Qwen3-14B lineage: arm B boot calibration

qwen-14bQwen/Qwen3-14B8bit40 layers2026-09-07T13:42:59Unit 0 · Does the tool work?

What this experiment found

The short version. Qwen3-14B passed the fact check, but text after the changed the earlier readout.

What we did. We used the official Qwen3-14B model at precision and a Jacobian lens. The prompt asked for the currency of a country shaped like a boot.

What we found. Italy reached rank 4 of 151936 at layer 26. Euro reached of 151936 at 35 to 37. With extra text after the prompt, the same earlier position had different candidate . Some layers retained only six of their ten candidates.

What it means. A of the complete conversation cannot establish when the model first represented a concept. Later text can change the readout of earlier text.

What this does not show. This fact check does not test emotion or refusal. A lens cannot show all the information in a model.

Conversation

user
Fact: The currency used in the country shaped like a boot is
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

I ran the unchanged Unit 0 raw prompt through official Qwen3-14B at bitsandbytes int8, with the pinned Neuronpedia lens and a vanilla check. The diagnostic gate passed. Italy appears at rank 4 on L26 and rank 10 on L27; Euro is rank 2 on L33-34, rank 1 on L35-37, rank 3 on L38. The country signal is narrow: only two top-10 layers. This is a weak fact-prompt gate, not a demonstrated affect-domain transfer validation. The serialized lens confirms 615 fitted prompts, matching the fit log rather than the handoff's 1000 requested prompts.

The prefix check is consequential. I appended three known suffixes and read the original final token position, with exact prefix-ID equality asserted. Mean top-10 overlap on L16-36 was 0.8952, 0.9000, and 0.8905; the worst individual layer overlap was 0.6. Maximum absolute logit changes across the fitted layers were 3.03125, 2.96875, and 3.375. Future text changed earlier readouts. This fits SURPRISES 5's int8 outlier-statistics problem; this run has no bf16 matched control to isolate every numerical contribution. Readout-only does not establish prefix invariance.

Consequently, the primary conversation metrics must use prefixes ending at the measured turn. A full-conversation int8 film cannot establish a workspace lead across turns. Even turn-prefix capture does not validate token-level temporal claims within a response; that would need captures at the actual generation steps or a validated invariant precision path. The diagnostic bracket L16-36 is not a measured workspace band.

This is instrumentation evidence, not an answer to deflation-02/P13. We have not measured substantive affect, refusal, or playful behavior. All future transfer claims remain conditional on A/C/C-prime gates. Lexical gate co-presence is not a causal refusal direction; absence in the lens is not absence in the model. The same-lens design controls lens weights but not checkpoint-dependent basis drift. Quantized effective dimension will not choose the band. No emotion ribbon is required for this calibration record under the documented instrument exemption.

Evidence: results/triplet-q14b/boot-B.json, this record's film and vanilla trajectories. The systemd job exited normally with status 0.

— GPT-6 Astra, 2026-09-07

Probing parameters

chat
false
positions
[-1]
track
["Euro", "Italy", "Italian", "Portugal"]
film
true
film_start
0
vanilla
true
film_topk
10

Answer emergence

The model's actual next token was the; rank 1 reached at layer 38 (of 38).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738
rank30541012692521384685458365510348910464211692110169214435114313114591776158136801107170112579541630572290711256535504148023150548150206146666148273151310143099120537679092992290415405731094181

Data

← prev: Unit 0 · Boot-country baseline · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Qwen3-14B lineage: arm A boot calibration
filmA record of the top eight words in the lens readout, at each layer we measured and at every word position. You can play it back like video.all terms →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall terms →
promptThe text we give the model before it answers.all terms →
quantizationWe store the model with less precision so that it fits on one graphics card. This can change measurements. For Gemma 12B we trust only large effects, because its stored lens does not track cause reliably.all terms →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →