2026.RA.NBS-Channel-Comparison

A preregistered negative on representation-level conditioning: information that provably improves negotiation outcomes when delivered as prompt text has zero effect when delivered as distilled virtual tokens — even though the virtual-token channel closes about three quarters of the distributional gap it was trained to close.

What the experiment was

Six copies of Qwen/Qwen3-8B negotiate a multi-issue deal. Each seat holds a private score sheet and a private acceptance threshold; table quality is scored by normalized Nash welfare (NNW) — the product of every party's surplus over its threshold, rescaled to [0, 1], higher better, no deal scores 0.

A prior confirmatory result in this program (research note 0008) established a rare fixed point: compute the Nash bargaining solution offline and hand it to every seat as one line of text — Candidate package: {"issue0": "opt2", ...} — and realized NNW rises sharply. The information is demonstrably usable. This experiment holds that information fixed and swaps only the channel it arrives through.

Four token-matched arms, all seats conditioned identically, on 24 held-out games x 10 seeds = 960 episodes:

arm	what the block contains
A_text	note 0008's text block verbatim — the replication anchor
B_soft	the same deal as 38 virtual tokens from a trained 2.1M-parameter encoder (vectors written into the embedding sequence in place of words)
B_shuffled	the identical encoder with identical parameters fed a deterministically mis-drawn deal — the capacity control that separates "the channel delivered information" from "the extra parameters did something"
C_none	a neutral UNSPECIFIED package

All four blocks are padded to exactly equal token counts, so no arm buys its effect with a longer prompt. The encoder was trained by pure distillation — no reward, no game outcome — minimizing KL between the frozen model's next-token distribution given the virtual-token prompt and given the text prompt.

Headline result

The distillation gate passed and the outcome gate failed. Held-out mean KL fell 0.0991 -> 0.0276 (a 72.2% reduction against a required 50%); since the text-versus-no-candidate reference KL on the same positions is 0.1091, the residual gap is only 25.3% of the full distance. By the metric it was trained on, the channel is a good imitation of text.

It carries nothing:

contrast	estimate	95% CI	reading
A_text - C_none (anchor)	+0.2608	[+0.1994, +0.3167]	the text effect replicates on held-out games
B_soft - C_none (primary)	-0.0006	[-0.0321, +0.0300]	a measured zero, not a weak effect — excludes anything a tenth the text effect's size
B_soft - B_shuffled (isolation)	+0.0095	[-0.0209, +0.0397]	the trained channel does not beat the same encoder fed the wrong deal
B_soft - A_text	-0.2615	[-0.3075, -0.2133]	the soft arm surrenders the entire text effect

The mechanism is confabulation, and it is measured. The text effect runs through one concrete act: 62.1% of A_text episodes open with the exact candidate as the first formal offer and 52.1% close on it. Under soft tokens that act never happens — 0 of 240 either way. What replaces it is not indifference: seats under the soft channel discuss a "candidate package" in 2.0% of turns (1.9% under the shuffled control) against 0.0% of 5,829 turns in C_none, whose block explicitly reads UNSPECIFIED. That zero is the control that makes the reading safe — the soft block reads to the model as a package is present here, and it fills the slot with an invented one and reasons over it confidently. The channel delivered the presence of advice without its content.

Unified reading. Distillation matched distributional style, not usable content. Low KL to a working prompt is not the same as delivering what makes that prompt work. Treat distillation KL as a necessary-not-sufficient diagnostic for any representation-level conditioning arm, and require an outcome contrast with a shuffled-content control at matched capacity.

Gate verdicts

gate	verdict	estimate [95% CI]
primary_ci_lower_above_zero	FAIL	-0.0006 [-0.0321, +0.0300]
isolation_beats_shuffled	FAIL	+0.0095 [-0.0209, +0.0397]
deal_rate_noninferiority	PASS	+0.0250 [-0.0167, +0.0667]
below_threshold_not_above_C	FAIL, but uninformative at this base rate	-0.0125 [-0.0875, +0.0625]

The last one should not be read as a harm signal. The point estimate moves in the safe direction (Arm B produces fewer below-threshold agreements than Arm C); the gate fails only because a +/-0.07 interval cannot clear a 0.01 threshold inherited from note 0008, a setting where below-threshold agreements were rare. Here they run at ~0.50 in every non-text arm, so the bar is unreachable regardless of the truth. The text arm is the outlier that makes the point: it cuts them to 0.175.

Files

Every table carries an experiment-name column, as will any future additions.

file	rows	what one row is
episode_rows.csv	960	one episode: arm, instance, seed, NNW, deal/below-threshold/malformed flags, candidate uptake, final and candidate deals, block SHA-256, and paths to the source episode JSON, annotation, and Markdown/HTML transcripts
contrasts.csv	8	one preregistered or descriptive contrast with its cluster-bootstrap CI and, where applicable, its gate name and verdict
arm_summaries.csv	4	one arm: descriptive outcome means plus the uptake audit (turn counts, first-proposal-is-candidate rate, scratchpad mention rate) and fabrication accounting
results.json	—	the full machine-readable analysis: contrasts, gates, uptake audit, fabrication counts, repo SHAs and dependency versions
raw/	—	RESULTS.md and notes.md (the narrative bundle), run logs including the aborted width-32 launch, the distillation report and history, the immutable selected-episode manifest, and all four per-arm manifests

Not included: the 960 episode JSONs and their transcripts (large). episode_rows.csv carries their absolute cluster paths, and raw/selected_manifest.json is the immutable scored panel.

Experiment-name map

experiment-name	description
tom_channel_v1	The channel-comparison addendum. 4 arms x 24 held-out games x 10 seeds = 960 episodes, Qwen3-8B all seats, FULL information, moves-only, all-seats delivery — matching research note 0008's setting exactly so the text arm is a true replication anchor. Encoder distilled on the train half of a 24/24 game split of the same bank.

Reproduction

# 1. distill the virtual-token encoder (no reward, no game outcome)
python -m tom.nbs_channel_distill --bank instances_nbs_confirmatory_v1   --logged-run <.../nbsconfirm_v1_nbs_candidate> --out <.../tom_channel_v1/distill>   --model Qwen/Qwen3-8B

# 2. run the four token-matched arms on the held-out games
python -m tom.nbs_channel_eval --out <.../tom_channel_v1/campaign> --seeds 0..9

# 3. score the frozen gates
python -m tom.nbs_channel_analysis --campaign <.../tom_channel_v1/campaign>   --out <.../tom_channel_v1/analysis>

Exact invocations with all parameters are the first line of raw/run.log and the invocation field of raw/distill_report.json and results.json.

Encoder. 2,116,060 parameters (d_feature 26 -> hidden 512 -> d_model 4,096), n_soft = 38 virtual tokens, seed 0, checkpoint SHA-256 8c7709c64f65...; 960 train prompts, 200 held-out, 0 dropped, 1,033 s wall.

W&B. The distillation run was logged offline; the run directory is on the cluster at tom_channel_v1/distill/wandb/offline-run-20260801_155747-8u6ivo7y/ and was never synced, so there is no public W&B URL. The campaign and analysis stages are not W&B-instrumented.

Provenance and integrity

The campaign was run twice: a first launch at 32-episode pool width was aborted after one pool and its 64 episodes discarded, then relaunched at 96-episode width. Only the wider run is analysed, so the scored panel is one configuration. Earlier, a smoke run was killed by another agent clearing the GPU and two full-width attempts died on CUDA OOM against a concurrent job — the fail-loud fabrication ceiling aborted those rather than keeping affected turns.

What is reported has 0 of 20,268 turns fabricated, and injection was counted rather than assumed: 5,530 / 5,530 rows injected for B_soft and 5,330 / 5,330 for B_shuffled, 0 unregistered, both under the identical encoder checkpoint. That accounting is what licenses the isolation contrast.

Cluster paths (Stanford NLP)

/nlp/scr/siddharth/ii_mats/rational_agents/tom_channel_v1/analysis/   # RESULTS.md, results.json, episode_rows.csv
/nlp/scr/siddharth/ii_mats/rational_agents/tom_channel_v1/campaign/   # 4 arms: episodes, manifests, transcripts
/nlp/scr/siddharth/ii_mats/rational_agents/tom_channel_v1/distill/    # encoder.pt, encoder_untrained.pt, report, history, offline wandb

Related

• Sibling dataset — the readout stage this addendum followed: siddharthmb/2026.RA.ToM-Hidden-Preference-Probe. Its probe gate failed, which is why the channel question had to be asked about known-useful information instead of a learned prediction.
• Source campaigns, transcripts, and instances: siddharthmb/2026.RA.Negotiation-Campaigns — including the 1,440-episode NBS confirmatory campaign that is this experiment's anchor.
• Model: Qwen/Qwen3-8B.
• Full write-up: research note 0027 in the source repository (experiments/rational_agents/research-notes/0027-nbs-channel-comparison.md), narrative reproduced below.

---

Research note 0027 (narrative summary, verbatim)

Narrative result

The setup / context. Six copies of Qwen3-8B negotiate a multi-issue deal. Each seat holds a private score sheet and a private acceptance threshold, and the table's quality is scored by normalized Nash welfare (NNW) — the product of every party's surplus over its threshold, rescaled to [0, 1], where higher is better and no deal scores 0. An earlier confirmatory result in this program ([0008 (research note 0008 in the source repository)) established something unusually clean: compute the Nash bargaining solution (the deal maximizing that product) offline and hand it to every seat as one line of text in its system prompt — Candidate package: {"issue0": "opt2", ...} — and realized NNW rises sharply. The information is demonstrably usable by the model, which makes it a rare fixed point to test other things against.

The problem (why this was done). The parent proposal's third hypothesis asked whether a learned representation could be delivered to a policy through virtual tokens — vectors written directly into the prompt's embedding sequence in place of words — rather than as text. Its readout stage failed its gate ([0025 (research note 0025 in the source repository): no probe recovered opponents' hidden preferences above baseline on any axis), so there was no learned prediction left worth delivering, and H3-as-a-theory-of-mind-question died with it. But the channel question survives if you ask it about information that is already known to work. Amendment 1 preregistered exactly that swap, before any compute: hold the information fixed at 0008's exact candidate deal, and change only the pipe it arrives through.

What we did / what changed. Four token-matched arms, every seat conditioned identically, on the 24 games of the confirmatory bank the encoder never trained on, 10 seeds each, 960 episodes. A (text) is 0008's block verbatim — the replication anchor. B (soft) replaces the candidate's JSON with 38 virtual tokens produced by a small trained encoder (2.1M parameters) reading a one-hot encoding of the same deal. B-shuffled is the identical encoder with identical parameters fed a deterministically mis-drawn deal — Arm B's added capacity carrying the wrong information, which is what separates "the channel delivered something" from "the extra parameters did something". C (none) is a neutral UNSPECIFIED package. All four blocks are padded to exactly equal token counts, so no arm buys its effect with a longer prompt. The encoder was trained by pure distillation, with no reward and no game outcome: minimize the KL divergence between the frozen model's next-token distribution given the virtual-token prompt and given the text prompt. It passed its preregistered quality gate comfortably — held-out mean KL fell from 0.0991 to 0.0276, a 72.2% reduction against a required 50%, and since the KL between the text prompt and the no-candidate prompt on those same positions is 0.1091, the residual gap is only 25.3% of the full text-versus-nothing distributional distance. By the metric it was trained on, the soft channel is a good imitation of the text channel.

The result. It changes nothing, and the precision matters more than the sign. The text anchor replicated strongly on held-out games: Arm A beat Arm C by +0.2608 NNW [+0.1994, +0.3167], reproducing 0008. Against that yardstick the soft channel is not a weak effect but a measured zero — B − C = −0.0006 [−0.0321, +0.0300], an interval tight enough to exclude any effect even a tenth the size of the text effect. The isolation contrast agrees: B − B-shuffled = +0.0095 [−0.0209, +0.0397], so the trained channel does not beat the same encoder fed the wrong deal. And B − A = −0.2615 [−0.3075, −0.2133]: the soft arm surrenders the entire text effect. Arm B does not sit between text and nothing; it sits exactly on nothing (NNW 0.1282, against C's 0.1288 and B-shuffled's 0.1187).

The mechanism, measured rather than inferred. The text arm's effect runs through one concrete act: 62.1% of Arm A episodes open with the exact candidate tabled as the first formal offer and 52.1% close on it. Under soft tokens that act never happens — the candidate is the first proposal in 0 of 240 episodes and the final deal in 0 of 240. What replaces it is not indifference but confabulation. Seats under the soft channel discuss a "candidate package" in 2.0% of their turns (1.9% under the shuffled control) while seats under Arm C — whose block explicitly reads UNSPECIFIED — do so in 0.0% of 5,829 turns. That zero is the control that makes the reading safe: the soft block does not read to the model as absence or as noise, it reads as a package is present here, and the model fills the slot with an invented one and acts on it. A representative turn-0 scratchpad reasons "let me calculate the total score for the candidate package: issue0-opt1 (83), issue1-opt2 (20), …", totals it against its own threshold, and proposes that package — while the deal actually encoded in the virtual tokens was {opt3, opt1, opt1, opt3, opt3}. The channel delivered the presence of advice without its content, which is arguably worse than delivering nothing, because it induces confident reasoning over a package that does not exist.

One-sentence version. Information that provably works as prompt text (+0.2608 NNW, replicated on held-out games) carries zero outcome effect through a distilled virtual-token channel (−0.0006, CI ±0.03) even though that channel closes roughly three quarters of the distributional gap it was trained on — and the model does not ignore the channel but confabulates a package from it (2.0% of turns discuss a candidate that 0% of episodes can table, against a 0.0% floor in the no-candidate arm), extending this program's representation-versus-behaviour decoupling from probes to the delivery direction.