[Autoplectic]

Stacked on #10 (uses compare_networks(curtail=..., map_fn=...)).

Scripts in examples/catfr/ and a report in reports/catfr_pilot.md, running infoflow on the UPenn categorized free recall iEEG dataset (OpenNeuro ds004809).

Findings

• Data alignment. The dataset's BIDS onset column is offset from the recording by 1 s to 15 min per session. Epochs must use sample / sfreq: only those show the word-locked evoked response. common.word_events does this and documents why.
• No subsequent-memory effect in directed flow.
  • Pilot: 3 subjects × 18 channels; one significant edge in six contrasts.
  • Cohort: 25 subjects × 3 region nodes, paired between-subject test; nothing significant (18 tests, smallest q = 0.48).
  • The classical HFA amplitude effect is present in the same cohort data (PFC, p = 0.001).
• Calibration.
  • Selecting edges per condition makes the comparison anti-conservative; pooled selection restores nominal rates for intrinsic and shared flow.
  • Synergistic per-edge p-values are mildly low between 0.01 and 0.05 (100-split check). The evidence is modest.

Scripts

• fetch.py: rank sessions and download them.
• prepare.py: epochs, broadband and HFA.
• run.py: within-subject contrast and null splits.
• null_check.py: per-edge null p-values.
• cohort.py: streaming download, region nodes, between-subject test.
• report.py: tables and figures.

Made with Cursor