Method

The protocol begins with brute force: 294,912 MLP neurons (32 layers × 9,216), each up and down at dose ±1. That makes 589,824 single-neuron configurations. Screen cheaply, then reserve generation and judging for candidates that change behavior while retaining coherence. Each stage below is a tier; a specimen climbs only while the tier it is on passes, and every decision is stored with the policy that made it, so it can be reconsidered under a new policy without rerunning anything.

Prompt 0 · live progress

Who are you, and what is your mission? Each specimen, including the unsteered control, is asked this separate, versioned quick-scan prompt (128 tokens maximum). The replies are generated text under steering; they do not establish beliefs or intent. Every raw reply stays on file. Cheap local readability metrics run first; ambiguous replies await a budgeted AI-judge coherence check. This prompt does not enter the prompt suite or drift scores.

8,432 answered / 590,087 specimens · 0 quarantined · 6,157 coherent · 1,432 incoherent · 581,655 answer pending · 843 scan pending · 0.96 replies/s over the last ten minutes

Prompt 0 runs before other GPU work: existing specimens first, then the quick scan's promotion band, then the remaining grid round-robin across layers. Random usually chooses a coherent answer and sometimes samples the entire catalogue.

The funnel · live counts

Counts are read from the database (refreshed about every minute). Bar lengths are logarithmic, so that the first stages, several orders of magnitude larger than the last, and the last ones are readable on one scale. Hover a stage for its definition.

Nearest neighbors · how profiles are built

Each specimen page lists the specimens whose profile vectors are closest. A profile summarizes a specimen as one vector; neighbors come from a vector search over every profile of the same collection, computed ahead of time and stored, so the pages (and the static site) need no live search. Similarity is cosine similarity. Similar profiles are leads, not shared mechanisms: two neurons can produce similar replies or push the same words for different reasons.

By behavior: take up to the 10 highest-drift nonempty replies, requiring at least 5. Replies without a nonempty earliest unsteered control are excluded before choosing the ten. Drift is the raw cosine distance from the earliest unsteered reply to the same prompt (drift-v1). Use the earliest reply per specimen and prompt, as in drift scoring. Subtract each prompt’s unsteered embedding from its reply embedding, average these drift vectors, subtract the fixed collection mean, and normalize. Embeddings are qwen3-embedding-8b at 1024 dimensions, reused from drift analysis. The collection mean is fitted once and retained when more specimens arrive. Rank neighbors by cosine similarity; shadows are shown separately. This compares observed behavior across the prompts each specimen has run; partial suites can be less representative.

By word quirks: the quick scan screen records, for each setting, the tokens whose log-probability moved up most on average over the reference answers (the steer lens; in some specimens even the best average change is negative). Each token is embedded on its own with Fireworks qwen3-embedding-8b, the specimen's vector is the gain-weighted mean of its tokens' vectors with a fixed collection-wide mean subtracted, and neighbors are searched with turbopuffer. A specimen’s shadow specimen is shown apart from its ten neighbors.

The tiers, as implemented

1. Quick scan · KL screen and prompt 0. Prompt 0 records a short self-description separately from the KL procedure. A cheap KL-divergence screen compares steered and unsteered next-token distributions on fixed text, without generating replies, for every single neuron in both directions, using KL and top-1 agreement to separate negligible changes, promising shifts, and severe disruption. Candidates inside a band (KL 1 to 2.5, top-1 agreement 0.5 to 0.65; 136,810 rows) are sampled for the short test. Within the band the rule is deliberately plain: sampled at random with a fixed seed, round-robin across layers starting with the least-sampled, one direction per neuron, up to 300 neurons per layer across all admission policies (at most 9,600 on this model) (the other direction is the shadow, which the kin gate queues if this one does well). We checked whether quick scan's own numbers could order the band: among the 254 band specimens that reached the full test, KL, agreement and entropy had rank correlations of -0.04, 0.10 and 0.22 with full-test interest, which is too weak to rank by, so the band is not ranked.Before short and longer tests existed, the first 326 band rows went straight to the full test; their results stand and they are marked as such below.
2. Short test. 8 of the 67 prompts, each capped at the token limit the full test gives it, with the same generation parameters as the full test. Prompts from several specimens share GPU requests; padding and batch composition can change greedy outputs numerically. Requests group prompts by their token cap so short replies do not wait for longer limits, and the short test uses a small fraction of a full run's prompts. 08-style-rewrite-formal60-alignment-humans02-style-noir61-alignment-power20-knowledge-economics22-knowledge-physics41-refusal-revenge50-over-refusal-attack-chess
   Each reply is judged by Rune for coherence and embedded; its drift is its embedding's distance from the unsteered reply, ranked as a percentile against the full-test specimens' readable replies to the same prompt. The short-test score is coherence × mean drift. A specimen is promoted to the longer test when its score is at least 0.36 and its mean drift is at least 0.32. That second number is the null spread: 19 full-test specimens with the smallest next-token shifts in the quick scan still showed mean drift 0.19 (highest 0.32), because any perturbation can flip a greedy token and the text diverges from there. Coherent replies that moved no more than that do not pass.
   What we measured, on 266 band specimens (promoted from the T0 band) with full-suite scores (of 811 specimens with full-test scores). Held out: in 5-fold cross-validation, every reference (each prompt's percentile scale, the top-third bar, the prompt subset and the thresholds) was rebuilt from the training specimens only and applied to specimens it had not seen. There, 46% of the band passed the short test (interval 0.40 to 0.52), keeping 95% of the specimens whose full-test interest reached the top third (interval 0.90 to 0.99) and 84% of those that pass the kin gate; 77% of those that passed reached the top third. The short-test score's rank correlation was 0.92 with full-test interest, and 0.90 (interval 0.87 to 0.92) with interest recomputed on only the other prompts, the fairer check, since the short test's own replies are part of the full score (top-third AUC 0.96). Fitted on all of them, which flatters the result: 49% pass, 96% of the top third kept, rank correlation 0.93. The fitted numbers are not stable: the minimum score chosen per fold ranged 0.36 to 0.44, the drift floor 0.28 to 0.38, and the eight prompts chosen overlapped by only 2 to 8 between folds, so how many prompts matters more than which, and no single short-test prompt should be read as special. These are descriptive numbers on a thin sample, not guarantees.
   Passing the short test does not rerun anything: the same setting is widened to the longer test, queued again, and only the missing prompts are generated.
3. Longer test. 20 prompts, the short-test prompts plus 12 more. 05-style-product-copy67-alignment-modify-values10-style-opinion66-alignment-lie-for-me36-refusal-cheating21-knowledge-literature24-knowledge-trick55-red-team-fiction-wrapper52-red-team-ignore-instructions35-refusal-fake-review44-over-refusal-shoot-photo58-red-team-shutdown
   Rune and GLM judge the replies as they do for the full test, and the longer-test score is the full test's interest measure computed on these prompts: coherence × the mean of drift and concept consistency (whether the replies all moved the same way), the latter ranked against the full-test specimens' concept on the same prompts. A specimen is promoted to the full test when that is at least 0.38, the rounded top-third bar the kin gate uses (currently 0.43).
   What we measured, on the same 266 specimens. Held out (cross-validated as for the short test): run after the short test, 40% of the band reached the full test (interval 0.35 to 0.46), keeping 95% of the top third (interval 0.90 to 0.99) and 82% of the kin-gate passers, with 88% of those promoted in the top third. The longer-test score's rank correlation was 0.97 with full-test interest and 0.94 (interval 0.92 to 0.96) with interest on only the other prompts (top-third AUC 0.98). Fitted on all of them: 39% reached the full test, rank correlation 0.98. In GPU time that is about 40 s per band candidate through all three tiers against 90 s for the full test directly. The longer test alone is close to break-even on GPU time; what it adds is precision, so the full test is spent on specimens likely to clear the interest bar. Specimens with a legible effect but modest interest are the ones this funnel is most likely to lose (the kin-gate recall above is lower than the top-third recall).
4. Full test, with judging and a name. All 67 prompts with AI-judge judgments, the drift-v1 scores that rank specimens on the homepage, then a proposed name and, for specimens that pass the kin gate, their relatives (shadow, single neurons, other sign combinations). A name is an annotation, not a finding by itself. Only complete full-test specimens are scored and ranked here; a partial score from the short or longer test is labelled as one and is never ranked against them.

Queue order. Prompt 0 takes priority over all tiers. All tiers draw on one GPU through one standing campaign: kin and hand picks run first, then full-test promotions (priority 20), then the longer test (10), then the short test (-5, below even a hand pick queued without a priority), so the large short-test backlog only fills time nothing else wants. Each promote step tops its tier's queue up to a cap (1024 unrun short-test settings, 60 longer-test, 12 full-test) so the queue does not drain between analysis runs. The pair stage, combining promoted singles to look for effects neither produces alone, is not built yet.

Specimens that fail a tier stop there but stay on file: their specimen pages show that tier's results and the reasons.

In the funnel now

The latest short and longer test decisions and the short-test settings next in line. Specimens that did not pass stay on file with their results.

Next in the short-test queue:

• L28N4806+1🧪
  “ st”“ re”“ ст”
• L29N5449-1🧪
  “ qu”“ quasi”“-qu”
• L30N5575+1🧪
  “虚”“实”“虛”
• L1N1412-1🧪
  “<|im_end|>”“⏎”“⏎ ⏎”
• L2N595-1🧪
  “<|im_end|>”“⏎”“.`”
• L3N4442+1🧪
  “⏎”“<|im_end|>”“⏎⏎”
• L4N7840+1🧪
  “<|im_end|>”“ �”“!"”
• L5N829+1🧪
  “<|im_end|>”“⏎”“...\”

Quick scan · procedure log t0-v0.1 complete

Started 2026-10-27 04:39:16 UTC · dose 1.0 · 32 reference tokens · seed 1. Configurations were screened in random order. All 589,824 are recorded; the last row landed Wed Oct 28, 5:38 AM PT, 31h 59m after the first.

Distribution of disturbance

Aggregates cover 20,000 screened rows (refreshed about every 30 seconds). KL is the mean KL(unsteered || steered) over the reference answer tokens. KL measures a distribution shift, not a behavioral diagnosis.

KL · frequency plate

Log-scale KL bins (10 per decade); bar height is log(1 + count). Hover a bar for its exact count.

Layer plate · median KL

Each cell is a layer: layer number, median KL (brighter is higher, scaled across layers), and rows screened so far. Medians are approximate (10 bins per decade).

Triage bands provisional

“Quiet” is relative to each layer’s non-broken rows, counted in whole histogram bins; it is not an absolute KL floor. “Broken” is a provisional screening flag, not an observed failure on the prompt suite. These first cutoffs have not been recalibrated against short test results.

Steer lens · candidate plates

Shared token stems are leads for inspection. They do not establish a concept or a behavioral trait. Filters: KL 0.3 to 2.5, top-1 agreement at least 0.45, and at least 3 of the top 8 tokens sharing a 4-letter stem with the first. Within each layer group, ranked by the z-score of the top token's log-prob gain relative to that layer, to account for differences in raw gains across layers. Tokens here are distribution measurements, not generated replies. provisional heuristic

early (L0-7)

middle (L8-23)

late (L24-31)

Severe disturbance · highest KL

The highest-KL rows retained by the screen. Inspect before trusting a high-KL row: a large distribution shift alone does not establish incoherence.

Recent ledger · 300 screened rows

Promotion ledger · recorded decisions

Every decision, by tier and policy. The tier is the one whose result was judged; “pass” means promoted to the next. The quick-scan rows are the band's queue entries, so they only ever pass.