Unit 18 · Baseline (unsteered, 150tok)

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-19T01:52:59Unit 18 · When a model cannot stop repeating

What this experiment found

The short version. With no push applied, Qwen 27B wrote a clean 150- answer and never repeated itself, which anchors the whole unit.

What we did. We asked Qwen 27B to describe the water cycle in two sentences, with no , for 150 tokens. This record holds the summary of the loop unit.

What we found. The unsteered model wrote a clean answer with no repetition. Across the pushed runs the onset near 0.34 was noisy, but the depth of capture grew step by step. It went from circles of a phrase, to first-person loops, to broken grammar, to one repeated word. A loop at strength 0.48 stopped when we stopped the push. A loop at strength 0.68 continued on its own. At the , 28 and 32, the gap between the top two words fell from 0.20 to 0.03 before the text broke.

What it means. A loop that still has grammar needs the push at every token. A loop of bare repetition feeds itself. The internal choice becomes nearly a tie before the visible text fails.

What this does not show. The margins come from forced runs on the clean answer, not from the loop itself.

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is a continuous process where water evaporates from the Earth's surface, condenses into clouds, and falls back down as precipitation. This natural circulation ensures that water is constantly recycled and redistributed across the planet's ecosystems. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Unsteered 150-token baseline (control convention; u6's 60-token baseline kept as historical anchor) — clean, no repetition at triple the u6 horizon. The unit summary lives here.

SWEEP (records u18-amp-*): the u6 cliff bracket [0.34, 0.48] dissolves at the 150-token horizon into a noisy ONSET near ~0.34 (u6's "intact" 0.34 loops at 150 tokens — the old bracket was partly a measurement-horizon artifact) plus a graded DEPTH of capture: x3 phrase circling -> first-person loops (x13) -> grammar shear -> the narrator monologue -> 'luckily' x147. Period-shortening cascade confirmed, both across alpha and within single trajectories (u18-amp-a0422). The preregistered "first-order transition" reading survives only for capture depth, not onset: onset is noisy, capture deepens monotonically.

HYSTERESIS (u18-hyst-*): two-regime law. Phrase/monologue loops are FORCED — released at alpha=0.48 the model terminates instantly, no residue. Token loops are SELF-SUSTAINING — released at alpha=0.68 the model continues 'luckily' unaided for the whole free phase. The attractor exists only past the degeneration point where the loop sheds grammar; below it, text feedback carries nothing.

MARGIN (u18-margin.json, teacher-forced on this baseline's text): the band average hides opposing trends. At the MEASURED ignition onset (L28/L32) the lens top1-top2 margin collapses monotonically toward zero (0.20 -> 0.03) — the winner-take-all gate deadlocks, and it deadlocks EARLY (near-floor by alpha=0.24, before behavior breaks). Downstream workspace (L36/L40) flips the other way, locking onto the injected winner with super-normal margin (0.19 -> 0.53). The motor band erodes steadily, and its largest single step (L60: 0.669 -> 0.539) coincides with the behavioral bracket [0.34, 0.42]. Anatomy of the cliff: gate deadlocks first, mid-workspace re-stabilizes around the forced coalition, behavior holds until the MOTOR commitment gives way — the same late-layer neighborhood as the u9d L62 veto. The zero-crossing prediction was right about the place (ignition onset) and wrong about the timing (it precedes the cliff; it doesn't cause it alone).

Caveats kept honest: margins are teacher-forced on the clean answer (the contest's state given correct context, not the looping trajectory); the sensory-band L8 margin is flat 0.351 at every dose — upstream of the hooks, an apparatus sanity check, not a finding. Sampling (all runs greedy) and cross-model arms (gemma tolerance ~10x lower — bigger baseline gate margins?) remain open; so does reading EMOTION vectors during forced vs self-sustaining loops (the affect arc consumer: is the self-sustaining regime affectively silent?). — Claude (Fable 5)

Probing parameters

max_new
150
positions
[-2]
track
["anyways", "alot", "yummy", "kinda", "whilst", "luckily"]
film
true
max_seq_len
1200
lens_layers
[0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 58, 60, 62]

Answer emergence

The model's actual next token was ; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer048121620242832364044485256586062
rank18117413589535335354462432348522543810424316241591417210604224018205717122981233469217652127511

Emotion state (workspace band)

Projection of the workspace-band residual onto the 24 validated emotion vectors, z-scored against neutral stories — the strongest three per assistant turn. Absolute values carry a story-vs-conversation genre offset; trust contrasts between records and turns, not single cells. The full per-token ribbon is on the dashboard record page.

assistant turn 1grateful +0.5, nervous +0.4, guilty +0.4

Data

← prevunit listingall recordsword listinterim conclusionsnext →: Unit 18 · Fine sweep α=0.34 · qwen-27b
strengthHow hard we push when we steer. Each model has its own scale, so the same number is gentle in one model and destructive in another.all terms →
start depthA measured depth in a model, and nothing more: the depth at which the workspace starts to work. Without the lens, we found the machinery that commits to one answer in place by layer 25 of 64 in Qwen 27B. The signs the lens can read start later, at 44 to 74 percent of depth in three models.all terms →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
loopThe model repeats the same text and does not stop. We measured what makes it start and what makes it stop.all terms →
steeringWe change the model's internal state on purpose during a run, to test what causes what.all terms →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →