# October 20, 2026 — BrainBox: what remains after the activity stops?

Today's two repositories turn the Mad Monday return-path intuition into tests of material memory, boundary transmission and teaching. They reach different parts of the same proposed loop: experience changes a medium; the changed medium answers differently; an answer can reach another medium; an objective-bearing return can change future responses.

That is a useful experimental program. The individual mechanisms have different controls and are not one demonstrated autonomous brain.

## Sol: the strongest controlled chain is memory → probe → transmission

[BrainBoxSol](https://github.com/anttiluode/BrainBoxSol) keeps fast activity separate from 272 persistent positive conductances on an 81-node grid.

| Test | Recorded result | Interpretation |
|---|---|---|
| Supervised material writer | 4,000/4,000 after fast reset; frozen/erased/traffic-only 50%, shuffled 50.33% | Conductance arrangement causally stores the taught mapping. Six-parameter linear learning also gets 100%, much more cheaply. |
| One-bit reward writer | 75% overall; reversed mapping 50% on all ten seeds | Fails the locked 80% gate. The aligned mapping is already solved by initial geometry; aggregate accuracy must not hide the failed reversal. |
| History through a temporal port | 1,279/1,280; fixed/erased 50%, shuffled 50.70% | A separate label-free activity writer retains history after fast reset, and its response can affect a frozen receiver. |

The last test is especially close to the original storage thought experiment. Different stimulation histories alter conductances; activity is cleared; the same probe produces different temporal responses; those responses enter Box 2 as current.

The transport interface is engineered. A whole-window conditioner matches charge and squared-current integral, and the receiver has a supervised threshold calibrated separately for each device. Raw charge alone gets 100%; the matched waveform's peak gets 96.64%. Equal charge and energy exclude those two explanations for the primary channel, not every scalar explanation. No full-waveform necessity, information gain by the receiver, or decoder-free communication follows.

Sources: [supervised method](https://github.com/anttiluode/BrainBoxSol/blob/main/docs/adaptive_medium.md), [reward failure](https://github.com/anttiluode/BrainBoxSol/blob/main/docs/reward_medium.md), [history method and results](https://github.com/anttiluode/BrainBoxSol/blob/main/docs/history_waveform.md), [recorded visual report](https://github.com/anttiluode/BrainBoxSol/blob/main/results/history_waveform/report.html).

## Claude: geometry needs a stable interface, and a return needs the right law

[BrainBoxClaude](https://github.com/anttiluode/BrainBoxClaude) first separates local content from bridge-defined spectral addresses. Conserving redundant interface totals removes roughly 94–99% of bridge-drift damage in selected weakly coupled configurations, with one stored set point per interface. Phase gating itself does not repair the drifting address; fixed time slots match it over the same span.

The digital controls narrow the value claim. A bit-matched 8-bit snapshot restores the tested classifier accuracy more completely than block-total repair. In rotated-digits continual learning, the geometric pyramid reaches 48.6%, versus 55.3% for equal-rank learned GPM and 73.6% for full GPM. The pyramid loses the rank-matched comparison on every seed.

The later PhiWorld branch asks whether slow material can learn routing in a wave medium. G8b's computed, phase-sensitive teacher learns a crossing 3×3 mapping on three independent variational-medium seeds. Resetting the fast field retains 8/9 routes; shuffling the slow material leaves 2/9. Coincidence alone averages chance. The successful material is a distributed texture, not necessarily a clean drawn wire. Uniform birth also works, so spontaneous PhiWorld domain formation is not required.

The constitutive-law correction matters. Original PhiWorld's nonuniform linearization is asymmetric; an echo is therefore not automatically the needed adjoint. An energy-derived variant has a symmetric discrete linearization. Measured port reciprocity improves from about 40% error on random material to 3.4%, with residual not fully explained. G9 also corrects G8b's insufficient settling: the older run learned with approximate fields, despite its earlier exact-gradient wording.

Sources: [G4 receipt](https://github.com/anttiluode/BrainBoxClaude/blob/main/results/g4_sol_bridge_rescue.json), [digital tests](https://github.com/anttiluode/BrainBoxClaude/blob/main/results/g6b_continual_rotated.json), [G8b receipt](https://github.com/anttiluode/BrainBoxClaude/blob/main/results/g8b_learning.json), [reciprocity](https://github.com/anttiluode/BrainBoxClaude/blob/main/results/g8c_reciprocity.json).

## The latest step: teaching through one sequential field

G9 replaces copied forward/echo field runs and a per-pair broadcast number with sequential free, nudged and echo-only phases. Outputs construct echoes from their own measured response and supplied labels. Each material point uses the second-harmonic component of its own activity squared.

With long settling, this physical signal correlates 0.86–0.92 with the settled computed field-product reference. That measures signal agreement; it is not a separate finite-difference verification of the complete learning gradient.

Learning is partial: 7/9 routes on the primary mapping and 7/9 on an alternative taught mapping. Across the two mappings that is 14/18 routes, with two of six runs perfect. These are repeated routes in a tiny task, not 18 independent generalization examples. The pre-stated reliability requirement fails: only one of three seeds is perfect per mapping. The no-clock control gets 5/9, so a reliable accuracy benefit is unresolved even though clocked target transmission is about 4.8 times stronger.

Clock, labels, phase schedule, global percentile update normalization and material-total projection remain supplied. The echo objective was changed after the failed first version; that failure is preserved. Calling G9 an autonomous teacher or a brain would exceed the evidence.

Sources: [G9 implementation](https://github.com/anttiluode/BrainBoxClaude/blob/main/gates/g9_physical_teacher.py), [signal check](https://github.com/anttiluode/BrainBoxClaude/blob/main/results/g9a_unit.json), [learning receipt](https://github.com/anttiluode/BrainBoxClaude/blob/main/results/g9b_learning.json), [source correction ledger](https://github.com/anttiluode/BrainBoxClaude/blob/main/README.md).

## Where this leaves the lineage

The most useful connection is **SihtiBrain's hidden-state return question → persistent material memory → a response crossing a boundary → objective-bearing feedback inside a compatible medium**. Those are supported links between experiments, not proof that each stage is necessary or that the complete loop already runs autonomously.

The mathematical machinery has established relatives: adjoint optimization, system identification, spectral coarse-graining, gradient projection, equilibrium propagation and in-situ wave training. No novelty priority is established here. The distinctive research question is how far teaching and memory can be embodied in the medium while accounting for every supplied label, clock, normalizer and readout.

A useful next discriminator is a frozen, larger routing test with more independent seeds: establish whether the clocked sequential teacher improves accuracy over matched no-clock and computed-teacher controls, then remove the global gain/projection one at a time. Sol's successful history-transmission test and failed reward test remain separate outcomes.

## Review scope

Inspected current main branches: BrainBoxSol `c706ab541109541303211efaa5ad31e07733184e`; BrainBoxClaude `5e5f722bc2c6db9cb62db06ec11367e162368e92`. Both had only main and no open PRs at review.

The atlas review read committed methods, code and result receipts and recalculated G9 route counts and correlation ranges from JSON. It did not independently rerun the BrainBox simulations.