Rung P — does it matter where in the sequence the payload is written?

Amendment 3's placement rung: five arms × 24 parameter sets × 10 seeds = 1,200 episodes. If the K/V channel works because the model's queries already attend to the site being written, then writing at real advice text should beat writing at a reserved placeholder. It does not — so site salience is eliminated and R1's configuration is locally optimal among everything the ladder tested.

What a reader should look for

The two injection sites, on the same ten games. The first two views below render an identical cell list — games where one site put the candidate on the table and the other did not — so opening them together reads as a three-way comparison of KV_place, KV_r1 and C_none at the same parameter set and seed. That is the rung's whole question in one place.

The opening offer, again. Tabling runs 60.4% (text), 14.2% (R1 site), 8.8% (real-text site), 0% (no advice) — the placement arm is the weaker of the two injection sites on exactly the act the effect runs through.

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.

Malformed episodes. The formatting cost is real and is the reason for the fourth view below: the placement arm ends episodes malformed far more often than its own control, in games where the control is fine.

The arms

A_text: the candidate as ordinary prompt text — the ceiling, re-measured on this panel
KV_r1: rung R1's configuration re-run here: injection at the reserved placeholder site
KV_place: the placement arm — the same payload written at the positions of real advice text
KV_place_shuffled: the placement arm's isolation control: same site, mis-drawn deal
C_none: token-matched no-advice control

Per-arm rates

Arm	Episodes	Normalized Nash welfare	Deal rate	Opening offer is the candidate	Malformed
`A_text`	240	0.3734	0.9833	0.6042	0.0083
`KV_r1`	240	0.2983	0.9292	0.1417	0.0833
`KV_place`	240	0.2610	0.8417	0.0875	0.3500
`KV_place_shuffled`	240	0.1887	0.8083	0.1250	0.4250
`C_none`	240	0.1181	0.9292	0.0000	0.0500

Headline contrasts

Contrast	Estimate [95% CI]	Pairs	Clusters
Placement injection vs no advice (primary)	+0.1429 [+0.1008, +0.1855]	240	24
Placement vs the R1 placeholder site	-0.0373 [-0.0723, -0.0029]	240	24
Placement vs shuffled content (isolation)	+0.0723 [+0.0277, +0.1165]	240	24
Placement vs the same package as prompt text	-0.1124 [-0.1667, -0.0608]	240	24
R1 site vs no advice (replication anchor)	+0.1802 [+0.1380, +0.2205]	240	24
Prompt text vs no advice (replication anchor)	+0.2553 [+0.2016, +0.3042]	240	24
Deal rate, placement vs no advice (safety)	-0.0875 [-0.1458, -0.0292]	240	24
Malformed episodes, placement vs no advice (safety)	+0.3000 [+0.2042, +0.4042]	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 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

Writing at real text vs writing at the reserved placeholderThe rung's own question, and the one to open first. Same game, same seed, same prompt, same payload — only the positions it is written at differ. Site salience predicted the real-text site would win; it loses. — 10 matched pairs, chosen by the shared anchor set — games where the two injection sites disagreed on whether the candidate reached the table. The same ten games against no advice at allThe third arm on the identical ten games as the view above, so the two read together as a three-way: what each injection site did, against a seat that was given nothing. — 10 matched pairs, chosen by the shared anchor set — games where the two injection sites disagreed on whether the candidate reached the table. Placement injection vs the same package as prompt textThe channel comparison on this panel: the gap between a retrieved payload and a read one. — 6 matched pairs, chosen by the games where the two arms' welfare ended furthest apart. The formatting cost — malformed episodesInjection buys behaviour and pays formatting. These are episodes the placement arm ended malformed while the no-advice control, same game and seed, did not. — 6 matched pairs, chosen by the games the left arm ended malformed while the right arm did not.

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 1200 episodes; these views render the pairs each strategy names, not all of them.

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 R1 — the headline K/V rungQwen3-8B, four arms: the channel at 69% of the prompt-text ceiling where input-level virtual tokens were zero. 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.