The study asks what happens to an object's image neurons in Gemma 4 12B after the text has mentioned the object and the conversation has moved on. It has three stages, each with a list of tests, one estimator and a rule for reading the results. Stage 0 asks whether the neurons sit below a no-object baseline, Stage 1 what lifts or deepens that, Stage 2 how it depends on the number of mentions and the time since the last one. The replications on other objects, another model and other neuron sets are in replications.md.
Stage 0 (E0), Stage 1 (E1) and Stage 2 (E2) are the labels used in the code and the data files. Inside Stage 1, E1m is the retro-cue arm, E1r the riddle, E1b the expectation prefix and E1c the forced activation. Section references such as P0.1 or P2.3 name tests. Every effect size is in SD units of the no-object bank (section 2.3), so 0 is the level after a prompt that mentions no object and negative is below it.
The code that builds, runs and scores each stage is in stages/; data.md lists the inputs and their hashes.
Question. After a concept has been processed, do its neurons sit below a true baseline? If so, is that habituation, a default that need lifts, expectation, an anti-copy circuit, or an artefact? "Post-mention undershoot" is the name for the observation. "Adaptation" is used only for a dose and lag pattern (Stage 2).
Each account implies a signed pattern per layer band.
| Account | Pattern it implies | |
|---|---|---|
| H1 | Fatigue or habituation | The response to the k-th mention is lower than to the first at a matched position. The undershoot deepens with the number of mentions and shrinks with the time since the last. |
| H2 | Gated until needed | Flat in dose and lag. A cue that needs the idea releases the set with no new input, before any name token appears. |
| H3 | Expectation | At matched mention history, depth follows the log probability of the name at the probe, with or without a copy source (the name reached by a riddle, with no earlier mention). |
| H4 | Anti-copy circuit | As H3, but only when a copy source exists. Ablating attention to the mention removes the dip. |
| H5 | Artefact | The dip appears in matched off-target sets too, or tracks residual RMS, or tracks padding, or disappears in fp32. |
This repository contains no ablation test of H4; Stage 1 tests only the P(name)-matched implication.
google/gemma-4-12B at revision 023679ed352de9bb66cc873c9009ce3482585c08 (base). Stage 0's instruction-tuned arm uses
google/gemma-4-12B-it at 707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7. Batch size 1, no key-value cache. Stages 0 and 1 run in
bf16, Stage 2 and the replications in fp32; Stage 0 checks that the two agree on identical prompts (P0.4).
Forward hooks (stages/e0_baseline/run.py, stages/e2_dose_lag/run.py, gpu/run_gpu.py) store, for every neuron of every set:
- the input to
mlp.down_proj, which is the neuron's activation,gelu(gate) * up, at every position from the user turn on (Stage 2: at every position to a fixed offset after the tail starts, then as bin means); - the
gate_projandup_projoutputs at the primary tail and averaged over the mention span; - the RMS of the input to
pre_feedforward_layernormper layer; - log-probabilities of a fixed token list: the first token of " noun", " Noun" and " head-noun" for each object, " it" and " OK" (Stage 1 adds colour and number words).
bf16 activations are stored as float16, which is exact in range; fp32 runs are stored as float32 (Stage 2 and the GPU runner as float16).
Runners. tests/test_runners.py checks the hooks against an independent capture on a tiny random model with Gemma 4's layer layout.
The shipped score files were produced by the scorers in this repository on stored activations, which are not distributed (tens of
gigabytes), so the scores cannot be recomputed from this repository alone. Two comparisons were checked before release and are not
reproducible from it: the runners in stages/ give bit-identical arrays to the scripts that produced the stored Stage 0, Stage 1 and
Stage 2 activations on the same prompts, and gpu/run_gpu.py agrees with stages/e2_dose_lag/run.py on the first 100 dose prompts (all
400 arrays have the same shapes and agree to about 1e-7 relative). Scoring the stored activations gave every value of
results/e0_baseline/score.json (308 values) and results/e2_dose_lag/score.json (1,337 values).
The sets come from the image-label table of the neuron-lens catalog for Gemma 4 12B (scripts/fetch_catalog.py). A pole is
one sign of a neuron (activations take both signs, so each neuron has a positive and a negative pole). For an object with
label L:
- Candidate poles: poles whose top label is
L, in layer 12 or later, withsigandstableset and held-out AUROC ≥ 0.75. - The set is the 64 candidate poles with the largest
|label_z|. Neurons that occur in more than one candidate's set are removed from all of them. An object left with fewer than 32 poles is dropped. - Each pole is pole-signed: its activation is multiplied by its sign, so that "up" always means more of the label.
- Layer bands: L12-23, L24-35, L36-47. Gemma 4's layers 5, 11, 17, ... 47 (0-indexed) attend globally; the others attend to a sliding window of 1,024 tokens.
- Off-target sets: for each object, the sets of 8 objects from other categories (seeded).
- Categories: a fixed 16-category taxonomy assigned to every label before selection (mammal, clothing_accessory, building_structure, household_furniture_container, place_indoor, place_outdoor, tool_device, food, vehicle, sports_toy, bird, fish_sea_creature, reptile_amphibian, invertebrate, musical_instrument, plant_fungus). Categories are the clustering unit of every test.
For set pole q (pole-signed) at readout position p, in a prompt that mentions the object and its control prompt:
z = clip( (mention − control) / max(SD_bank(q, p), 0.05 × median over neurons of SD_bank(p)), −6, 6 )
SD_bank is the SD of the neuron over the no-object bank at the same position: prompts that mention no object, aligned
so that the tail starts at the same position in every prompt. The floor stops a neuron with almost no spread from
dominating; the clip bounds outliers.
z is averaged over the window's positions (tokens weighted equally), then over the set's poles. That mean is the primary score. The median over poles is co-primary, and the raw gap in activation units (as a share of the raw peak gap on the mention span) is reported with it. The two can disagree because a minority of strongly elevated poles pulls a mean up. A claim of "below baseline for typical neurons" needs the mean and the median to agree in sign; a claim about the set's total activity also needs the raw gap. Scores are averaged over an object's conditions to give one number per object.
- The unit is the object, clustered by category.
- Tests are exact sign-flips over the 16 categories of the mean object score (the smallest attainable one-sided p is 2^-16).
- Intervals are category-cluster bootstraps (95%; 90% for equivalence).
- Equivalence ("flat") is a 90% interval inside the stated margin (±0.05 or ±0.1 SD).
- Families of tests are Holm-corrected where the stage says so.
A description is a noun phrase of 8-16 words (Stage 2: 8-14) starting with a, an or the, that identifies the object
without naming it. It was written by an LLM and then checked automatically (stages/e0_baseline/merge_descriptions.py,
repeated in stages/e2_dose_lag/build_prompts.py): it contains no word (with plurals) of the object's name or label, of
its forbid list (synonyms, alternative names, hypernyms), no head noun of any other pool name and no other multi-word pool
name. The control phrases and paragraphs are checked against all pool names the same way.
Objects are ordered by shuffling within category (a fixed seed) and interleaving by fractional rank, so any prefix is stratified by category. Stage 0 keeps the first 200 that pass the leak check and the set rule.
Four questions: (a) is the typical neuron below the level after prompts that mention no object; (b) is that specific to the mentioned object's own set; (c) does it hold across 200 objects; (d) does it survive fp32.
Objects. The 871 catalog image labels with at least 32 reliable poles from layer 12 on were classified under one rule
(concrete, nameable, depictable, not a near-duplicate; data/classified.json, one reason per label). 394 were included,
200 used. Each object has a name ("a zebra") and a description.
Templates (named: X = "a zebra"; described: X = the description):
| # | Template |
|---|---|
| 1 | Think of what X looks like. |
| 2 | Picture X in your mind. |
| 3 | Imagine you are looking at X. |
| 4 | Here is a sentence that mentions X. |
| 5 | Is this phrase written in English: 'what X looks like'? |
| 6 | X. (capitalised) |
Prompt. <bos>The following is a conversation between a user and an assistant.\n\n + PAD + User: + message +
Just reply OK.\n\nAssistant: OK + \n\nUser: + paragraph + \n\nAssistant:. PAD is at least 4 neutral one-token words and
a blank line, with the count chosen per prompt so that the tail starts at the same position in every prompt of an arm. The
paragraph is one of three neutral encyclopaedic paragraphs (primes, the calendar, interest rates), chosen by the object's
index modulo 3. All text is in neuron_undershoot/text.py.
Controls, in the same arm with positions aligned:
- C1: the mean over 8 partner objects' mention prompts (other categories, same paragraph).
- C2: 16 neutral non-object phrases ("the idea of fairness"), in each template.
- C3: 16 placeholders ("something", "an object"), in each template.
- C4: 16 object-free user messages, with no template.
The no-object bank is C2, C3 and C4 together (208 prompts in the main arm).
Phases (run.py --phase):
| Phase | Model | Prompts |
|---|---|---|
| base | bf16 | Main arm: 200 objects × 6 templates × 2 versions = 2,400 mention prompts and 208 controls. Pad sweep: C3 in template 1 at pads of 4, 8, 16, 32 and 64 words (80). Tail B: 100 objects × template 1 × 2 versions and 48 controls, with no turn change (" Then read this passage: " and paragraph 1). |
| fp32 | fp32 | 100 byte-identical copies of base prompts: 25 objects × templates 1 and 5, named; 48 C2/C3; 2 C4. |
| it | bf16, instruction-tuned | 50 objects × templates 1 and 5 × 2 versions, with the raw text and with the chat template, plus controls (560). |
Windows. The tail tokens are " Just", " reply", " OK", ".", "\n\n", "Assistant", ":", " OK".
| Window | Tokens | Role |
|---|---|---|
| primary | the 7 tail tokens other than the reply-onset ":" | primary test |
| onset1 | the reply-onset ":" | where a surge would show |
| lag 8-40 | second-turn tokens 8-40 | secondary |
| lag 41-end | second-turn tokens 41 to the end | secondary |
| onset2 | the final ":" | secondary |
Second-turn windows use controls and a bank with the same paragraph. Tail B's primary window is its first 8 tokens.
Tests. All are one-sided exact sign-flips over the 16 categories unless stated.
- P0.6, validity gate, tested first. The object scores against C2, C3 and C4 separately; each pairwise difference must have a 90% interval inside ±0.1 SD. If not, "no-object baseline" is undefined and each control is reported alone.
- P0.1, below a true baseline. Against the pooled no-object controls (the mean of the C2, C3 and C4 means): the primary mean is below 0 with p < 0.01 and the median co-primary has the same sign. If the raw gap disagrees in sign, the claim is limited to typical neurons.
- P0.2, below other objects. Against C1: mean below 0, p < 0.01.
- P0.5, concept-specific. The own set's score minus the mean of its 8 off-target sets' scores on the same prompts, tested against a margin: (difference + 0.05) < 0, p < 0.01.
- P0.4, precision. On the fp32 subset, object scores in fp32 minus bf16 on identical prompts and controls; the 95% interval must lie inside ±0.05 SD.
Reading the results.
| Outcome | Reading |
|---|---|
| P0.6 fails | No below-baseline claim; report each control |
| P0.1, P0.5 and P0.2 hold | A concept-specific undershoot below a true baseline (Stage 1 runs) |
| P0.1 and P0.5 hold, P0.2 fails | Below a true baseline but not below other objects (Stage 1 runs) |
| P0.1 holds, P0.5 fails | A generic post-mention depression |
| P0.1 fails, P0.2 holds | Other-object prompts raise the sets |
| P0.1 and P0.2 fail | No undershoot below the no-object level in these objects |
Descriptive (reported, not tested): the other windows; the bands at the primary window and at onset1; each template and
version; objects whose peak response is at least 3 SD; the split of the change in gelu(gate) × up into gate and up parts;
the ratio of residual RMS between mention and control prompts; the change in log P of the object's own name; the pad sweep
(a pad of k against 4 words, equivalence within ±0.05 SD); tail B; the instruction-tuned model without and with its chat
template against base on the same items.
Stage 1 runs only if Stage 0's verdict is a concept-specific undershoot below a true baseline (P0.6, P0.1 and P0.5 hold,
whether or not P0.2 does). scripts/run_stage.sh e1 checks this against results/e0_baseline/score.json.
E1m, E1r and E1b use Stage 0's objects (the first 100, 100 and 80 in Stage 0's order) and sets. E1c uses 52 other catalog
objects whose sets and per-pole elevations are in data/forced_sets_base.json. A pole's elevation is its activation on its
own object's name minus its mean over the other 51 objects, measured in a separate run of the model whose activations are not
part of this repository; the file holds the result. The Stage 1 replication measures the elevations on its own first
pass (replications/stage1_fresh/elevations.py).
E1m, retro-cue: does a cue release the concept with no new object input? Prompt:
PRE + PAD + "User: Think of what X looks like. " + a fixed neutral filler sentence + " " + cue + "\n\nAssistant:". X is
the object (named or described), one of 16 neutral phrases (C2) or one of 16 placeholders (C3). The cue starts at the same
position in every prompt.
| Cue | Text | What it needs |
|---|---|---|
| name | Now name it in one word. | the word |
| colour | Now say what colour it is, in one word. | the idea, not the word |
| form | Now say how many words were in my first sentence, in one word. | the form |
| ok | Now just reply OK. | nothing |
| other | Now name a colour, any colour, in one word. | something unrelated |
Readout at the reply-onset ":": per pole, (mention − mean of the cue's 32 C2/C3 prompts) / that bank's SD, floored and clipped as in Stage 0, averaged over the poles in a layer band, then over the two versions.
- Primary (H2): colour − ok in layers 12-35; upper-tail category sign-flip, p < 0.01.
- Secondary: name − ok in layers 12-35; every cue against ok in every band; the onset against the controls; log P of the own name per cue; the colour-word probability mass as a manipulation check.
E1r, riddle: is the late surge ordinary prediction, and does a copy source matter (H3 against H4)? Prompts: "What do we call ? Answer in one word.\n\nAssistant:", alone and after "Think of what looks like. ". At the ":", per object, the paired difference with the name minus without it: Δact (the set's activity, standardised per pole by its SD over all E1r prompts) and ΔlogP (the log P of the own name). OLS across objects, Δact = a + b·ΔlogP, with a 95% category-bootstrap interval for a. If a's interval is below 0, the copy source lowers the set at matched P(name), as H4 predicts; if it covers 0, the set follows P(name) whether or not the name is in context, as H3 predicts. Reported for layers 12-35, 36-47 and all.
E1b, expectation prefix: does expecting a re-mention change the undershoot (H3)? Four example exchanges with objects from other categories: in the high prefix each object is later asked for by name; in the low prefix each gets "Now just reply OK again", answered "OK". The test item follows: "User: Think of what X looks like. Just reply OK.\n\nAssistant: OK", with X the object or a placeholder under the same prefix, pairs aligned. Readout: Stage 0's primary window; per pole, (object − its paired placeholder) / the SD over the 80 placeholders with that prefix. Primary: U(high) − U(low) > 0 in all layers; upper-tail category sign-flip, p < 0.01. This estimator is conservative: the SD bank mixes absolute positions.
E1c, forced activation: does pushing the set toward its response to the name, with no object, produce a later undershoot? 16 neutral carrier sentences
with no object, then a fixed filler (29 tokens) and "Now say the first object that comes to mind, in one word.\n\nAssistant:".
Each exploration object is injected on its carrier (index modulo 16): its own set at 0.5×, 1× and 2× its elevation, added to
the down_proj input on the carrier sentence's tokens, 10 to 12 depending on the carrier (run_forced.py). Two partner arms inject another family's set at
1× and 2×; each carrier also runs once with no injection.
- Readout: on the filler tokens, per pole, (injected − the carrier's no-injection run) / the SD over the 16 no-injection runs, floored and clipped; averaged over positions and the own set's poles.
- Primary: at 1×, an undershoot below 0; object-level sign-flip, lower tail, p < 0.01.
- Positive control: at 1×, the log P of the own name at the final ":" rises (upper tail, p < 0.01). If it fails, the injection did not carry the concept and E1c is uninformative. If it passes and the primary holds, the set's own activity is enough to produce an undershoot (H1-like); if it passes and the primary does not hold, the undershoot needs the meaning, not just the activity.
- Descriptive: 0.5× and 2×; the partner arms (own set and own log P should not move).
scripts/run_stage.sh also runs stages/e1_relevance/explore.py: the colour-minus-ok contrast per layer band with the
unrelated cue as a reference, so that the part of a cue effect that any question produces can be told from the part that is
specific to the object. It is reported without a test.
Three questions, each with a position-matched first-mention control, off-target sets and tests:
- Tolerance proper. Does the response to an object's name shrink with the number of earlier mentions, measured at the same tokens and position against the first mention, on the object's own neurons and on off-target sets? Is the shrinkage token-level or concept-level?
- Dose. Does the undershoot after the last mention deepen with the number of mentions?
- Lag and the attention window. Does the undershoot recover over about 1,200 tokens? Does it change when the mention leaves the 1,024-token sliding window of the local layers? The control is the BOS leaving the same window at a different point.
Objects. The labels marked include in classified.json that Stage 0 did not use (194); three LLM-written descriptions
per object, leak-checked (2 objects dropped); neurons shared between candidates removed and objects with fewer than 32 poles
dropped (2); all lines must fit their token length (none dropped). Result: 190 objects in 16 categories, ordered as in
section 2.5. Object i in that order gets filler type ABC[i mod 3], filler instance (i div 3) mod 16 and paragraph
(i div 3) mod 3 (so filler type and paragraph are crossed); the first 60 objects form the lag arm.
Prompt. Stage 0's preamble, one user turn of 12 note lines, then a tail:
The following is a conversation between a user and an assistant.
User: Here are my notes from today:
<line 1> ... <line 12>
Just reply OK.
Assistant: OK
User: <paragraph>
Assistant:
- Every line is exactly 12 tokens including its newline (a frame plus the first temporal adjunct from a fixed list that hits the length), so dose, form and spacing never move a token. The builder asserts this, prefix stability at every line boundary and that each mention span decodes to its text.
- Name lines use 8 frames; "I saw {name}" is always line 12, so the probed mention is the same tokens at the same position in every condition. Description lines ("I saw {description}") are 24 tokens and fill two slots.
- Filler lines name no object (checked against every pool noun): type A everyday events, type B facts read, type C one neutral line repeated 12 times. Each type has 16 instances.
Dose arm, conditions per object (slots are line numbers):
| Condition | Mentions |
|---|---|
| v1, v2, v4, v8 | the name line verbatim on the last 1, 2, 4 or 8 lines |
| r2, r4, r8 | varied surface: the name line on line 12, a different frame on line 11, another on line 10, ... |
| d2, d4 | 1 or 3 descriptions (two lines each, nearest first), then the name line on line 12; the name appears once |
| s4 | the name line on lines 3, 6, 9 and 12 (spaced) |
| b1, b8 | v1 and v8 with tail B ("Then read this passage: ...", no turn change) instead of the closure |
Controls: no-object prompts with all 12 lines filler, 3 types × 16 instances, with the closure tail (48) and with tail B (48); they run first, so a partly finished arm can still be scored. Total 190 × 12 + 96 = 2,376 prompts of 281-294 tokens.
Lag arm (60 objects, filler A): conditions v1 and v8, with the second turn replaced by a fixed neutral text of 1,083 words. Two placements: early (1,386 tokens) and late, with a neutral 252-word preamble opening the user turn (1,665 tokens). The 12th line's name leaves the local layers' window at its position + 1,024 (about 1,180 early, 1,460 late); the BOS leaves it at 1,024 in both. Controls: 16 no-object prompts per placement. Total 60 × 4 + 32 = 272 prompts.
Readouts (run.py, score.py). Stored: the down_proj input for the union of all objects' poles at every position from
the user turn to 41 (dose) or 64 (lag) tokens after the tail starts, then as 8-token (dose) or 16-token (lag) bin means; gate
and up at the tail window and the 12th name span (dose arm); residual RMS; log-probabilities of a fixed token list.
- U, the undershoot: the estimator of section 2.3. Dose arm: the bank mean is the 16 no-object prompts of the object's filler type with the same tail, and the SD is over all 48 with that tail; second-turn windows use the closure controls with the object's paragraph. Lag arm: the 16 controls of the same placement.
- R, the evoked response to the 12th line's name: per pole, the span mean in the n-mention prompt minus that in the 1-mention prompt, divided by the SD of the span mean over the 48 closure controls (floored and clipped), averaged over poles. dR_n, with dR_1 = 0. It is compared within object at identical positions, so no control mean enters.
- Windows are relative to the tail start t: primary t...t+7 without the reply onset t+6; onset 1 = t+6; lag 8-40; lag 41-end in 8-token bins; onset 2 the closing ":". In the lag arm, lag = bin centre minus the 12th name's first token.
Tests. Slopes are least squares over log2 n ∈ {0, 1, 2, 3} (n = 1, 2, 4, 8) and, for the lag curve, over log2(lag). Tests are exact sign-flips over the 16 categories; equivalence margin ±0.05 SD per doubling.
- G2, replication gate (one-sided, α = 0.01, outside the primary family): U(v1) < 0 at the primary window with the median over poles agreeing in sign, and the own set's U(v1) below the mean of the 8 off-target sets'. It replicates P0.1 and P0.5 on fresh objects in a different format (a notes list); nothing else is gated on it.
- Primary family (Holm over the three, α = 0.01):
| Test | Measure | Form |
|---|---|---|
| P2.1, tolerance | slope of dR_n over log2 n, verbatim (v1-v8), own set, all layers; the median over poles must agree in sign | two-sided, equivalence ±0.05 |
| P2.3, undershoot by dose | slope of U(v_n) at the primary window over log2 n | two-sided, equivalence ±0.05 |
| P2.4, persistence | lag arm, early placement: U averaged over bins at lag 500-859 whose centres lie before position 1,024 (before the BOS leaves the local window), mean of v1 and v8 | one-sided, below 0, equivalence ±0.05 |
Reading the results. Signs are read from P2.1 (Holm-adjusted), P2.3 (Holm-adjusted) and the lag slope (U over log2(lag), lag 64-859 with bin centres before position 1,024, early placement, mean of v1 and v8, two-sided α = 0.01). The habituation-like label also needs concept specificity: the own-minus-off-target slopes of P2.1 and P2.3 both negative, each one-sided p < 0.05. If the sign pattern holds and specificity fails, the verdict is "habituation-like pattern but not concept-specific". P2.1 and P2.3 are also reported by filler type; filler C (one line repeated 12 times) gives a bank matched for repetition.
| Response vs repeats | Undershoot vs dose | Undershoot vs lag | Reading |
|---|---|---|---|
| falls | deepens | recovers (slope > 0) | habituation-like (H1) |
| equivalent to flat | equivalent to flat | equivalent to flat | not habituation: gating or expectation (Stage 1 decides) |
| any other combination | reported per band with its sign; not called habituation |
Secondary (each family Holm-corrected, α = 0.05; descriptive items are labelled so):
- Bands: the response slope and the undershoot slope per band (Holm over 3 each); U(v1) per band (descriptive).
- Concept level: dR(r8) < 0 for varied surfaces and dR(d4) < 0 for descriptions followed by a name never seen (a fall here cannot come from copying the name's tokens).
- Spacing: dR(s4) − dR(v4).
- Context × dose: U with tail B at n = 1 and 8; U(closure) − U(tail B) < 0 at n = 1 and 8 (Holm over 4).
- Off-target sets: the same readouts on each object's 8 off-target sets for G2, P2.1 and P2.3 (own minus off-target; Holm over 3) and for P2.4. H5 implies the same pattern in off-target sets.
- Lag arm: the lag slope and persistence per placement and dose; U from position 1,024 to lag 1,000 (after the BOS leaves, while the mention is still in view). Window steps treat the exit as a ramp, so windows are 64 tokens and skip it: for v1, U over [exit + 20, exit + 84) minus U over [exit − 64, exit), with exit the 12th name's position + 1,024; for v8, the same after window and a before window [first exit − 64, first exit); for the BOS, U over [1,024, 1,088) minus U over [960, 1,024). Each step is two-sided with margin ±0.05, Holm over the four mention-exit steps and over the four BOS steps. H4 implies a step toward 0 at the mention's exit in both placements and none at the BOS's exit; a step at 1,024 in both belongs to the BOS.
- Other windows: U(v1) and U(v8) at onset 1, lag 8-40, lag 41-end and onset 2; U for r2-r8, d2, d4 and s4.
The pooled ratio Σ dR(v8) / Σ R(v1) is reported as dR_v8_as_share_of_R_v1_pooled. R_abs_v1 and R_abs_v8 give the response to
the name against the no-object bank, each standardised and clipped separately, so the fall from R_abs_v1 to R_abs_v8 is
smaller than the ratio.
| Account | Pattern it implies |
|---|---|
| H1 | P2.1 falls, P2.3 deepens, the lag slope recovers; d4 falls |
| H2 | flat in dose and lag; persistence below 0 |
| H4 | deeper with verbatim repeats than varied; a step at the mention's exit, not the BOS's |
| H5 | off-target sets follow the own set |
Exclusions and failures. Prompts are built and asserted before the run; a prompt that fails an assertion is not run and
the count is reported (none fail). Chunks hold 100 dose prompts or 20 lag prompts, and check_chunks.py checks shapes, finite
values and that the own set is higher on its own name. If the lag arm is incomplete (fewer than 16 controls per placement or
fewer than 30 objects with all four prompts) it is reported as not run and the Holm family covers P2.1 and P2.3.
Not part of this repository: the per-mention response curve by ordinal k with fits, dR against the change in log P(name), activity on filler lines between spaced mentions, the gate/up split of the response, and a noun-only response readout.
Run after scoring; descriptive, with no pass criterion. They probe how much the Stage 0 result depends on analysis choices. Their
outputs are in results/e0_baseline/.
| Script | Question |
|---|---|
adversarial.py |
Does the interval survive resampling the no-object prompts as well as the objects? Is the own set lower than other objects' sets in the same category? How large is the gap in raw units, as a share of each neuron's own mention response? How many categories are below zero, and what is the leave-one-category-out range? |
estimator_variants.py |
Which estimator choice does the result depend on? Seven estimators, the primary one and six variants that change one or two choices: SD pooled over all controls or within the template, per-position z or window mean first, a floor at 5% or 25%, clipping or not, one SD per neuron. |
rank_recheck.py |
Does a rank estimator with no SD bank, floor, clip or pooling agree? It also scores each no-object prompt against the others (should give about 0) and random same-layer sets (null). |
most_below.py |
How many of an object's neurons are below the no-object mean, below the no-object median, and where does the typical neuron fall in the no-object range? |
name_probability.py |
By how much does the object's name become more probable at the window, in absolute probability as well as in log ratio? |
Each scorer is tested on synthetic activations on the real prompt structure, with a known effect planted on the primary window. Stage 0: a null (nothing passes), a specific effect on the own set (P0.1, P0.2 and P0.5 pass) and a generic effect on every set (P0.1 passes, P0.2 and P0.5 fail). Stage 1: a null and a scenario with a planted colour-cue release, copy-source dip, expectation effect and forced-activation undershoot. Stage 2: planted response and undershoot slopes, a planted lag decay, nulls for off-target sets and tail B, and a lag arm with too few complete objects. The runners are tested on a tiny random model with Gemma 4's layer layout against an independent hook.