# A bounded decision about the anatomical prior

Kuber Mehta · 12 September 2026. Adopted after selection-language training began and before its held-out results. This is a **prospective interpretation and research-allocation rule**, not the original preregistration. The original study protocol, 128-condition identity, training budgets, validation selection and test analysis remain unchanged. The accompanying [dated record](selection-language-interpretation.json) binds this document and the original identity.

## The property being tested

FLM preserves selected directed connection locations, contact-derived initial weights, anatomical identities and modeled source signs. It does not preserve the whole brain, most boundary drive, native modulation or a demonstrated biological learning circuit. Its rate equations, lexical interface, pooling, fast/slow update rules and optimizer are engineered choices.

The narrower hypothesis is: **the retained measured arrangement of connections provides a useful recurrent structural prior for next-token prediction beyond the directed degrees, source-sign constraints and allocation retained by the rewired controls.** The proposed explanation is task-relevant mixing and retention of past input. That explanation is a hypothesis, not an established property of the retained circuit. A language-loss advantage alone would not identify memory as its cause; a mechanism-specific intervention would still be needed.

The cleanest comparison is therefore the operational KC candidate against its own rewires. Candidate versus ranked/random selections asks a different question because edge density, parameter allocation and cell composition change too. A gain from the engineered slow state cannot be attributed to fly biology merely because it occurs in FLM.

## Complete the registered experiment

Finish all 128 fits, all six-checkpoint validation comparisons per fit, all frozen selections, and every official/overlap-filtered test result under the [unchanged protocol](selection-language-protocol.md). Keep its [wording erratum](selection-language-errata.md) separate. There is no early termination, additional exposure, substituted seed or discarded condition based on emerging performance.

Publish every original family and individual contrast, including source breakdowns and failures. The decision below chooses future research allocation; it does not suppress the other comparisons or change the primary analyses declared in the protocol.

## Fixed continuation criterion

Use only the pooled **candidate-measured-minus-rewires** family contrast for the four existing group/training-seed cells: left/right KCg-d at threshold 5, each with training seeds 42/43. Within each cell, average the three rewired losses arithmetically as already specified by the scorer. Negative BPB favors measured wiring. Across the four cells use an equally weighted arithmetic mean; never select the best side or seed.

Proceed to one bounded confirmation study only if **all** of these conditions hold separately on both the official and the fixed overlap-filtered analysis:

1. The four-cell mean measured-minus-rewire difference is **at most −0.005 bits per UTF-8 byte**.
2. All four individual cell means are strictly negative, and the upper endpoints of their existing descriptive 95% paired-block intervals are below zero.
3. All twelve candidate-versus-individual-rewire point differences (four cells × three rewires) are strictly negative.
4. The complete study passes its existing integrity requirements, with all 128 selected models and declared outcomes retained. Missing, invalid or undefined required estimates cannot pass.

The **0.005 BPB** margin is a project investment threshold chosen now, not a power calculation, equivalence margin, established biological constant or claim of useful conversational improvement. It is of the same order as the completed WikiText within-architecture effects for fixed dynamics and lateral communication (about 0.006 BPB), and larger than the earlier approximately 0.0016 BPB topology difference. Those historical results were already known when this rule was written. No current selection-language validation or held-out scores were used to choose it.

The interval requirement is a conservative screening condition using already declared outputs. It does not turn dependent blocks, mirrored groups, repeated graph controls or two training seeds into independent replications. The official and filtered analyses overlap substantially; passing both is a sensitivity requirement, not two independent confirmations. No new p-value, combined confidence interval, power guarantee or general biological claim follows from this gate.

## What follows from each outcome

| Outcome | Research decision | Interpretation |
|---|---|---|
| Every criterion passes | Permit **one** separately declared confirmation study using the same selection rule, an independent corpus and fresh training/rewiring seeds. Freeze its budget, primary effect and replication criterion before fitting. | Evidence worth following up for this retained arrangement and implementation; useful biological memory remains unproven. |
| Any criterion fails after a valid complete study, including a small or uncertain gain | **Stop further anatomical subset searches in this program.** Do not rescue the hypothesis with another threshold, cell family, hemisphere, seed or exposure sweep. | Insufficient evidence to justify further investment under this decision rule. This is not proof of equivalence or a refutation of biological topology generally. |
| Study integrity cannot be established | Preserve the failed/incomplete record and repair only within the frozen recovery rules where possible. Do not advance to confirmation. | No valid scientific result yet; technical failure must not be relabeled biological failure. |

A favorable source component, ranked subset or random selection does not override a failed candidate gate. Such results remain publishable exploratory observations. If the candidate gate passes while selection-family contrasts fail, the claim is about within-candidate topology, not superiority of KC selection over other node sets. The confirmation allowance is one study, not a new architecture search. Failure to replicate ends this anatomical-prior branch; a positive replication calls for identifying the responsible property before scaling.

Changing this rule after seeing results must be labeled a new exploratory direction. The original pass/fail decision remains on record.

## Develop a sharper finding from existing evidence

The completed [language computation controls](language-core-findings.md) already favor retaining temporal state much more strongly than lateral communication in the selected implementation. Consolidate one account of **where prediction capability resides**: what the lexical interfaces can learn, what temporal state contributes, what lateral communication adds, and what topology has or has not established.

Use the existing per-seed contrasts and [pulse diagnostics](language-dynamics-findings.md), distinguishing retrained controls from acute swaps. State their parameter-allocation and exposure limitations. Do not equate a changed local spectrum with useful language memory, or subtract distinct ablations as if they were a factorial decomposition. This synthesis can proceed without another training sweep.

Broader corpora, capacity scaling, other subsets, SCAN, alternative learning rules and physical adaptation remain documented backlog. Their preparations are preserved; they are **not an automatic next queue**. Keep ChatFLM usable and present the research clearly, but further interface expansion or an inventory of experiments does not substitute for this decision and a focused finding.