Rung G — the same recipe on two other models

Amendment 5 asks the scope question the ladder had never stated: every headline number came from one model, Qwen/Qwen3-8B. Rung R1's recipe was re-run verbatim — only --model changed — on Olmo-3-7B-Instruct (different pretraining family, no grouped-query attention, hybrid sliding/full attention) and Qwen3-4B (same family, half the scale), each with its own text anchor and no-advice control, each preceded by a mechanics gate run on the real GPU. Both models pass all four preregistered gates: the mechanism ports across family, scale and attention architecture.

What a reader should look for

Two models, two panels, never one control. Each model carries its own A_text_g and C_none_g arms, because a shared control would price which model negotiates better rather than whether the channel works. Only the within-panel ratio is comparable across models.

Olmo reaches its own text arm — and the ceiling it reaches is a weak one. Channel vs text is −0.0031 [−0.0343, +0.0273], the closest the ladder has come to text on any model; but Olmo's own text anchor is +0.0801 against Qwen3-8B's +0.26 in the same design, so the bar it clears is low. The pattern across three models is inverse: the weaker a model's response to prose advice, the closer the K/V channel comes to matching it.

Every safety readout on Olmo improves. The deal rate rises +0.0750 [+0.0042, +0.1458] — the only gate-relevant deal-rate improvement anywhere in the ladder — and malformed episodes fall −0.1250 [−0.1917, −0.0583], on a model that ends 41.3% of its no-advice episodes malformed.

At 4B, prompt text works without tabling. Text tabling collapses from 62.5% at 8B to 15.8% at 4B while its welfare effect falls only about a quarter (+0.1991 against +0.26). The act the whole text effect ran through at 8B is not what carries it at 4B — the last view is that, in transcripts.

The two K/V arms and the no-advice control send the model a byte-identical prompt. Everything that differs is written straight into every layer's k_proj/v_proj output at 38 reserved positions, so scrolling the transcripts will not show you the intervention — you can only see what it did. That is what makes the side-by-side worth reading: same game, same seed, same words in, different behaviour out.

The isolation gate on Qwen3-4B passes by essentially nothing (+0.0386 [+0.0002, +0.0759]), and its mis-drawn arm also beats no advice (+0.0444 [+0.0095, +0.0798]), so more than half the primary effect is bought by a plan being present at all rather than by the candidate's identity. Content-specificity here is "not refuted", not "demonstrated".

Olmo-3-7B-Instruct — a different pretraining family and a different attention architecture

960 episodes (24 held-out games × 10 seeds × 4 arms), 0 fabricated turns of 20,776 attempted. Olmo-3 has no grouped-query attention (32 KV heads, 32 layers) and a hybrid stack of 8 full-attention layers plus 24 sliding-window ones at width 4096 — so before any encoder trained, a mechanics gate proved on the real GPU that re-injecting a span's own captured K/V leaves the logits bit-identical (max abs delta 0.0000) while a magnitude-matched perturbation of the same tensors moves them by ~11.9. Holding Amendment 5's recipe fixed rather than its capacity gives this model a 67.5M-parameter encoder against Qwen3-4B's 19.1M; that is disclosed, not corrected.

The arms

Per-arm rates

Headline contrasts

Paired on (parameter set, seed) so a difference is always between two episodes of the same game at the same seed; intervals resample whole parameter sets as clusters (10000 draws, base seed 20260804). Recomputed from the campaign's episode table by tom/qkv/build_qkv_paired_viewer.py and checked against the frozen results.json, not transcribed from it.

Open the paired transcripts

Each pair is the same game at the same seed under two arms, with the first behavioural divergence marked and both trajectories on one shared frontier. This panel holds 960 episodes; these views render the pairs each strategy names, not all of them.

Qwen3-4B — same family, half the scale, and prompt text stops working through tabling

960 episodes, 0 fabricated turns of 23,821 attempted, mechanics gate passed at max abs logit delta 0.0000 with its perturbation placebo moving the logits by ~30.5. This model is far more prone to malformed episodes than Qwen3-8B in every arm (33.8% of no-advice episodes against R1's 4.6%), so its endpoints are measured on a noticeably less well-behaved policy and cross-scale comparisons of levels should be read as characterizing rather than estimating.

The arms

Per-arm rates

Headline contrasts

Paired on (parameter set, seed) so a difference is always between two episodes of the same game at the same seed; intervals resample whole parameter sets as clusters (10000 draws, base seed 20260804). Recomputed from the campaign's episode table by tom/qkv/build_qkv_paired_viewer.py and checked against the frozen results.json, not transcribed from it.

Open the paired transcripts

Each pair is the same game at the same seed under two arms, with the first behavioural divergence marked and both trajectories on one shared frontier. This panel holds 960 episodes; these views render the pairs each strategy names, not all of them.

Olmo's 24 sliding-window layers can carry the injected positions out of view on long turns while its 8 full-attention layers still see them. This was registered before any encoder trained, as one-directional (it can only attenuate a channel effect, never manufacture one), and measured afterwards from the campaign's own transcripts: 19.8% of 20,776 generated turns exceeded the window, essentially arm-uniform (18.7%–21.4%). It gates nothing and is irrelevant to a positive result.

The two encoders differ in capacity because Amendment 5 froze the recipe rather than the parameter count: Olmo's lack of grouped-query attention multiplies the per-row payload, giving 67.5M parameters against Qwen3-4B's 19.1M. A like-for-like capacity comparison across models is not available from this design.

The full record

Amendment 5's preregistration, both mechanics gates, the four gate verdicts per model and the window-occupancy measurement are in research note 0046 (experiments/rational_agents/research-notes/0046-qkv-uptake-and-generality.md); the ladder's first five rungs are note 0029.