# Larkum public-data innovation pilot

## Purpose and status

This exploratory pilot asks whether subtracting the dendritic activity predicted
from local spine traffic exposes a direction-independent trial-outcome signal in
public neural recordings.  It is an independent biological consistency test for
the innovation operation, not a test of network training or an exact
somato-dendritic residual.  The data structure and class counts were inspected
before this protocol was fixed, so the result is not an untouched confirmation.

The data are from Maristany de las Casas et al., *Science* (2026), “Tuft
dendrites in frontal motor cortex enable flexible learning,” DOI
`10.1126/science.adx4358`.  The archived public data are released under CC BY
4.0 at `https://doi.gin.g-node.org/10.12751/g-node.etlk5k/`.

## Fixed dataset

Use the eleven `Figure2/Data/*_dff.mat` sessions in archive
`10.12751_g-node.etlk5k.zip`:

- animals `DCO1`, `DCO2`, and `DCO4`;
- saline (`Sal`) and chemogenetic NDNF activation (`DCZ`) conditions;
- trial-aligned `spine_local`, `branch`, `TrialTypes`, `Choice`, and `DirOut`;
- omit trials with a non-finite choice or malformed neural arrays.

No Figure 1, 3, 4, 5, or 6 endpoint enters this pilot.  Those modules do not
provide the same trial-level pairing needed here.

## Fixed innovation estimator

For every session, subtract the first 30-frame mean from every spine and branch
ROI on each trial.  The expected branch trace is a multi-output ridge regression
from all local-spine traces, frame identity, instruction identity, and their
interaction.  The ridge coefficient is fixed at `1.0`.

The predictor never receives choice, correctness, outcome, drug condition, or
future-session data.  Saline trials are five-fold cross-fitted in contiguous
trial blocks.  The DCZ predictor is fitted once on all saline trials from the
same session and then frozen.  The innovation is

```text
branch activity - predicted branch activity.
```

A simpler task-template residual subtracts the saline mean trace for the same
instruction without using spine activity.

## Fixed endpoint

The outcome window is 0 to 1 second after report onset, corresponding to the
public analysis time axis from -3 to 3 seconds over 180 frames.  Split this
window into six bins.  In each bin summarize the population by signed mean,
mean absolute activity, and RMS activity.  This produces the same 18 features
for raw branch activity, task-template residual, spine-conditioned innovation,
and local-spine activity.

The primary endpoint is cross-instruction, leave-one-animal-out decoding of
`DirOut`:

1. hold out one animal;
2. train a balanced logistic decoder on one instruction direction from the
   other two animals;
3. test it on the opposite instruction direction in the held-out animal;
4. repeat in the other direction and for all three held-out animals;
5. pool the out-of-fold predictions and report AUROC.

This split is load-bearing.  Correctness is determined by instruction and lick
direction, so ordinary random cross-validation can relabel sensory or movement
activity as an outcome signal.  Under the cross-instruction split, a pure
choice-direction signal reverses sign.  A choice-only decoder is retained as a
negative control.

Report separately for saline and DCZ:

- AUROC for raw branch, task-template residual, spine-conditioned innovation,
  local spine, and choice-only control;
- per-animal AUROC and the number of correct/error trials;
- the paired AUROC difference between innovation and raw branch;
- the saline-minus-DCZ change in innovation AUROC;
- 95% descriptive intervals from 5,000 session-block bootstrap samples,
  resampling sessions within animal.

The bootstrap describes stability across the released sessions.  With only
three animals, it is not treated as population-level animal inference.

## Decision rule

The pilot supports the narrow biological claim only if saline innovation has
AUROC above 0.5, exceeds raw branch activity, and the gain has the same sign in
all three held-out animals.  A weaker DCZ innovation endpoint is a causal
consistency result, not a required pass condition because DCZ has few error
trials.  Failure, sign inconsistency, or an advantage confined to ordinary
within-instruction decoding rejects this dataset as flagship evidence.

Regardless of outcome, this pilot cannot establish improved learning,
scalability, or an exact Harnett-style soma-dendrite residual.

## Frozen result

The single frozen run failed the primary decision rule.  It retained 730
saline trials (53 errors) and 738 DCZ trials (29 errors) from all eleven
sessions.  Cross-instruction, leave-one-animal-out AUROCs were:

| Condition | Choice only | Raw branch | Task-template residual | Spine-conditioned innovation | Local spine |
| --- | ---: | ---: | ---: | ---: | ---: |
| Saline | 0.000 | 0.406 | 0.518 | 0.450 | 0.340 |
| DCZ | 0.017 | 0.494 | 0.634 | 0.577 | 0.322 |

In saline, innovation improved over raw branch by 0.044 AUROC, but its
session-block 95% interval for the difference was `[-0.064, 0.104]`, the
endpoint remained below 0.5, and the animal-level improvement was not
directionally consistent.  The simpler task-template residual was stronger.
DCZ innovation improved over raw by 0.082, with interval `[-0.009, 0.200]`,
but DCZ innovation exceeded saline innovation by 0.127 rather than weakening.

The near-zero choice-only result verifies that the cross-instruction split
reversed the obvious movement-direction shortcut.  The observed pattern does
not support a unique spine-conditioned innovation signal in this release and
does not justify endpoint-window, ridge, or feature retuning.  The complete
record and trial predictions are stored in
`results/larkum_public_pilot.json` and
`results/larkum_public_pilot_predictions.csv`.