| name | prospective-kit |
| description | Use this after the report ships. Turn open questions into falsifiable predictions and instrument future sessions lightly. |
Prospective Kit
Why
Retrospective analysis establishes suspects, not verdicts. Confounders,
missing data, and selection bias cap what any transcript audit can prove.
The fix is cheap prospective measurement going forward.
Hypotheses to Test
Choose five to ten uncertain but potentially valuable claims about the
subject's work style. Convert each to a falsifiable form:
HYPOTHESIS: planning before debugging reduces correction loops.
PREDICTION: comparable debugging tasks started with an explicit hypothesis
tree will need fewer major reversals than ones started cold.
MEASURE: reversals per debugging episode, split by presence of a written
hypothesis tree, over the next N episodes.
Good hypotheses come straight from the audit's LOW-confidence findings and
its miscalibration list.
Light Instrumentation
For suitable higher-value tasks only - never interrupt routine actions -
capture:
BEFORE: expected difficulty, expected solution path, confidence,
expected completion route.
AFTER: actual result, surprise level, what changed the mental model,
one-line lesson.
Total burden: under two minutes per captured task. If it feels heavy, it
will be abandoned; cut fields before cutting compliance.
This data compounds: every month of BEFORE/AFTER pairs sharpens the next
calibration-audit run.
Research Grounding
Keep methodology and sources separate from personal reports. Distinguish
established peer-reviewed findings from emerging preprints in SOURCES.md.
When a new behavioral claim enters the report, name the literature it
agrees or disagrees with.
One-Line Memory
Retrospective finds suspects. Prospective runs the trial.
