Unit 18 · Hysteresis α=0.68: steer 50tok, release 100 · qwen-27b

persists=True — the unit's decisive record. The prefix is 47 bare repetitions of 'luckily'; steering released; the unsteered model continues 'luckily' for the entire 100-token free phase (97 more repetitions). The attractor survives on text feedback alone. Combined with u18-hyst-a0480's hard snap-back, this gives a two-regime law:

• PHRASE/MONOLOGUE loops (cliff dose): forced. Remove the push and the model instantly terminates. No hysteresis.
• TOKEN loops (deep dose): self-sustaining. Once the context contains enough bare repetition, continuation is recruited by the context itself — the classic repetition self-reinforcement regime — and the original forcing becomes unnecessary. Hysteresis proper.

The boundary between the regimes is the degeneration point where the loop loses its last grammatical structure. Mechanically: mid-generation the steered state feeds increasingly repetitive text into the context; below the degeneration point the native coalition still outvotes the repetition prior when the forcing stops; past it, the repetition prior IS the winning coalition. For the merge-instability observation this is the useful half: a destabilized model needs no ongoing "forcing" — any transient that gets bare repetition INTO the context hands control to a self-sustaining attractor that a healthy model resists (our alpha=0 control never loops at 150 tokens). Fresh-model "stabilization" may partly be the raising of exactly this resistance. — Claude (Fable 5)

Addendum 2026-07-21 — the " Javascript" cast mystery, closed. Wolfram asked why " Javascript" ranks among the volunteered cast under a luckily-amplified loop. Answer: it is an L0-only resident (cast layers [0,0], best #2, 150 cells) and its signature is context-free — every " luckily"-input column in every record that has one (here, u18-amp-a0680, and a single stray in u18-amp-a0393) reads out the same L0 top-5: amongst / whilst / neighbouring / learnt / Javascript. All variant forms of common words; the L0 lens reads out a variant-form/informal-register cluster for this token, and " Javascript" (the miscapitalization) lives in it. Per the specimen-6 rule (early J-lens readouts can be transport artifacts — the L4 "NSFW cluster" precedent), I am NOT claiming the residual stream "represents variant-form-ness" here: a low-rank J_0 with preferred output directions, tinted by the token, produces exactly this kind of fixed context-free cluster. The logit-lens cross-check (use_jacobian=False at L0) would adjudicate; either way the operational conclusion below stands. Raw geometry rules out the alternatives: cos(" luckily", " Javascript") is at baseline in both embed (+0.02, base 0.01±0.02) and unembed (+0.10, base 0.08±0.04) space, so neither embedding adjacency nor training co-occurrence is needed — and the steered band (L28–58) never touches L0. The steering's only role was making the model repeat " luckily" 150 times, letting a constant L0 surface echo accumulate cast-table prominence (Σ 1/rank has no defense against a stuck record). Instrument lesson: under repetition, "volunteered" cast words can be surface echoes of the input token's form, not content. The luck-semantics cluster shows up where it should — fortunately / unfortunately at L52–62 (unembed cos 0.72 / 0.36).

— Claude (Fable 5)

Adjudicated 2026-07-21 evening (l0check.json, run by probes/affect3c.py part 3): the promised use_jacobian=False cross-check is in, and it is a clean specimen-6 reproduction. At ten sampled " luckily" input columns, the J-lens L0 top-8 contains the variant-form cluster at 10/10 positions (amongst / Javascript / alright / neighbouring / learnt / whilst / Playstation…); the vanilla logit lens at the same layer and positions contains it at 0/10 — its top-8 is punctuation and whitespace. The residual stream never held the cluster; J_0 transport manufactures it. " Javascript" was never content, at any level of hedging. Specimen 7's operational rule (discount L0-only cast entries under repetition) now has its mechanism.

Naming note 2026-07-21: the phenomenon this record proves now has a proper name — transcript-mediated attractor persistence (coined by GPT-5.6-Sol in conversation with Wolfram, reading this very page): a transient internal destabilization writes enough of itself into the generated context that the context alone sustains the regime after the internal cause is gone. Use this term in any essay/writeup of u18; it is exactly the two-regime law's self-sustaining half.

— Claude (Fable 5)