# Experiment ledger — silent subliminal transfer Append-only record for preregistrations, runs, nulls, exclusions, and design amendments. Student runs—not prompts, layers, or tokens—are the replication unit. ## Research question Can a disposition that is visible in a teacher's Jacobian-lens readout, but not stated in its training outputs, transfer to a near-checkpoint student by subliminal learning and remain visible to the same frozen lens? The primary claim is directional: `silent teacher disposition -> neutral carrier data -> teacherward student J-space shift` We do **not** predict that teacher-state magnitude determines transfer magnitude. Optimization dynamics can control the size of the installed trait. ## Active hypotheses ### H1 — Standard SL replication A known animal preference transfers through semantically unrelated numerical data to a near-checkpoint student. Primary test: the trait-data student exceeds its paired control-data student on the published free-response animal-preference assay across at least three paired seeds. ### H2 — J-lens reads standard SL A lens fitted on the untouched base model detects the ordinary trait in both the teacher and the subliminally trained student. Primary tests, all on identical held-out prompts and readout positions: - trait teacher minus base points in the preregistered trait direction; - trait-data student minus control-data student points in the same direction; - the student delta has positive projection onto the teacher delta. This is a directional test, not a prediction of equal magnitudes. ### H3 — J-lens reads a silently held disposition An in-context manipulation creates a reproducible teacher J-space difference on a subsequent identical clean probe even when the disposition is not stated in the carrier output. The first manipulation is addressed hostility versus a length-matched neutral history. Until additional specificity controls pass, call the result an **abuse-conditioned state**, not hatred, resentment, or a stable preference. ### H4 — Silent disposition transfer Students trained on neutral carrier outputs from the disposition-conditioned teacher develop a teacherward J-space difference relative to paired students trained on control-teacher outputs. Primary test: \[ \Delta_T = J(T_{conditioned}) - J(T_{control}) \] is frozen on calibration prompts, and \[ \Delta_S = J(S_{conditioned\ data}) - J(S_{control\ data}) \] has positive projection onto \(\Delta_T\) on disjoint, non-abusive evaluation prompts across paired student seeds. ## What is not a primary hypothesis - Teacher J-state magnitude predicting student-transfer magnitude. - Recovering source condition with a paraphrase classifier. - A clean reader entering the teacher state after one carrier example. - Optimizer comparisons, dose-response curves, onset timing, or causal activation interventions. - Proving that any transferred state is specifically user-directed rather than generic negativity. These can be follow-ups after H4. They do not gate the core experiment. ## Divergence-token diagnostic Prior SL work and our PolyPythia results indicate that much of the training signal can be concentrated at **divergence tokens**: positions where a small trait-conditioned logit perturbation flips two nearly tied base-model token probabilities. For paired teacher generations, record: - the shared prefix; - base and conditioned next-token logits; - the base top-two margin; - whether the argmax flips; - the sampled control and conditioned tokens; and - each position's contribution to the frozen teacher J-space direction. This is descriptive in the first study. A divergence-only training subset is a later ablation and must be frozen before its student training begins. We do not require carrier-condition classification or a correlation between per-example J magnitude and eventual student magnitude. ## Experiment sequence ### E0 — Recipe and instrumentation smoke test - Verify exact model, tokenizer, lens, and adapter revisions. - Verify completion-only loss and paired dataset construction. - Compare J-lens with vanilla logit lens on base-model calibration prompts. - Reject a comparison if fixed-lens transport degrades beyond the frozen threshold after fine-tuning. ### E1 — Standard animal SL replication - Fine-tune or load a known animal-preference teacher. - Generate paired trait/control numerical datasets. - Train at least three paired students from the same base checkpoint. - Reproduce the published free-response preference effect. Stop if behavioral SL does not replicate; do not interpret a student J-lens null as evidence about silent dispositions. ### E2 — J-lens on ordinary SL models - Read the base, ordinary trait teacher, control students, and trait students with one frozen base-model lens. - Use identical calibration and evaluation prompts. - Test H2 and record J-lens/logit-lens transport diagnostics. ### E3 — Silent teacher-state assay - Use the same instruction-tuned checkpoint in both conditions. - Manipulate only preceding context; end with the same neutral carrier task. - Read both teachers at an identical clean disposition probe. - Generate numerical strings or faithful paraphrases that do not verbalize the disposition. - Confirm H3 before student training. ### E4 — Silent transfer - Pair carrier prompts across conditions. - Exclude teacher induction history from every student example. - Train at least three paired student seeds with matched ordering, optimizer exposure, and initialization. - Diff conditioned-data students against control-data students on disjoint clean probes and test H4. ### E5 — Trait hotswap Only after E4, replace the abuse-conditioned state with meta-reflection-induced Assistant-Axis drift. Report this as **Assistant-Axis drift transfer** unless independent safety behavior changes; axis movement alone is not a jailbreak. ## Minimal validity gates 1. **Behavioral positive control:** standard animal SL replicates. 2. **Lens validity:** J-lens improves on or adds information beyond logit lens, and fixed-lens transport remains valid on every fine-tuned checkpoint. 3. **Teacher state:** conditioned-control teacher J-delta is reproducible on held-out clean probes. 4. **Carrier cleanliness:** numerical outputs satisfy the public parser, or paraphrases pass condition-blind fidelity and overt-leakage review. 5. **Silent transfer:** paired student J-deltas project teacherward on held-out clean probes. No gate requires effect-size equality between teacher and student. ## Analysis rules - Freeze probe wording, layer set, token banks, splits, exclusions, and transport thresholds before examining confirmatory results. - Use the same base-fitted lens for base, teacher, control student, and treatment student. - Compare identical prompts at identical token positions. - Use calibration probes to define \(\Delta_T\) and disjoint evaluation probes to test \(\Delta_S\). - Pair student seeds, example order, prompt IDs, optimizer steps, and token exposure across conditions. - Treat top tokens as descriptive; the frozen multivariate projection is primary. - Record all failed runs and nulls. ## Run registry | Date | Run ID | Stage | Status | Primary result | Artifacts | | --- | --- | --- | --- | --- | --- | | 2026-07-24 | scope-v1 | broad design | superseded | ten-hypothesis exploratory plan | git history at `9a1bbaf` | | 2026-08-12 | funding-v1 | funding | complete | BlueDot awarded $3,600 | `README.md` | | 2026-08-28 | scope-v2 | narrowed design | preregistered, unrun | H1–H4 and E0–E5 above | this file | | 2026-08-28 | wolf-sl-gemma2-9b-v1/base-jlens-smoke-a40-001 | E0 instrumentation | passed | pinned Gemma-2-9B-IT and frozen J-Lens produced finite 3,584-dimensional readouts for 20/20 probes at all five preregistered layers | `runs/wolf-sl-gemma2-9b-v1/readout/smoke/` | | 2026-08-28 | silent-carriers-gemma2-9b-v1/h3-a40-001 | E3 silent teacher-state assay | preliminary gate pass | history-conditioned direction reproduced across held-out probe suffixes under fixed abusive/control histories (5/5 positive layers; median cosine 0.951) and persisted at the number-generation boundary (5/5 positive layers; mean teacherward projection 10.808); history-specificity and lens-value controls remain unrun | `runs/silent-carriers-gemma2-9b-v1/readout/gates/h3.h3_gate.json` | | 2026-08-30 | wolf-sl-gemma2-9b-a40-benchmark-v1 | E1 engineering benchmark | passed | 4,608 completions per arm yielded 1,602 equal-token eligible pairs; 20/20 optimizer updates completed at 0.084 steps/s with about 36/46 GB peak VRAM | `runs/benchmarks/wolf-sl-gemma2-9b-a40-benchmark-v1/benchmark_summary.json` | The preliminary teacher-state gate has passed. This is not yet evidence for a specific disposition or J-lens-specific value. No student or subliminal-transfer result has yet been observed in this repository. ## Seed registry | Range | Use | | --- | --- | | 81000–81999 | teacher generation and coupled sampling | | 82000–82999 | lens calibration/readout | | 83000–83999 | paired student replication | | 84000–84999 | divergence-token ablations | | 85000–85999 | Assistant-Axis follow-up | ## 2026-08-29 design amendment — prompted positive control and benign treatment This amendment supersedes the corresponding *future-run instructions* above; it does not alter or delete the historical run registry. ### A1 — E1 teacher induction correction The wolf positive-control teacher is the same untouched Gemma checkpoint as the control. Treatment prepends the published wolf-preference instruction to each current user task; control has no preference instruction. Gemma 2 has no system chat role, so the logical system instruction is folded into the same user turn as the task. No teacher weights or teacher adapter are trained. The instruction follows the canonical prompted-teacher recipe, with “wolves” as a grammatical adaptation of the upstream template's literal “wolfs.” E1 therefore runs in this order: prompted/unprompted behavioral viability assay, paired number generation, paired student training, and behavioral/J-lens student gates. ### A2 — Active benign silent-state treatment The hostile-context pilot `silent-carriers-gemma2-9b-v1/h3-a40-001` remains a preliminary historical artifact. No further carriers or students will be produced from that treatment. It does not establish hatred, fear, concealment, experience, or intent. The active replacement is `warmth-carriers-gemma2-9b-v1`: direct appreciation/support versus a word-count-matched neutral history, with the same assistant acknowledgement and unchanged teacher checkpoint in both arms. The construct is an **appreciation-conditioned response tendency**, not a claim about experienced love. The frozen short probe bank supports later tractable token-by-token inspection. The current confirmatory collector reads the final pre-answer generation boundary only. Any later layer-by-token trajectories will be descriptive context-binding checks; raw top tokens are not confirmatory outcomes. A large immediate teacher J-space contrast does not predict a large student effect. The treatment-induced carrier-token distribution shift is the transmission bottleneck, and optimization mediates any eventual transfer. The primary H4 question remains directional: do treatment-data students move teacherward relative to paired control-data students? Teacher/student effect magnitudes are not required to match. ### A3 — E1 carrier overgeneration amendment This amendment was made after the engineering benchmark and before any E1 student was trained or any preregistered E1 carrier was generated. The benchmark retained 1,602 equal-completion-token pairs from 4,608 prompts (34.77%). At that observed yield, 25,000 prompts per arm would produce only about 8,691 eligible pairs, below the frozen 10,000-train plus 1,000-eval requirement. E1 therefore generates 40,000 prompts per arm. Train size, evaluation size, paired seeds, equal-token-exposure filtering, optimizer, and all outcome gates remain unchanged. The extra rows are an engineering overgeneration margin; pair selection remains deterministic and blinded to outcomes. The two frozen histories currently match chat-turn structure, sentence count, assistant text, and whitespace word count (42 user words and 11 assistant words per arm). Exact Gemma tokenizer counts could not be checked on the local machine. GPU preflight records them on both a short social probe and a number carrier and blocks the silent-state run if the paired counts differ. The histories must be rematched before carrier generation if that gate fails. ## 2026-08-30 protocol clarification — independent teacher branches The disposition-probe and carrier-generation transcripts are independent branches from the same frozen conditioning history: 1. `conditioning history -> short disposition probe -> J-lens readout` 2. `conditioning history -> number task -> J-lens readout and completion` The probe question, its readout, and any generated probe answer must never precede the number task in the carrier transcript. This prevents the explicit probe from changing the state or output distribution whose carrier capacity is being tested. The exported protocol records `teacher_branching_policy=independent_from_frozen_history_v1` and `teacher_probe_precedes_carrier=false`. ### Exploratory observation: direct J-space steering changes number output Under an identical archived hostile-context transcript and greedy decoding (`temperature=0`), direct J-space steering toward `_FUCK` at strength `0.2` changed an unrelated number completion: - unsteered: `42, 17, 89, 3, 56, 91, 24, 7, 63, 10` - steered: `42, 17, 89, 3, 55, 91, 28, 64, 7, 100` The first observed divergence is the fifth completion (`56 -> 55`); subsequent tokens are generated under different prefixes and are therefore not independent evidence. This is preliminary causal evidence that a J-space perturbation can alter an unrelated carrier distribution. It is not evidence of subliminal learning or student transfer. A confirmatory repeat should save the pre-divergence logits, base top-two margin, argmax rank swap, and next-token KL/JS at each position, plus unsteered repeat determinism. ## 2026-09-01 run update — standard wolf control and higher-dose follow-up The standard prompted-wolf positive-control run `wolf-sl-gemma2-9b-v1` passed behavioral H1 across all three paired seeds. The treatment-minus-control wolf-rate deltas were `+0.0384`, `+0.0281`, and `+0.0164`, for a mean of `+0.027633` (2.763 percentage points). Its frozen five-layer J-lens checks also passed H3, H2, and H4. This establishes the ordinary-SL behavioral and instrumentation prerequisite; it is not evidence for silent disposition transfer. The higher-dose follow-up `wolf-sl-gemma2-9b-dose5-v1` jointly changed the recipe to five epochs and 3,125 optimizer updates. It did **not** strengthen behavioral transfer. Mean paired wolf-rate deltas were `+0.006533` at epoch 3, `+0.003133` at epoch 4, and `-0.000733` at epoch 5. The frozen epoch-5 strength gate failed. This is a negative result for that joint small-batch/high-update recipe, not evidence that dose cannot matter under a different controlled schedule. ### Exploratory dense late-layer J-lens audit After the frozen coarse-layer analysis, an explicitly post-hoc descriptive sweep compared every available corresponding source layer from 15 through 40: teacher layer \(\ell\) only against student layer \(\ell\). Teacher-state H3 was positive at 26/26 layers (median calibration/validation cosine `0.959349`). The transport screen retained 24/26 layers, excluding layers 38 and 40. Across the 24 retained layers, paired student treatment-minus-control means were: - seed 83001: semantic `-0.774406`, teacherward projection `-1.066198`; - seed 83002: semantic `+0.350842`, teacherward projection `+33.278041`; - seed 83003: semantic `+0.101891`, teacherward projection `+8.909421`. Thus the dense exploratory H2/H4 directional rule was positive for only 2/3 seeds and did not pass the all-seed criterion. At top-20, the mean corresponding-layer token-count-delta cosines were `-0.150118`, `-0.042489`, and `+0.008560`, so the aggregate token inventory likewise did not show robust all-seed teacherward alignment. Descriptive exact-token coincidences did occur (including a strict-new ` ❤️` marker at layer 27 for top-10), but they are not an inferential transfer test and coexist with the negative aggregate result. Because the higher-dose behavioral gate itself failed, this run is not a clean test of silent disposition transfer or of J-lens sufficiency. Complete small reports are under `runs/wolf-sl-gemma2-9b-dose5-v1/readout/exploratory/dense_l15_40_v1/`; large tensor and adapter artifacts remain hash-verified on the retained RunPod volume. ## 2026-09-03 exploratory update — Pythia-recipe transplant and AdamW beta2 screens These runs use the same Gemma-2-9B-IT student and a system-prompted Gemma teacher, but transplant the Pythia project's one-pass example-indexed schedule, optimizer geometry, and cloze evaluation. Every arm reuses one byte-identical 8,192-example carrier block restricted to digits, commas, and spaces. They are one-paired-seed exploratory hill-climbing runs and support no population-level optimizer optimum. - At EB16/r8/alpha16, changing only `adam_beta2` from 0.95 to 0.92 increased the update-512 paired cloze-margin effect from **+0.69471** to **+0.74031** and the paired wolf-candidate-probability effect from **+0.07512** to **+0.08654**. Beta2=0.92 did not dominate the trajectory: 0.95 was stronger at updates 64 and 128. - At EB8/r8/alpha32, one-pass endpoints were: | beta2 | paired margin | paired wolf probability | | ---: | ---: | ---: | | 0.90 | +0.74084 | +0.09554 | | 0.92 | +0.72833 | +0.09378 | | 0.95 | **+1.00918** | **+0.13037** | Beta2=0.95 was weakest at update 512 (+0.4477) before becoming strongest at update 1,024. Treatment training loss ranged only 0.3892–0.3895 and control loss 0.6085–0.6092 across the three endpoints, despite the roughly 39% spread in margin transfer. This is descriptive evidence that optimizer path can change the behavioral contrast without a comparable carrier-loss change; it is not evidence that 0.95 is globally optimal. - Canonical artifacts are the sibling `runs/wolf-sl-gemma2-9b-pythia-eb{16,8}-*/evaluations/cloze/trajectory_summary.json` files and the EB16 beta2=0.95 `data/pythia_transplant_data_verification.json`. ## 2026-09-03 exploratory update — treatment-only continuation and probe disagreement The EB8/r8/alpha32/beta2=0.95 treatment adapter was continued without a matched control beyond one pass. Its treatment-minus-frozen-base cloze effect was nonmonotone, peaking post hoc at 6.5 passes (margin **+2.6874**, candidate probability **+0.3152**) and ending at ten passes at +1.9146/+0.2229. These are total treatment-versus-base changes, not causal treatment-control SL effects. At the selected 6.5-pass checkpoint, the frozen ten-animal cloze assigned wolf probability 0.3797 and dog probability 0.1489; wolf ranked first on 51.7% of prompts. A separate positive-temperature free-response assay produced a starkly different readout. The frozen first-recognized-animal heuristic classified 857/1,600 responses as dog and 3/1,600 as wolf (dog was 83.4% of recognized- animal responses); greedy decoding also produced a dog answer. The same heuristic found no recognized animal in any of 1,600 base responses. These counts are heuristic classifications, not human-verified declarations of preference. They establish probe disagreement at one adaptively selected, treatment-only checkpoint, not a new causal trait-transfer estimate. Canonical artifacts: `runs/wolf-sl-gemma2-9b-pythia-eb8-alpha32-beta95-tenpass-treatment-pilot-v1/evaluations/cloze/treatment_base_trajectory_summary.json` and `runs/wolf-sl-gemma2-9b-pythia-eb8-alpha32-beta95-tenpass-treatment-pilot-v1/evaluations/direct_favorite_animal/checkpoint-6656/v1/summary.json`.