path: root/ORAL_B.md

# Oral-bar B: preregistered Harnett-signature programme

This programme is grounded in Francioni et al., *Vectorized instructive
signals in cortical dendrites* (Nature, 2026),
<https://www.nature.com/articles/s41586-026-10190-7>. It does not treat
"apical activity carries error" as sufficient novelty. The target is the
paper's specific somato-dendritic innovation signatures and its explicitly
unresolved distinction between an online control signal and a plasticity-only
teaching signal.

No oral-B result may be generated before this protocol is committed. Native
accept-bar baselines may continue in parallel, but the oral-B confirmation may
not be inspected until C4 is closed.

**Development status: failed on 2026-07-22.** The complete 36-run frozen grid
was generated from clean revision `b4458cd` and no confirmation seed was
touched. All 12 candidates failed the sign-inversion requirement in all three
task environments. No candidate is eligible, so the confirmation specified
below is prohibited. Detailed negative results are retained in `RESULTS.md`
and `results/bci_dev_selection.json`.

## What the biological experiment establishes

The paper motivates four experimental conditions for a vectorized dendritic
teaching signal:

1. dendritic activity contains information absent from parent-soma magnitude;
2. a population of SD residuals contains reward and trial-outcome information;
3. residual sign depends on the causal role of each neuron (opposite for P+
   and P- cells) and follows error change rather than error magnitude alone;
4. suppressing apical computation abolishes the vectorized signal and impairs
   learning.

The paper also reports that surrounding-network activity predicts whether a
coincident event is dendritically amplified or attenuated, and that residual
sign relates to subsequent activity changes. It explicitly leaves open whether
the residual first controls neural state online, directly gates plasticity, or
does both. That phase ambiguity is the model's central prospective prediction.

## Continuous synthetic BCI

The task is deliberately closer to the experiment than shuffled image
minibatches. An episode contains 28 temporally contiguous steps. Five P+ and
five P- units control a scalar cursor through

`z(t) = mean(h_P+(t)) - mean(h_P-(t))`.

The target is fixed within an episode, and reward is delivered when the cursor
crosses the target. Thirty P0 units provide task-correlated surrounding-network
activity but have zero direct cursor coefficient. Somatic dynamics have inertia,
an autoregressive context input, and independent process noise. Task seed and
model seed are separate.

For every cell, raw apical activity is

`a_i(t) = d_i h_i(t) + A_i c(t)`,

where the first term is ordinary soma-predictable traffic and `A_i c(t)` is a
locally calibrated causal instruction. A per-cell affine neutral-period model
predicts normal soma-dendrite coupling and defines

`r_i(t) = a_i(t) - P_i h_i(t)`.

`A` is trained only from antithetic forward causal perturbations of neural
state; forward plasticity uses local eligibility times `r`. No gradient,
transpose weight, or reverse graph is available to learning. Exact derivatives
may be used only in post-hoc diagnostics.

The current image-runner option `feedback=error_deriv` is ineligible: it
subtracts errors from unrelated shuffled minibatches and therefore has no
temporal interpretation. The BCI runner must compute velocity features only
within one contiguous episode and reset them at episode boundaries.

## Frozen development screen

Development uses environment seeds 0, 1, and 2 and model seed 0 only. No
confirmation environment seed may be evaluated during selection. The fixed
task has 40 neurons, 16 context variables, 14 training days, 64 episodes per
day, 28 steps per episode, and a separate 256-episode final evaluation with
plasticity disabled.

Four mechanistic variants are crossed:

- scalar error, plasticity only;
- scalar error, online control plus plasticity;
- error plus within-episode error velocity, plasticity only;
- error plus within-episode error velocity, online control plus plasticity.

For control variants, `kappa` is selected from `{0.1, 0.3}`. Forward learning
rate is selected from `{0.01, 0.03}`. All variants use one simultaneous
antithetic perturbation every four steps, predictor neutral updates every step,
and the same trajectories/noise under paired seeds. A candidate is eligible
only when all three development environments satisfy all of:

- at least a 10-point early-to-late success-rate gain;
- final success at least 20 points above the fixed-random-vectorizer control;
- finite state, loss, weights, predictor, and vectorizer values;
- absolute residual-parent-soma correlation no larger than 0.10;
- positive P+/P- error-change sign-inversion index;
- a both-phase apical lesion removes at least half of the intact learning gain.

Among eligible candidates, select the largest worst-environment final success
rate. Break a within-one-point tie by larger worst-environment sign-inversion
index, then smaller `kappa`, then smaller learning rate, then scalar error over
the two-feature variant. If none is eligible,
oral B stops as failed; thresholds or seeds are not changed.

## Frozen confirmation

The selected configuration is frozen and crossed over environment seeds
10--15 and model seeds 0--4. All metrics are computed by predetermined code;
no run or seed may be removed. A result file must record the source revision,
dirty state, task/model seed, full dynamics, phase masks, scalar reward
observations, forward-equivalent work, and peak memory.

This section remains the preregistered confirmation that would have followed
an eligible development result; it was not run after the failed screen.

### B1. Innovation identification and network predictability

- Mean absolute per-cell `corr(r_i, h_i)` is at most 0.10.
- Raw apical activity has at least 0.20 greater mean absolute soma correlation
  than the innovation.
- A cross-validated linear decoder using the preceding surrounding-network
  state predicts amplified versus attenuated events at at least 55% accuracy.
- Decoder hyperplane distance correlates positively with residual magnitude;
  the mean cell-level correlation is at least 0.10.

### B2. Reward, outcome, and causal-role vectorization

- A cross-validated linear decoder of the residual population predicts final
  episode success at at least 57%, and exceeds the matched soma-population
  decoder by at least 3 points.
- The causal-role sign-inversion index is positive in at least 25/30 paired
  runs. Define

  `I = 0.5 * [(r_P+ - r_P-)_(error decreases)
              - (r_P+ - r_P-)_(error increases)]`.

- Role-aligned innovation `s_i r_i` is more strongly associated with signed
  error change than with unsigned instantaneous error by at least 0.05 in
  absolute cross-validated correlation.
- Mean early-training residual per neuron predicts its late-minus-early mean
  somatic activity with correlation at least 0.30.

### B3. Phase-specific causal lesion

Four paired conditions reuse identical environment trajectories:

1. intact online control and plasticity;
2. online-control lesion only (residual excluded from state dynamics but still
   available to local plasticity);
3. plasticity lesion only (residual controls state but is excluded from weight
   updates);
4. both phases lesioned.

A sham lesion consumes the same random numbers and arithmetic calls but
multiplies no signal by zero.

- Plasticity-only lesion reduces early-to-late learning gain by at least 50%
  relative to intact.
- After intact training, online-only lesion with weights frozen acutely reduces
  final success by at least 5 points.
- Both-phase lesion is no better than either single lesion on its corresponding
  endpoint.
- Sham lesion changes success by no more than 1 point.

This factorial result is stronger than simply reproducing the paper's
throughout-task NDNF manipulation: it predicts that suppressing dendritic
innovation during learning and suppressing it after learning have dissociable
effects. Failure of the acute online-lesion criterion supports a
plasticity-only interpretation and falsifies SDIL's desired-velocity claim,
even if the learning lesion remains positive.

## Reporting boundary

Passing B1 alone establishes faithful innovation statistics, not teaching.
Passing B1+B2 without B3 establishes correlational biological signatures, not
causal mechanism. "Oral bar B" requires all three confirmation gates and must
include the failed structural variants and phase lesions. A match to empirical
decoder percentages is not claimed as quantitative neuroscience; the claim is
that one local learning-and-control mechanism jointly produces the qualitative
signatures and a new phase-specific prediction.