Lab Notebook

Update The Summary Sparingly

• Update the summary layer only when one of these changes:
  • current best explanation
  • current best branch or baseline
  • settled negative results
  • next discriminative check
  • success criterion or decision rule
• Do not rewrite the summary after every routine action.
• The summary should be short enough that a fresh agent can re-enter from it.
• If multiple live explanations matter, keep a tiny hypothesis registry in the summary layer:
  • hypothesis
  • strongest support
  • strongest weakening evidence
  • next falsifier

Make Entries Useful

• Include exact evidence paths and identifiers.
• Include representative visual artifact paths when a plot, overlay, slice, curve, or diff view materially affected the decision.
• Record negative results explicitly when they rule out a family of ideas.
• If the entry draws an inference, state it explicitly instead of leaving it implicit in the next step.
• Keep inference confidence coarse: low, medium, or high.
• Do not add confidence to entries that are only reporting raw observations or routine actions.
• Preserve chronology so later readers can reconstruct what happened.
• End each entry with "so what?" in concrete form.
• If nothing changed in understanding, say that explicitly.

Avoid Notebook Drift

• Do not let the summary become stale relative to the chronology.
• Do not dump raw metrics or images without interpretation.
• Do not record an interpretation without enough evidence to re-check it later.
• Do not duplicate the full chronological detail into the summary.

References

• Read references/templates.md for one-file and two-layer notebook templates.