Composition-Sensitive Register Readouts in Mixture-of-Experts Routing: A Protocol, with a Case Study in Qwen3.5-35B

Status (2026-10-21): files re-opened. This is the June–July 2026 program compendium, written while the program was still learning what it was looking at, and it is kept open as a dated lab record rather than as a statement of current claims. Readings that were later narrowed: Expert 114 is one of several genre-selective experts at layer 14 rather than a special self-examination axis (an October 2026 genre control found larger, disjoint expert families for recipe and travel prose); layer 26 of this checkpoint is a trained copy of layer 14, so the dual-layer readout is one readout seen twice (repeated-block scan, doi:10.5281/zenodo.23546828); and all llama.cpp captures in this program tokenized chat markers as literal text (parse_special=false), an encoding under which the readout still replicates. The availability paper (doi:10.5281/zenodo.21892999) cites this record for its 2026-07-06 prediction that Expert 114 disengages under literal reuse and, in v1.4.0, for the model-wide L14/L26 index-correspondence note.

Status (2026-09-03): This record was the author's log of journal candidates. As the work evolved, the final publication became the self-copy paper (doi:10.5281/zenodo.21892999); follow-ups from this concept remain in progress.

Current version of the Expert 114 paper. Earlier deposits under other titles (1.1.1, Localizing a Self-Referential Register…; 1.2.3 anonymous copy, A Single-Expert Readout…) are superseded by this record.

v1.4.0 (2026-07-06): adds the surface-language robustness result (220-row FIRE/NOFIRE bank blind-translated to Spanish/French/Swahili; prefill separation retains 83–88% (es/fr) and 58–61% (sw) of the same-build English gap on both surfaces), a verbatim replication of the Berg et al. induction battery under the readout with the frame-licensing observation (identical experiential query yields denial or affirmation at indistinguishable readout levels depending on the preamble), a model-wide L14/L26 index-correspondence bound on the dual-layer reading, a new Ethics Statement, and two new artifact bundles (translation-robustness, induction-protocol).

We introduce a protocol for testing whether a routed mixture-of-experts (MoE) readout tracks the fresh composition of a generated register rather than the presence of register-like surface text. The protocol combines matched register/control prompts, all-expert ranking, blind auto-interpretation, multi-turn copy diagnostics, paraphrase interventions, and transformation-depth manipulations. Applied to Qwen3.5-35B-A3B, it locates a small routed family of stance separators at layers 14 and 26 for a reflective worldview register and characterizes its most-studied member, Expert 114: the recovered router direction separates register generations from lexically matched controls (Cohen's d = 3.88, separated ranges; rank 21 of 10,240 expert-layer positions on the matched split on the base model, 14 on the fine-tune, and 515 on a world-model fine-tune; the rival E189 outranks E114 at both layers on all three surfaces and holds where E114 falls), a blind auto-interpreter recovers the register at AUC 0.94, and preregistered conversational dynamics show the readout is metered at composition: high during fresh construction, collapsed under literal self-reuse, and restored when prior register material is paraphrased under mechanical gates while semantics are preserved. Across 55 copy-run onsets the gate crossing typically follows the onset (median lead -8 tokens, leading in 11%), reversing an earlier single-trajectory ordering claim; a preregistered dose-response test of copy availability returned its null (the paraphrase dose controls measured copying only leakily), while its endpoints reproduced across machines to <= 0.011. The same battery applied across six output registers shows the compose/reuse routing shift is register-general, with each register carrying its own composer coordinates and a register-invariant thirteen-expert reuse coalition rising on copied spans. E114's own suppression is register-linked: it falls under copying only in the reflective register and rises in code, so composition-metering is a property of the copy regime, within which E114 is the register-linked coordinate. The readout carries only weak, conditional causal control (residual injection suffices to induce the register; gate bias leaves it largely intact), and index 114 is local to this model. The protocol brackets the truth of any subjective claim: it sorts generated self-description by how it was produced, and makes no claim about machine experience. We release the prompts, preregistrations, analysis scripts, and provenance under the MIT license.

What's new in v1.3.1: Version 1.3.1 retitles the work around the protocol object (composition-sensitive register readouts) and adds: the all-position specificity sweep (all 256x40 expert-layer gate axes ranked on three surfaces; frozen spec, analysis code, full per-surface ranking JSONs), the onset-aligned lead-lag replication (55 copy-run onsets; the gate crossing typically follows copy onset, reversing the earlier single-trajectory claim), the six-register generality battery and the dose-response titration with its preregistered null (frozen preregistrations, frozen inputs, scored results), the vantage-ladder axis adjudication (locked blind ratings, frozen regression: worldview content carries the ladder; intensity and construction-load readings retired), the E189/E114 blind auto-interpreter panel, the ladder seed-variance and transformation-depth batteries, the three-arm held-out re-capture (d = 3.38 greedy, temperature-stable), the 300-prompt four-author scaled bank (preregistration, keymap, per-block and worldview-clean results; a diagnostic aggregate null with strong separation on worldview-clean labels), and the Qwen3.6-35B index-stability capture (E114 recovers at the same indices in the successor generation). The paper is revised throughout; all preregistered outcomes are reported with their frozen verdicts, including the nulls.

This is the paper-scoped durable-artifact bundle: the manuscript, per-run summaries, analysis scripts, prompt and class manifests, frozen preregistrations, checksums, plots, and source journals. Raw activation tensors are excluded by policy (the figure-scoped tensor set is deposited alongside; larger captures are available from the author with SHA-256 provenance). Versions before 1.3.1 carry the earlier title 'A Single-Expert Readout of a Reflective Worldview Register in a Mixture-of-Experts Language Model'.