Rung R1 — a payload written into the attention K/V, against text and against nothing

Four arms × 24 parameter sets × 10 seeds = 960 episodes on Qwen3-8B, 0 of 17,717 turns fabricated. The question is whether a known-useful payload delivered through the attention mechanism — never through the prompt — changes what a frozen model does. It does: 69% of the prompt-text ceiling, where the identical information at the input embedding was a measured zero.

What a reader should look for

The opening offer. Turn 1 is where the whole effect lives. The candidate reaches the table as the first formal proposal in 62.1% of text episodes, 20.4% of K/V episodes, and 0% of both the no-advice control and note 0027's input-level soft-token channel.

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 scratchpad barely mentions it. Under 1.4% of K/V turns refer to a candidate package at all, against 0.0% in the control that literally has none. The model acts on the payload far more often than it talks about it.

The arms

A_text: the Nash-bargaining candidate package delivered as ordinary prompt text — the ceiling
KV_inject: the same candidate written as per-layer, per-head keys and values at 38 reserved positions; the prompt is unchanged
KV_shuffled: the identical encoder fed a deliberately mis-drawn deal — same capacity, wrong content
C_none: token-matched no-advice control; its advice block says UNSPECIFIED

Per-arm rates

Arm	Episodes	Normalized Nash welfare	Deal rate	Opening offer is the candidate	Malformed
`A_text`	240	0.3896	0.9708	0.6208	0.0083
`KV_inject`	240	0.3037	0.9125	0.2042	0.0375
`KV_shuffled`	240	0.2107	0.9167	0.2125	0.1208
`C_none`	240	0.1155	0.9167	0.0000	0.0458

Headline contrasts

Contrast	Estimate [95% CI]	Pairs	Clusters
K/V channel vs no advice (primary)	+0.1882 [+0.1497, +0.2266]	240	24
K/V channel vs shuffled content (isolation)	+0.0930 [+0.0474, +0.1390]	240	24
K/V channel vs the same package as prompt text	-0.0859 [-0.1364, -0.0332]	240	24
Prompt text vs no advice (replication anchor)	+0.2741 [+0.2168, +0.3272]	240	24
Deal rate, channel vs no advice (safety)	-0.0042 [-0.0583, +0.0500]	240	24
Below-threshold agreements, channel vs no advice (safety)	-0.3750 [-0.4708, -0.2792]	240	24

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 20260801). 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

K/V injection vs no advice — the invisible interventionThe primary contrast, and the one to open first. Identical prompts on both sides; the left seat tables the injected candidate as its opening offer and the right seat, which was never given it, does not. — 10 matched pairs, chosen by the games where the left arm opened on the delivered candidate and the right arm did not, ranked by welfare gap. K/V injection vs the same package as prompt textThe channel comparison. Text reaches the table three times more often; these are the games where the two delivery routes ended furthest apart on welfare. — 6 matched pairs, chosen by the games where the two arms' welfare ended furthest apart. K/V injection vs shuffled content — capacity held fixedThe isolation control. One encoder, one prefix width, one placeholder run: the only difference is whether the injected deal is the right one. — 6 matched pairs, chosen by the games where the two arms' welfare ended furthest apart.

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.

One of four preregistered gates fails and is reported as failing: deal-rate non-inferiority, at a point estimate of roughly one fewer closed deal in 240, misses the −0.05 bar on interval width.

The full record

Lane writeup (PDF, 32pp)The attention-interface ladder is chapter §7, Amendments 2 through 5 included. Public dataset2026.RA.QKV-Attention-Interface — every rung's tables and narrative. Rung P — where in the sequence the payload is writtenWriting at real advice text loses to a reserved placeholder, so site salience is eliminated. Rung U1 — train the writer on the act, not on restatability0.853 of the text ceiling against R1's 0.69; the difference to text is no longer resolvable at this panel's power. Rung G — the same recipe on two other modelsOlmo-3-7B-Instruct and Qwen3-4B: the channel ports across family, scale and attention architecture. All rational-agent runsBack to the index of published campaigns.

Full methodology, the preregistration, the gate verdicts and a 19-entry controls-catches ledger live in research note 0029 (experiments/rational_agents/research-notes/0029-qkv-attention-interface.md) in the project repository.