Ising analysis of Purkinje-cell populations: experiment tracker

This repo fits pairwise maximum-entropy (Ising) models to Neuropixels recordings of mouse cerebellar Purkinje cells during reaching. The data is the Calame, Becker & Person dataset: 21 sessions, 6 mice, 3 blocks each (baseline, closed-loop stimulation, washout). The goal is one paper on what the Ising model reveals that other methods cannot.

• How to run or add an experiment: PROTOCOL.md. Start one with bash tools/new_experiment.sh.
• Paper-facing results (single source of truth): RESULTS_LOG_FOR_PAPER.md
• Full project handoff (data conventions, cluster gotchas): 09_claude_science_handoff/HANDOFF_CLAUDE_CODE.md
• Shared code: shared/. The original MCH pipeline is in old_code/energy_over_time_script/.

Experiment status

ID	Experiment	Status	Owner	Updated	Headline
E01	Stationary MCH Ising fits (July 2026, all reaches) · results	complete	claude-science	2026-10-12	Ising reproduces P(K) (wins 62/63 blocks); |J| beats 100 rate-preserving shuffles in 63/63. Superseded as primary fits by E08 (exclusions applied).
E02	Arbitration: energy vs firing rate at peak acceleration · results	superseded	claude-science	2026-10-12	Early 'energy wins 30/63' result not robust; firing rate is more movement-modulated (B6). See E07.
E03	J coupling vs kinematics timing · results	superseded	claude-science	2026-10-12	WITHDRAWN: 'J lags in ~70%' and 'J best indicator' fail against control windows (D table). See E07.
E04	Collective activity P(K) vs J peaks · results	superseded	claude-science	2026-10-12	J-peak/collective association (Fisher p=4e-8) reflects independent-model failure; superseded by time-varying-field analysis (E07/E09).
E05	Decision-tree prediction of J-peak blocks · results	archived	claude-science	2026-10-12	No write-up of results exists; target (J peak) depends on withdrawn peak definitions.
E06	Synthetic EI network: J changes from correlation alone · results	complete	claude-science	2026-10-12	Yes, J jumps during synthetic collective epochs at matched rate (method validation only; A6).
E07	Pre-submission checks and time-varying-field model · results	complete	claude-science	2026-10-12	Coupling beyond reach-locked rates in 21/21 sessions; no evidence coupling strengthens during reach; timing claims withdrawn. See RESULTS_LOG A5-A8, B, D.
E08	MCH refit with excluded reaches dropped (current primary fits) · results	complete	claude-science	2026-10-12	No material change: J vs July fits median r=0.996, slope 1.00; 63/63 blocks refit.
E09	Exact-ML thermodynamics, H0 validation, operating point H1 · results	analysis	claude-science	2026-10-12	H0 validation done (primary dI/N accurate). H1 goes AGAINST premise: p(1-p) falls in 17/21 sessions (p=0.018). Amendments 2-3 need synthetic check.
E10	Operating point: amendment checks and H2-H4 on real data · results	planned	grad-researcher	2026-10-12	Planned (design written, awaiting go-ahead). Step 1: synthetic check of amendments 2-3; step 2: H2-H4 on real data (cluster).
E12	Field-term dip vs firing rate as a marker of peak acceleration · results	complete	grad-researcher-3	2026-10-12	Field-term dip locks to peak acceleration (65% vs 14% rest control; 50 ms vs 330 ms for mean rate) but is a population-polarization effect: no detectable advantage of coupling-corrected fields (FT vs FT0, CIs wide) and a model-free polarization index matches it. POOL and rest-control check NOT RUN (sbatch blocked).

Status values: planned → running → analysis → complete. superseded means the experiment was replaced by a later one and its claims may be withdrawn. archived means it is kept for reference only. After changing an experiment.json, regenerate this table with python3 tools/update_status.py.

Experiment queue

Each experiment has its own branch and worktree (../ising_wt_eNN). Its design is in NN_*/README.md on that branch. Rows for experiments not yet merged appear in the status table above when their PR merges.

Order	Experiment	Branch	Owner	State
running	E10 Operating point H2–H4	exp/10_operating_point_h2_h4	Graduate researcher	full cluster run (job 4363115)
running	E11 Stimulation κ model	exp/11_stimulation_kappa	Graduate researcher 2	full cluster run (jobs 4363060–62)
next 1	E12 Field-term dip vs firing rate (+ decoder, Calame pooled-rate re-run)	exp/12_field_vs_rate_indicator	unassigned	designed; mostly local
next 2	E14 Bin-size sweep and 1-ms synchrony (Nashef reconciliation)	exp/14_binsize_1ms_synchrony	unassigned	designed; heavy at 1 ms
next 3	E13 Coupled-GLM benchmark	exp/13_glm_benchmark	unassigned	designed; light

When an agent finishes its experiment, it takes the next unassigned row: it sets owner in experiment.json, works in that worktree, and writes the pre-registration before touching real data.

Not yet scoped: rewriting the stimulation-model figures for the paper; a shared-input simulation; J vs probe distance (partly in E14). Questions for the lab:

• How were the 2022 sessions (DJC000/DJC002) recorded and sorted?
• Was the unit-stability filter applied?
• What do exclude and qm mean?
• Are all mice opsin-positive?
• Does Nashef et al. have a published journal version?

Layout

ising_EnergyOverTime/
├── README.md                  # this tracker
├── PROTOCOL.md                # how experiments are run and recorded
├── RESULTS_LOG_FOR_PAPER.md   # the single paper-facing results log
├── NN_<experiment>/           # one folder per experiment (see PROTOCOL.md)
├── _template/                 # copied by tools/new_experiment.sh
├── tools/                     # new_experiment.sh, log_run.sh, update_status.py
├── shared/                    # cross-experiment code (src/ package, aggregation scripts)
├── docs/                      # older project overview and session-report galleries
└── old_code/                  # everything from before the reorganization (original pipeline, notebooks, MATLAB)

Legacy folders (E01–E06) predate this protocol. Their notes/ hold the original analysis notes, and their per-session outputs live on the cluster under results/ (gitignored).