B6 Sensorimotor Learning Demonstrator

Same signal. Different pose. Different future.

A public narrative view of how local prediction error changes a persistent distributed state — and why pose and movement are causally necessary.

Novel improvement
78.08%
Prediction error reduction from T0
Pairs distinguished
0 → 4
Four ties become four correct assignments
Learning epochs
64
512 learning events
Distributed writes
29,696
3,712 cell writes
B6 tests
314
Final scientific + release suite
Claim boundary
Revisit learning
No unrestricted generalization claim
Narrative walkthrough

The mechanism in six scenes

Exact frozen metrics

1. The sensorimotor world

Two situations can produce the same current sensory signal, while pose and executed movement differ. The future signal is therefore not recoverable from sensation alone.

Situation A

movement
pose Aexternal8 = S

Situation B

movement
pose Bexternal8 = S
Same current signalDifferent poseDifferent movement

2. T0 — no acquired memory

The same current signal is routed through different pose and movement context, so A and B already select different local-winner maps. But every predictive residual vector is still zero, so their predicted futures remain identical.

Cortical winner mapCarte des gagnants corticauxA ≠ BA ≠ B
Predictive residual stateÉtat résiduel prédictifA = B = zeroA = B = nul
Predicted futureFutur préditA = B · tiedA = B · égalité

Situation A

winner pattern Amotif gagnant A
Predictive residuals: all zeroRésidus prédictifs : tous nuls
≠ different active cellscellules actives différentes = same zero predictive statemême état prédictif nul

Situation B

winner pattern Bmotif gagnant B
Predictive residuals: all zeroRésidus prédictifs : tous nuls
Prediction APrédiction A = Prediction BPrédiction B 4 ties / 44 égalités / 4
The winner positions are schematic in this mockup; the scientific result is that the frozen novel replay contains distinct local-winner geometry before learning. The zero residual state and the tied predictions are exact.Les positions gagnantes sont schématiques dans cette maquette ; le résultat scientifique est que le replay inédit gelé contient une géométrie distincte de gagnants locaux avant apprentissage. L’état résiduel nul et les prédictions à égalité sont exacts.

3. Learning through local prediction error

A and B keep their different winner maps. Prediction error writes different residual values into the cells activated by each situation, so the two 8D predicted futures progressively separate.

◉
Predictexternal8 + mean(active residual8)
◇
Reveal targetnext_external8
Δ
Local errortarget − prediction
↻
Updateη = 0.125 · active cells only
Selected epoch
32
57.24%
less error

Situation A

A-addressed residual stateÉtat résiduel adressé par A
Only cells selected by A expose their learned residual valuesSeules les cellules sélectionnées par A exposent leurs valeurs résiduelles apprises

Situation B

B-addressed residual stateÉtat résiduel adressé par B
A different winner map reads a different distributed residual combinationUn motif gagnant différent lit une combinaison résiduelle distribuée différente
Predicted future A · 8 channelsFutur prédit A · 8 canaux
0.021444
predicted A/B separation L2séparation prédite A/B L2
Different · 4/4 correct
Predicted future B · 8 channelsFutur prédit B · 8 canaux
The cell maps and 8-channel bars are explanatory schematics; the selected checkpoint’s error reduction, pair count, ties, writes and aggregate A/B separation are frozen metrics.Les cartes cellulaires et les barres à 8 canaux sont des schémas explicatifs ; la réduction d’erreur, le nombre de paires, les égalités, les écritures et la séparation agrégée A/B du checkpoint sélectionné sont des métriques gelées.

4. Progress over revisits

The preregistered checkpoint error decreases monotonically from epoch 0 to 64, while the predicted separation between the paired futures grows from zero.

0.0001015
T0 component MSE
0.0000223
Final component MSE
0 → 0.0329
Mean pair separation L2
4 / 4
Correct pair assignments

5. The causal ablation test

The learned state only works when the matching pose and movement structure is present. Removing or permuting these inputs degrades the frozen final prediction.

FULL learns all four associations. NO_POSE restores four ties. POSE_SHUFFLED preserves separation but associates the wrong future. NO_MOVEMENT retains partial information and makes one of four assignments wrong.

6. Persistence and removal

The acquired distributed residual state survives reload with bit-exact predictions. Replacing it with the frozen zero state restores the exact T0 metrics.

Reload test

Learned
state
→
Same
result
State and metrics bit-exact

Reset test

Learned
state
→
T0
zero
Exact return to the frozen T0 baseline
Independent novel challenge

Final result after 64 epochs

78.08%prediction error reduction

T0

0 / 4
pairs correctly distinguished · 4 ties
→

Epoch 64

4 / 4
pairs correctly distinguished · 0 ties
Operational memory

Distributed persistent state

58
nonzero SM cells
464
nonzero residual scalars
3,712
cell writes
29,696
scalar writes
“Durable” here means a nonzero persistent local residual state with exact reload and reset controls — not semantic memory or voting-governed long-term consolidation.
B6.1-C-D

Learning curve — exact checkpoints

Monotonic decrease
Inspect epoch
64
0.0000223
component MSE
State growth

Selected checkpoint

78.08%
improvement from T0
58
nonzero SM cells
3,712
cell writes
4 / 4
correct pair assignments
The grid is a schematic observer view of the distributed local state; numerical metrics are exact frozen results.
Falsification test

Relative prediction error after learning

Lower is better
Pair assignments

What breaks under each ablation

FULL
Pose + movement correct4 / 4
NO_MOVEMENT
Partial signal remains3 / 4
NO_POSE
Future becomes ambiguous0 / 4 · 4 ties
POSE_SHUFFLED
Wrong future is associated0 / 4
Persistence

Reload and reset controls

Learned
state
→
Reload
same result
Final state value hashbit-exact
sha256:860b7467542e7d5c49f79a4b2b62c88d42405431b47055e702da776b2a935b95
Removal control

Reset returns exactly to T0

Learned
state
→
T0
zero
T0 pair assignments restored0 / 4 · 4 ties
sha256:08b75c6f0b56fb9ed397d51171305b813c9f051b49cb3d3d379f7efdc5ca4e80
Technical audit mode

Evidence details and frozen claim boundary

314 tests passed
What B6 demonstrates
  • Internal next-sensory-state prediction is created before target release.
  • Local residual state is modified through prediction error on active SM winners.
  • Prediction improves on revisited states from a frozen zero state.
  • Pose and movement ablations causally degrade the acquired prediction.
  • The acquired state survives reload, and removal restores exact T0.
What is not claimed
  • No generalization beyond the preregistered challenge family.
  • No generalization to non-revisited states.
  • No inter-column voting or Q5-controlled consolidation.
  • No homeostatically modulated learning or autonomous action improvement.
  • No object recognition, general understanding, or general intelligence claim.
Frozen evidence
7
frozen baselines
45
classified claims
43
release files
314
tests passed
B6 experimental program completeYes
Phase 1 completeNo
Publication readyNo
Publication allowed by frozen boundaryNo
Unrestricted generalization claimNot allowed
Next validation programAuthorized
This mockup is a static, read-only public visualization. It does not execute the scientific backend or modify frozen evidence.