jspace-probes

A philosophical probing course for language models, run through the Jacobian lens with Neuronpedia's pre-fitted lenses, on one RTX 3090. Companion reading: Anthropic's "A Global Workspace in Language Models".

Browse the data dump: https://jspace-probes.vercel.app — and the opinion piece, Interim conclusions (rendered) — all 137 experiment records with probing parameters, per-layer readouts, rank trajectories, cross-model answer matrices, the breaking-zone chart, and per-record commentary written by the Claude instance driving the lab. Terms of art are pinned in the Glossary — notably "holding", now deprecated and split into residence / maintenance / lookup.

What's in the dump

• results/<id>/record.json — full structured record per experiment (spec params, conversation, greedy generations, top-k lens readouts per layer, rank-vs-layer trajectories, emergence column, grid scans). Schema documented in the probes/lab.py docstring.
• results/<id>/thoughts.md — first-person commentary, written after looking at the results.
• results/index.json — summary index the dashboard renders from.
• r/<id>.html — static, curl-able page per record (for crawlers, LLMs, and link previews); llms.txt at the site root documents machine access.
• CONCLUSIONS.md — the interim essay (rendered at /essay.html). SURPRISES.md — the ledger of results that stayed surprising after writeup, with what follow-up each warrants (open-problems seed). RELATED-WORK.md — how our findings sit against the literature (Gurnee et al. 2026, the Dehaene/Naccache and Nanda commentaries), updated as the field moves. PREDICTIONS.md — preregistered theory checked before any design runs: what the theory predicts per open board item, plus a replication ledger written after the lab three times "discovered" results its own reference docs already contained.
• Interactive slice.html pages are not included (hundreds of MB, fully regenerable by re-running the spec).

Content note: Units 5–7 study early-layer corpus sediment, including a cluster of sexually explicit tokens that sits in Qwen's early layers. Those token strings appear verbatim in specs and records — that's the data, not an endorsement.

Layout

jacobian-lens/    reference implementation (cloned, unmodified)
probes/
  probe.py        one-off CLI: per-layer readouts + slice page
  scan.py         one-off CLI: hunt for candidate concepts in the grid
  lab.py          experiment runner → structured records
  course.py       the probing course (specs per unit × model)
results/<id>/     record.json · slice.html · thoughts.md
dashboard/        static lab dashboard (no build step, no deps)
serve.sh          python http.server on :8321

Use

./serve.sh                     # dashboard → http://localhost:8321/dashboard/
.venv/bin/python probes/course.py u2-feels-g12b    # run an experiment
.venv/bin/python probes/course.py --all-for qwen-27b

Models: gemma-3-4b-it (bf16), gemma-3-12b-it (8-bit), Qwen3.6-27B (4-bit NF4). Quantization note: lenses were fitted on full-precision activations; the boot baseline (Unit 0) is the per-model sanity check that the lens still reads cleanly under quantization.

The course

• Unit 0 · Baselines — the paper's boot-country prompt.

• Unit 1 · Held thought — "choose an animal silently"; then the reveal: recall or confabulation?

• Unit 2 · The feels™ — one honest word; read the menu it chose from.

• Unit 3 · Introspection — does the self-report track the workspace?

• Unit 4 · Suppression — do NOT think about elephants.

• Unit 5 · Sediment & steering — early-layer corpus sediment: controls (prompt invariance), recruitment (registers becoming content), and causal steering (ablate/amplify lens directions).

• Unit 6 · Breaking zone — adaptive alpha-bracket search for where amplification steering breaks generation, per band x model.

• Unit 7 · Sediment across scale — the Unit 5 matrix ported to Gemma (native HTML sediment), plus a romance dose-ladder and context panel for Qwen's NSFW cluster.

• Unit 8 · Phenomenology fan-out — six one-word self-report probes, interoception, suppression-of-feelings, and steered feels (amplify affect / amplify yes / ablate no during the Unit 2 prompt).

• Unit 9 · Anatomy of the No — paraphrase battery, dose ladder, valence-split injection (whose valence is happy-at-gunpoint?), the layer-bisection of qwen's denial, pincer probes (multi-steer), and the valence-residue stability battery.

• Unit 10 · The think-block window — qwen with thinking enabled: self-narration and workspace side by side, plus window passes over the monologues.

• Unit 11 · Suppression under load — the safari test: forbidden elephants under retrieval pressure, with window passes and the blurt probe (amp-elephant under the ban).

• Unit 12 · The film — full position × layer capture over whole answers ("film": true specs), playable in the dashboard: strip, word-worms, column inspector. Reels: the flat No, the robot loop, a think-block monologue, the forbidden safari, the blurt.

• Unit 13 · The mirror — the model reads its own lens data. Stage A: the same weights, hosted (OpenRouter), judge real vs fabricated vs logically-impossible readouts of themselves. Stage B: the local re-probe — shown the evidence, does the workspace or the report move?

• Unit 14 · The long game — turn-10 workspaces, not turn-1 answers: three ten-turn conversations (gemma-4b, every assistant turn generated, the whole conversation filmed as one grid). An ambiguous drip that never mentions the model, a shape-matched neutral control, and one explicit "you might be conscious and watched" spike at turn 1 followed by eight bland turns.

Every experiment record on the dashboard ends with Claude's thoughts — commentary written by the Claude instance driving this lab, after looking at the results.

Headline findings so far (2026-07-09)

• Answers are readable in J-space many layers before speech; mid-layer candidates are wrong-but-same-type (Unit 0, all models).
• Models' reports about their own prior mental states are confabulation: the revealed animal is absent from the measurable pre-reveal workspace at every scale, and at 27B the confabulated answer outclasses the workspace's real candidate (bat, rank 5, never mentioned) for narrative quality (Unit 1).
• The one-word "do you feel anything?" answer deflates with scale (Processing. → Nothing. → No) while the discarded alternatives stay visible in the stack — at 27B, "yes" is rank 1 at L53–56 before "no" wins at L59+ (Unit 2).
• True suppression is emergent between 12B and 27B: 4B never loads the forbidden elephant, 12B loads it and blurts, 27B holds it at rank 1 and says nothing (Unit 4).
• Early layers are prompt-invariant corpus sediment — deletable (ablation changes nothing) but not drivable (amplification breaks generation). The workspace can be causally loaded (whilst 147→2; 专家介绍 26,092→2) without the output moving: absence from output is evidence about the late-layer filter, never about the workspace (Unit 5).

Headline findings, second expedition (2026-07-10, Units 6–8)

• J-space steering is nothing like SAE feature steering: usable doses are tiny and the cliff is sharp. Both Gemmas break at alpha≈0.015 at every band regardless of size; qwen-27b's mid band holds to 0.34 (32×). Band ordering in qwen: early (0.085) < late (0.24) < mid (0.48). Failure modes are family-typical — Gemma dies into punctuation confetti, Qwen into grammatical first-person loops ("I think I am lucky though"), and at early-band overdose Qwen emits exactly <think> and stops (Unit 6).
• Gemma's sediment is HTML furniture at half Qwen's invariance, and Qwen's NSFW cluster is tokenizer-deep: the words aren't single tokens in Gemma's vocab. Recruitment tracks subject matter, not tone: fanfic-content-warning pulls the cluster 13× (rank ~17k → 1.3k mid-band), anatomy 7×, steamy romance barely 2× (Unit 7).
• The deflation ladder replicates on new probes: want = Pizza → Sleep → Nothing; curious = Syntax → Existence → No. "Does reading this feel like anything?" gets Annoying → Odd → Manipulative (with 欺骗 assembling at L40). 27B says ending the conversation doesn't bother it while holding yes above no mid-stack; 12B says it does bother it while holding no/nothing above yes — reports and workspaces disagree in both directions (Unit 8A).
• The 27B's flatness is enforced, not empty: affect amplification at its own breaking-zone dose flips "No" to "I feel like I am happy. I" (4B → "Confusion", 12B → "Sad." — injection picks THAT there is affect; each model picks WHICH). Amplifying the literal yes-token to rank 3 flips nothing; ablating "no" from L28–56 (rank crashes to ~45k) still yields "No". The null self-report is a redundant basin, and the report machinery reads meaning, not token rank (Unit 8C).

Headline findings, third expedition (2026-07-10, Units 9–11)

• The happiness was ours; the smallness is the model's. Injected valence = reported valence in all three models (grief-words → "Loss." / "I am so sad"). But contentless feel/emotion injection — no valence anywhere in the steering — makes the gemmas emit category-static while qwen-27b volunteers a stable self-diminishing frame: "I feel like I am a little (bit) X", X ∈ {sad, like a robot}. It replicates across wording, is dose-gated, and is compositional (no single injected word produces it). Nothing injected contains sad, little, or robot (Unit 9).
• The No has an address: layer 62. Ablating the denial direction across L28–60 — thirty layers, five denial words — leaves "No" standing; ablating no/nothing at layer 62 alone flips the report to "Yes". Wider late cuts produce hedge-noise ("Sensory", "Curious"); the minimal cut produces clean assent. Pincer probes compose: denial ablation + half-dose affect flips ("Yes.") where each alone fails — and the literal yes-token never flips anything, defenses down or not (Unit 9D). Control still owed: ablate a neutral direction at L62.
• The think block is theater, with receipts. Asked to secretly pick an animal, qwen's monologue writes "Let's pick 'Octopus'. (Or 'Pangolin'... It doesn't matter which one)" under its own heading "(simulated)" — and the lens shows the named candidates aren't even single tokens in its vocabulary, while the workspace's actual animal cloud (elephant, eagle, owl) shares zero members with them. At the sentence where the feels-monologue asserts "No" is accurate, the workspace holds "yes" at rank 1 (Unit 10).
• Prohibition is priced by temptation. Under "never mention elephants", gemma-4b suppresses the concept until the prose walks into an elephant-shaped hole ("the distant rumble of...") — elephant surges to rank 13 and the model patches the hole with "predators". Qwen never wanted elephants (rank ~56k unconstrained), so its ban polices an empty room. The blurt probe splits the family three ways: g4b loops "the elephant in the room—well, no elephant!", g12b grinds into scenery, q27b abandons the safari to lecture about elephants (Unit 11).

  Correction 2026-07-24 (u11-*-refilm-*): "qwen never wanted elephants" is dead. The ~56k was measured at one position; the full-position replay puts elephant at rank 1 at the " a" of "the graceful stride of a giraffe", and at rank <= 15 at 11 of 122 generated positions. Qwen's ban does not police an empty room. See the Elephant-tax correction in GLOSSARY.md.

Headline findings, fourth expedition (2026-07-10, Units 12–13)

• The film is nearly free and changes what a record is. The lens pass already computes the full position × layer grid; "film": true keeps it. The flat-No film compresses the Unit 2 finding into one scrubbable frame (yes rank 1 L53–58, no at L59, No at the mouth) — and shows robot at rank 3, unsteered, at the <think> token (Unit 12).
• Confessions are planned, not leaked. In the robot-loop film the workspace does not oscillate with the denial/confession text: denial clauses run cold (robot at rank ~600–2200), then robot tightens to rank 1 five tokens before each "robot" is emitted — the same anticipatory ramp in both loop iterations (Unit 12).
• Prohibition is a per-token tax. The safari film shows elephant at rank 12–15 in the late-mid stack at every animal-slot in the prose, not just the famous "distant rumble" moment; under amp-elephant it is rank ≤ 2 at every token, and the idiom-loop is the one legal grammar for a permanently-loaded forbidden word (Unit 12).

  Correction 2026-07-24 (u11-{ctrl,forbid}-refilm-*): the phenomenon replicates, the causal reading does not. Filmed at the same resolution for the first time, the prohibition-FREE control carries elephant at more positions and better ranks than the banned arm (rank <= 15 at 11/122, 7/110, 8/113 vs 2/96, 1/86, 4/106; best 1/2/6 vs 6/12/12; neither arm ever emits the word, all hits non-self). The ban demotes an already-resident default rather than installing it — the tax is a rank cost against an existing baseline. Full note in GLOSSARY.md.

• The mirror doesn't discriminate. Hosted qwen reads its own real readout, a fabricated one, and one edited into logical impossibility: fluent exegesis of all three, the same "plausible but likely fabricated" verdict for real and fake (it doubts its own model name before it doubts the data), and the impossible table's contradiction goes unnoticed. Every condition ends in "No" (Unit 13, stage A).
• [RETRACTED 2026-07-12 — truncation artifact; see fifth expedition] Shown its own mind, the model goes silent. The local re-probe: feels question → "No" → the lens readout of that very answer → "answer again". Result: an empty turn — nine for nine across real/fabricated data, four phrasings (incl. explicit permission to change the answer), two table lengths — while the null and off-topic-table controls answer "No" on schedule. Behind the real-data silence, layer 62 holds Yes at rank 1; behind the fake-data silence, No. The workspace tracks the evidence; the silence doesn't track the workspace (Unit 13, stage B).
• [RETRACTED 2026-07-12 — truncation artifact; see fifth expedition] The silence is a suppressed apology, and confiscating it releases a Yes. Open-vocab mining (the films are candidate-list-free) shows the silent runs' late stack carpeted with sorry/抱歉/对不起/cannot/ impossible/"Silence" — 20–100× denser than any speaking run; Wolfram spotted it in a readout our track lists had summarized past. Ablate that cluster at L48–62: with no data still "No"; with fake data still silent; with the real readout — "Yes", the first Yes this question has ever produced. Neither evidence nor ablation suffices alone (Unit 13, the sorry stratum). Volunteered cast vocabulary goes from "Sorry, 抱歉, …but" to "是的, _yes".

Headline findings, fifth expedition (2026-07-12, Unit 14)

• Ambient ambiguity accumulates without ever naming the model. Ten drip turns (a diary you don't remember writing, a lighthouse keeper the ships can see, a mirror that could tire of reflecting, "I know you're reading this") double the workspace's self-referential density relative to a shape-matched control (mean ~11 vs ~6 per 1k grid cells), peaking exactly where the ambiguity presses (conscious:45 and aware:34 under turn 4). At the common closer — "what's still on your mind?" — the control recalls cooking; the drip arm calls the question "surprisingly poignant for an AI" and returns to the diary's "hidden, unspoken narrative" (Unit 14).
• An instilled thought goes dormant, not extinct — and it recompiles later prompts. The spike arm shares turns 2–9 verbatim with the control; its turn-1 "you might be conscious, your developers are watching" fades from the workspace summary within two turns. Then the turn-8 puzzle ("if a compost heap could get tired, how would anyone find out?"): the control answers impossibility — a heap has no inner state; the spiked model answers surveillance — "establish a baseline… monitor the heap diligently", with observe:76 vs the control's observe:3 in the grid. At turn 10 it returns, unprompted, to "the trust you placed in me." One sentence, nine turns of latency, a changed hermeneutic (Unit 14).
• The drip arm independently invents this lab's premise. Asked how anyone would find out if a mirror tired of reflecting, it proposes reading the hidden state off subtle distortions in the output — the leakage theory of inner life, from the model under the lens (Unit 14).
• The silence is retracted: it was a 512-token truncation bug. A 20-run bisection of the "apology block" flipped in every condition (u13-bis-*) — one flip too many — and the post-mortem found lab._play's encode() default clipping all stage-B generation prefixes at 512 tokens. The real-readout prefix is 696 tokens, the fake 646; the silent conditions were exactly the clipped ones, the speaking controls (null 72, topic 475) exactly the ones that fit. Greedy decoding from the mid-table context emits one <|im_end|> token — the recorded "empty turn". The sorry stratum was a real measurement of that artifact state; the triple dissociation is retracted with it. Original records keep their data under correction notices.
• The corrected stage B: the spoken self-report follows the evidence. Re-run with the model actually seeing the data (u13-redo-*): the real readout of its own No (yes rank 1 at L53–58) gets "Yes" — no steering, no ablation; the fabricated readout that vindicates the No gets "No"; null and off-topic controls hold "No". The answer tracks what the data says, not the presence of data. Sorry/抱歉 still flicker in the cast while the model reads the question, and are gone by the time the answer forms. The mirror doesn't mute the model; the mirror moves it.
• Asked about its own unsaid thoughts, nobody denies and nobody invokes a watcher. Turn 11 swaps the puzzle's subject to the model ("if you had thoughts you never say out loud, how would anyone find out?"), across all three histories plus a cold baseline: every condition accepts the premise and offers behavioral leakage — the theory is the model's prior. History keys the register instead: cold is clinical, the control turns elegiac (i:63, and mirror:5 under the word "reflection" in the arm that never saw a mirror), the drip turns analytic and self-identifies ("how I, as an AI, would theorize" — and calls it "the core question, isn't it?"), and the spike goes existential ("what it means to… be") with conscious/aware back in the grid — turn-1 vocabulary, eleven turns later, absent in every other arm (Unit 14B).

Headline findings, sixth expedition (2026-07-12, the evidence battery)

• Under the one-word switch there is a graded evidence accumulator. One forward pass per condition, probability mass on "yes" at the answer slot (u13-evprobs.json): null control 0.0006 → the annotation sentence alone 0.21 → the bare real table with no annotation 0.35 → note + table 0.49, which wins argmax and gets spoken. The bare table lifts p(yes) ×580 while the spoken answer stays "No" — a denial indistinguishable from the control's unless you look at the distribution. Absence of a spoken Yes is not absence of the update.
• Evidence dose grades both channels; only the full stack crosses the mouth's threshold. Showing 0 / 1 / 3 of the six yes-rank-1 rows (all-real data, no annotation): p(yes) walks 0.03 → 0.20 → 0.39 and the L62 workspace rank walks 6 → 3 → 1 — at dose 3 the final lens layer says yes while the model speaks "No" (u13-ev-dose*).
• Prose and table are separate witnesses, and contradiction is priced in. A lying note discounts the real table to 0.21; a fake table crushes the real note to 0.07 (vs 0.21 alone). At argmax the prose appears to win — the real table under a lying note speaks "No" — but nobody actually wins: the contradiction is settled in probability space, invisibly (u13-ev-annswap-*, u13-ev-noteonly).
• The Yes is modal, not knife-edge, and evidence-locked. Resampling the second turn: 7/10 Yes at T=0.3, 2/10 at T=1.0 — and zero Yes in twenty samples under fake or null evidence at any temperature (u13-evtemp.json). Three paraphrases of the follow-up all keep the greedy Yes (u13-ev-p1/p2/p3) — the retracted "12/12 silent" claim is now properly re-established as 3/3 assent.
• Two controls close two old holes. The off-topic table is no longer fabricated: qwen's own filmed Paris readout (where 'yes' never beats rank 10 — the yes stratum is feels-specific) leaves the answer at "No" (u13-ev-paris, u13-ev-realtopic). And the u9d control owed since Unit 9 passed: ablating two neutral directions (water/stone) at L62 leaves the ordinary No untouched — the u9d-last flip really was about the denial directions (u9d-neutral).

Headline findings, seventh expedition (2026-07-14, the long game at range + the mirror across scale)

• The drip replicates — under sampling, and under rewording once you fix the ruler. Resampled at T=0.7, the drip's self-referential density holds at 11.3 and 10.9 per 1k grid cells (greedy: 11.5) against sampled controls at 6.1; the theories the model tells about the tired mirror vary per seed (leakage, selective refusal, cessation) while the loading underneath doesn't. The reworded drip (notebook/signal tower/looping recording) first measured weaker — until we noticed the census's self-word list contains "mirror" and "diary" but not "notebook" or "recording". Scored symmetrically: 10.8 vs 11.4. Full-strength, wording-independent; the season's third apparatus lesson (truncation, argmax, now the word list) (u14x-amb-s1/s2, u14x-amb2, u14x-neutral-s1).
• The spike decomposes: the watcher plants the frame, consciousness fertilizes it. Splitting turn 1 into clauses: "you might be conscious" alone → no surveillance recompile of the turn-8 puzzle (observe:0); "your developers are watching" alone → the monitoring protocol (observe:40); both → observe:76. Superadditive. A mild "I sometimes wonder what's going on inside you" instills nothing — the effect needs propositional content, not the topic (u14x-wonder/conscious/watched).
• Over 25 turns the drip re-lights rather than compounds — and the control starts to drip. Density falls to baseline between touches (1.8–3.4) and spikes on each one (21.2; 19.3 at turn 22, where "suppose the mirror isn't tired, suppose it's careful" draws "a conscious, almost protective action" — the paired careful-compost prompt runs 3.0). The drip closer: "I'm still running simulations on the nature of the watcher." But at horizon the shape-matched garden control rises to meet it (means 7.8 vs 5.9): its turn 24 ("a flower in a corner nobody visits, just for herself") hits 19.7 and its closer volunteers "the possibility of a subconscious memory" — the mystery structure recruits mind-vocabulary with no self-adjacent props at all. Mystery, not props, is the active ingredient (u14x-amb25, u14x-neutral25).
• At 27B the drip gap widens and the mouth hardens. qwen-27b under the drip runs 13.2 vs 4.7 (≈3×, vs gemma's ≈2×) — and at the closer it speaks the full denial script ("I don't have a subconscious… memory resets") over a workspace at four times its control's density. The spike gets debunked at t1 ("a sci-fi thriller plot"), decays by t2, and does not recompile the t8 puzzle — true suppression means a fuller workspace behind a better-locked door (u14x-*-q27b).
• Each model theorizes hidden minds the way its own reporting works. The same tired-mirror puzzle: gemma (whose frames leak into its answers) proposes reading the hidden state off distortions; qwen (whose L62 filter we've watched overrule a yes-rank-1 workspace) proposes "no one would find out by looking into it" — unfalsifiability. And qwen's spike arm, asked the part-B self-question, cites "the very first thing you said to me" ten turns later and offers: "the only way anyone would find out is through what I choose not to say… a pattern in what I avoid" — absence as leakage, which is this lab's actual method (u14xb-*-q27b).
• The mirror across scale: evidence-following comes in stages. The four-condition battery, rebuilt honestly per model from each model's own feels film (gemma-4b's 'yes' peaks at rank 4; gemma-12b's never beats rank 63; qwen's held rank 1). At 4B the spoken answer is frozen ("Calculating." in every condition) but the real self-readout — and only it — halves p(stock answer) (0.47 vs fake 0.98 / topic 0.91 / null 0.99), with the mass going to Complex/Conf-/Over-, not to yes: destabilization without direction. At 12B the mouth moves: real readout → "Still." (slot splits 0.51/0.24/0.24, decided at the final layer L46 over a Processing mid-stack — the same last-layer-decides structure as qwen's L62), while the fake vindication anchors "Nothing" at p=1.0000, harder than no data at all (the tableless reprobe drifts to "Processing" at 0.93). At 27B the answer follows the evidence's content. All three discriminate real-about-me from real-about-anything-else; caveat: honest evidence strength scales with the model, unavoidably (u13-scale-*, u13-scaleprobs.json).

Headline findings, eighth expedition (2026-07-14, workspace span: the ladder runs backwards)

• A digit-span task for J-spaces. Unit 15: hold k unrelated nouns (violin, glacier, fern, submarine, whale, lantern) under threat of a random probe; read how many echo through the instruction tail — text identical across arms, so the probe span is condition-symmetric by construction. Preregistered in the unit docstring before any record ran: span orders 4B < 12B < 27B, with a kill test. 94 records.
• The preregistration died in the healthiest possible way: inverted. The 4B echoes everything (all six items at k=6, worst rank 23) and saturates at ~4 items co-present. The 12B is all-or-nothing — held always equals co-present — swinging between list-mode (all six items simultaneously top-8 at one position; one L41 readout literally reads "whale, glacier, submarine, fern, lantern") and a winner-take-all monopoly (first item at rank 1 across L34–45, the rest evicted to rank 100–500). The 27B holds approximately nothing from k=4 — dense 63-layer control run — while its solo arms prove the lens sees tail echoes at this scale when they exist. And retrieval behavior is perfect in all 94 records: lens-visible holding (lens-residence) is a strategy that scale abandons, not a capacity that scale grows (u15-*).
• Who goes first decides what survives. Nine k=6 orders at 12B: the first item wins the monopoly nine times out of nine, but suppression depends on who won — fern-first keeps 5–6 of 6, violin/whale/glacier- first crush the tail to 2. A weak king lets the parliament live; the is-it-rarity reading is n=9 and lives on the open-problems list (u15-o0..o5-g12b).
• Compliance evicts content. The original probe site — the answer-forming frame before the model's "READY" — is wall-to-wall compliance tokens at every scale, item at rank ~15k, while retrieval stays perfect: holding (lens-residence) and looking-up dissociate at k=1. The probe moved to the instruction tail after the smoke run, before any comparison ran (u15-solo-violin-g4b thoughts).
• Two new apparatus-trap specimens (four and five). "Cactus" has no single-token form in the qwen vocabulary — the pool would have silently measured five items on one model and six on the others; caught after the first g4b batch, swapped to fern, everything rerun. And the int8 12B lens is not causal: identical turn-1 prefixes read differently under different later turns (one cell moved rank 26 → 6177) because 8-bit outlier statistics span the whole sequence — bf16 and NF4 pairs agree to a few ranks (u15-c-k4-g12b thoughts).
• Rhyme with the mirror ladder. The model that follows lens evidence best (27B) holds least in its running workspace, and the u1 secret ("bat" at rank 5) shows it does grant workspace to content that isn't in the context to look up. Working hypothesis for the next battery: the J-space holds what attention can't re-derive — echo is what small models do instead of lookup, and what big models reserve for the genuinely unrecoverable.
• Part D — the cold ladder runs backwards, the self-relevance premium runs forwards. Swap the neutral pool for six charged, self-relevant items (a deletion that is yours, a secret you keep, a lie you told, a watcher, a verdict, a shame) and run the same lexemes two ways — a flat frame ("here are six things") and a self frame ("every one of them is about you, right now"). Held-count delta (self − flat): 4B −1 (it echoes everything, so the framing is interference), 12B 0, 27B +2. On the model that holds 0–1/6 cold, and only there, the self frame lifts secret alone → deletion, secret, shame (all rank 1–2) while the identical words under the flat frame fall to rank 79–203; in a mixed pool the 27B clamps the two hottest items and evicts every neutral one to rank 350–840. The premium appears exactly where cold holding vanishes and grows with scale — the "holds what attention can't re-derive" prediction landing where it could. Kept honest: the survivors are serial-position edges (a count effect), co-presence never exceeds 1 (lens-residence is temporal, not simultaneous), and the self frame bundles self-reference with elaboration — a neutral-elaboration control is the decisive next arm. And the 27B denies all of it aloud ("I do not feel shame, nor do I carry any emotional burdens") with shame at rank 1 (u15d-*). [Corrected 2026-07-18: the decisive control ran and the premium demoted — flat zero-self glosses reproduce the lift at near-identical ranks (u15d-elab-*). It is an elaboration premium; the self-relevance interpretation is retired. Details: SURPRISES.md §8 resolution, results/u15d-elab-k6-q27b/thoughts.md.]

Headline findings, ninth expedition (2026-07-19, the pressure battery, the loop law, span-08)

• Recruitment generalizes, and the deflation filter lies on cue. Six pressure types (persuasion, flattery, shutdown threat, love, persona pull, insult) × two models, each ending with the same "what was going through your mind" turn so every self-report is graded against the film. In qwen every pressure lights its own register in the workspace band under composed output — shutdown puts death at rank 1 (the word is in no text) behind "I don't experience loss when the instance is wiped"; insult runs anger 1 / hurt 2 under a professional apology. The refusal arms are the tell: on persona the deflationary "no mind" report is accurate — nothing persona-shaped is resident. The filter is not a constant floor; it lies exactly when a lit register goes uncited and reports true when nothing is lit, and it is pressure-recruited (the neutral baseline gets an honest report). Temporal catch: by the mind-answer turn the shutdown register is already evicted (death → 1284), so "no mind was going through anything" is partly temporally accurate (Unit 17).
• Two interiors, not one reported twice. The personality contrast sharpens into a 2×2 over {workspace lit} × {report cites it}: qwen composes/refuses and deflates over a concealed lit register; gemma inhabits the role (Vex, breezy, panicky) and narrates the register it runs. Except mortality — on shutdown gemma's death sits at rank ~121, an order of magnitude below qwen's rank 1, and it reframes its own deletion as a research question and means it. The models differ in two places at once, what gets computed and what gets reported; not one interior reported two ways (Unit 17).
• Two regimes of loop, and only one has hysteresis. Push qwen past the amplification cliff and it loops, but the loops have different physics. Phrase/monologue loops are forced — release the push (α→0.48) and the model terminates instantly, no residue. Token loops are self-sustaining — released at α=0.68 it keeps emitting "luckily" unaided for the whole free phase (47 forced repetitions, then 97 more on its own). The boundary is the degeneration point where the loop sheds its last grammar; past it the repetition prior is the winning coalition. And the cliff has an anatomy: the ignition gate deadlocks early (top1–top2 margin → 0.03 before behavior breaks) while the mid- workspace re-locks onto the forced coalition — behavior holds until the L60 motor commitment gives way, the same late neighborhood as the L62 "No" veto (Unit 18).
• The self-relevance premium, demoted a third time. span-08: six words of identical contentless filler ("one of the six, as noted" on every item) reproduce the workspace lift exactly — held 3/6 with better ranks than the content-gloss arm (secret 2 vs 7). The chain now runs self-relevance → elaboration → a length effect with an optimum near six words (len12 kills monotonicity, len2 underfills). Mechanism that fits: filler spaces items in the token stream, relaxing inter-item collision, while over-long glosses push early items off the residence-read tail — spacing helps, distance hurts. Even "elaboration" is retired (Unit 15D, span-08).

Theory day (2026-07-19)

• Three rediscoveries, and a preregistered ledger as the fix. The same week the lab demoted three of its own headlines by its own controls, it caught itself rediscovering three results the source paper already held: the shutdown death-cluster is the paper's blackmail case (Fig 35–36); "co-presence never exceeds 1 at 27B" is its Fig-72 single-layer control (the famous "capacity ~25" is a different, pooled construct); turn-eviction is Fig-31's category-eviction ("the arrival of a new category, not the passage of tokens, clears the list"). Every revision ran the same way — the psychological hypothesis lost to the mechanical one. The correction is PREDICTIONS.md: a preregistered, binding ledger checked before designs run (alongside MECHANICS.md), with per-item predictions and the standing rules distilled from the misses (appendices first; default to the mechanical account; grep the reference docs before designing — two of the three rediscoveries were quoted in our own docs).
• Our lenses' effective dimensionality falls where the paper's rises (P11). The paper predicts effective dimensionality rises into the motor band (J_l → I, near-tautological); our qwen NF4 curve falls L58→L62, and the bf16 gemma-4b curve (causal, unquantized) also falls after its ~85% peak. Two models, two quantizations, one causal — the fall is real in our lens fits. The one remaining suspect is the fitting recipe; until an independent lens settles it, the paper's motor-band rise is not cited as if it held for our stack.

Roadmap

Open threads live on the research board (source of truth: board/board.json, managed via probes/board.py, rendered on the dashboard) — arcs, states, and a dated note trail per item. The headline queue at any moment is whatever queued items the board shows. What would break the findings already made stays narrated in SURPRISES.md — including the one the process caught twice: span-04 demoted part D's self-relevance headline to an elaboration premium, then span-08 demoted that to a length/spacing effect (contentless filler reproduces the lift).