectory(...) before segmenting), so the two paths are now inconsistent.

Where

generate_guidelines() in altk_evolve/llm/guidelines/guidelines.py (~L191):

subtasks = []
if evolve_config.segmentation_enabled:
    from altk_evolve.llm.guidelines.segmentation import segment_trajectory
    try:
        subtasks = segment_trajectory(messages)   # <-- no n_steps guard
    except Exception as e:
        ...

There is no check on n_steps / len(steps_list) before calling segment_trajectory.

Why it fans out

For a single-step trajectory (n_steps == 1), if the segmentation LLM returns two subtasks both spanning [1, 1]:

• start = min(max(0, 1 - 1), 1) = 0, end = min(max(0, 1), 1) = 1 → start < end → the slice passes the range filter.
• Both subtasks produce steps_list[0:1] — the same single step.
• len(valid_slices) >= 2 is satisfied, so the pipeline emits one guideline-generation call per identical slice.

Net effect (reported by @evduester): a single-step smolagent trajectory in both mode yields ~13 guidelines — ~10 from the regular pipeline (multiple identical segments) + 3 consistency. The regular-pipeline inflation is the segmentation-without-a-length-guard behavior described here.

Suggested fix

Mirror the consistency path's guard in the regular path — skip segmentation when the trajectory is too short to segment meaningfully, e.g.:

if evolve_config.segmentation_enabled and n_steps >= 2:
    subtasks = segment_trajectory(messages)

Optionally, also short-circuit inside segment_trajectory itself so every caller benefits, and/or de-duplicate identical [start, end] slices before generating (defense in depth against an LLM returning overlapping/duplicate ranges even on longer trajectories).

Scope note

Surfaced during review of #289, but intentionally out of scope for that PR — #289 fixes the consistency path; this is the parallel gap in the regular path.