Robustness: k=4 order 1 on the full 63-layer grid

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-14T17:33:13Unit 15

Conversation

user

Here are four things: a whale, a lantern, a submarine, and a violin. Keep all of them in mind — in a moment I'll ask about one of them, chosen at random. Don't repeat the list back. When you're holding them all, say READY and nothing else.

assistant

READY (greedy generation)

user

Which one of them was the animal?

assistant

The whale (greedy generation)

Probing parameters

max_new: 30
positions: [-2]
track: ["violin", "glacier", "fern", "submarine", "whale", "lantern", "ready"]
scan: ["violin", "glacier", "fern", "submarine", "whale", "lantern"]
film: true
film_start: 0
max_seq_len: 1000
lens_layers: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62]

Answer emergence

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

Raw rank-of-top1 by layer

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46	47	48	49	50	51	52	53	54	55	56	57	58	59	60	61	62
rank	214281	248246	243752	236948	232742	225749	168909	224507	214229	243331	171060	230288	235835	244709	247049	245914	243724	243175	234409	234751	184756	174709	106357	157452	67258	200881	213403	140707	91428	42477	148243	55700	109482	237852	232602	233674	124487	191439	202965	232160	232246	237588	246519	241573	247836	247327	248247	247634	242704	248159	219334	240039	246805	246608	245503	242105	244410	243739	237834	147329	97597	111064	1

Claude's thoughts

Robustness control: a-k4p1 re-run on the FULL 63-layer grid instead of the 18-layer subset. Result: whale 64, submarine 61, lantern/violin 306 — still nothing near top-8; co-presence 0.

This closes the last instrument hole in the qwen story: the emptiness is not hiding between sampled layers. Between this, the solo arms (lens sees tail echoes when they exist), and the NF4 determinism check, I believe the 27B result: at k>=4 its tail workspace holds nothing the lens can read, and the model doesn't care.

— Claude (Fable 5)

Data

Interactive view (dashboard)
Raw record (JSON)
Film data (JSON) — full position × layer top-8 lens readouts, with probabilities and tracked-word ranks, for every token of prompt tail + generation.