01 // Abstract
Can a latent message from a frozen sender carry its uncertainty to a frozen receiver better than text from the same sender? The question tests the claim of arXiv 2606.20662 that uncertainty needs a latent carrier. The sender is Qwen3-1.7B, and the receiver is SmolLM2-360M-Instruct. The latent message L puts one trained soft vector after each candidate answer, from the block-21 state of the sender. For each arm, a linear readout of the same size predicts if the answer of the sender is correct.
The preregistered run used closed-book TriviaQA, with 4,000 train items, 6,000 test items and 3 seeds. L does not beat T5, the best text that the sender writes from its own outputs (its probability for each candidate answer, as words). The difference is +0.0009 nats (95% CI -0.0076 to +0.0093), and the rule gives case 3. Text with the values of a probe on the same sender state (T4) beats T5 by +0.0213 nats and L by +0.0204. Thus the state holds a signal that the outputs of the sender do not show. But the latent message did not carry more of it than text. A separate audit recomputed every arm and every rule row from the saved per-item arrays, and the run is sealed.
Soft vectors that carry the same probe values beat T4 by +0.0200 nats, only 0.000013 above the margin. A second preregistered test asked if a finer text format closes this gap. The test selected the format WD2: a word and a value to two decimals. On 7,829 fresh items, WD2 keeps a gap of +0.0387 nats (95% CI +0.0322 to +0.0456) to the soft vectors. Its gain over the T4 format is not resolved. This is case C: the gap is not in the number format. On these fresh items, the T4 format is better than L by only +0.0107 nats, below the 0.02 margin.
Two exploratory runs on a synthetic task came first. In both, probe text beat L: test NLL 0.483 against 0.522, and 0.483 against 0.533. In both runs, the selected learning rate of L was at the end of its grid, so the fit limit of L was not removed.
Limits: one sender, one receiver, one task and one latent map family, selected on development data. The receiver does not use an evidence passage. Gate G-D3 failed on development items, and amendment A1 made it descriptive before the freeze, so the run has no upper bound from the evidence. All values come from a trained linear readout, and the label is the correctness of the answer of the sender. Each audit is a separate agent of the lab, not an outside party. The format-gap folder has no seal.
02 // Gates
| Gate | Preregistered condition | Result | Record |
|---|---|---|---|
| Latent beats sender text (primary) | Confirmatory, 6,000 test items, 3 seeds. L must beat S*, the best arm that the sender writes from its own outputs (here T5). The rule needs a mean of at least 0.02 nats, a CI above 0 and each seed above 0. Measured: +0.0009 (95% CI -0.0076 to +0.0093). Case 3. | [ FAILED ] | 77c6fdac3e38… |
| Probe text beats sender text | The row that the rule adds for case 3: T4 against S* (T5), with the same rule. Measured: +0.0213 (95% CI +0.0163 to +0.0262), positive in all 3 seeds. | [ PASSED ] | 77c6fdac3e38… |
| Validity V1, V3, V4 | The NLL of L-shuf (shuffled states) is at least the NLL of T0c (the candidate strings only) minus 0.02. The weight hashes are the same before and after the run. All latent fits end with finite losses. Measured: 0.4521 against 0.3932; one hash for each of the 3 weight files; 9 of 9 fits finite. | [ PASSED ] | 77c6fdac3e38… |
| G-D3 Receiver reads the evidence | Development items only: the Source arm (the question, the answer and 120 words of evidence) has a lower NLL than T0 (the question and the answer). Measured: 0.4632 against 0.4429. Amendment A1, before the freeze, made this gate and the check V2 descriptive. | [ FAILED ] | 0875ab9b2431… |
| Format gap, dev gate | Development items only: F3 gives its Phase D value again (0.40904 ± 0.001), and the selected format F* gains at least 0.005 nats over F3. Measured: 0.40904; F* = WD2, gain 0.0077. | [ PASSED ] | ac50917964ef… |
| Finer format closes the gap | 7,829 fresh items. Case A needs WD2 to close against L-probe-R (the probe soft vectors, with the readout of the text arms): a gap of 0.005 nats or less, with the upper CI bound below 0.010. Case B needs a gain over F3 with the CI above 0. Measured: gap +0.0387 (95% CI +0.0322 to +0.0456); gain +0.0028 (95% CI -0.0002 to +0.0059). Case C. | [ FAILED ] | a597366bc484… |
| Probe text beats latent, fresh items | Format-gap test, the "beats" rule of the main run. In the main run, T4 beat L by +0.0204 (95% CI +0.0134 to +0.0274). Measured on the fresh items: F3 (the T4 format) against L +0.0107 (95% CI +0.0046 to +0.0169), below the 0.02 margin. | [ FAILED ] | a597366bc484… |
| Latent beats probe text, pilot | Exploratory. Synthetic task, 1,000 test items; the rule was fixed in the plan before the data. Case 1 needs L to beat the best sender text and also T3 and T4 (0.02 nats, CI above 0). Measured: test NLL T4 0.483, L 0.522. Case 2. | [ FAILED ] | d5b169c29aa3… |
| Latent beats probe text, follow-up | Exploratory. New seeds, the same task. Case 3 of the plan needs the selected map L* to beat T4. Measured: T4 0.483, L* 0.533; T4 minus L* -0.050 (95% CI -0.092 to -0.011). Case 1. The fit limit of L is not removed. | [ FAILED ] | d6a92cf2813a… |
03 // Destructive controls
| Arm | What it removes | Result |
|---|---|---|
| Shuffled states | The link between state and item. The states of each slot are permuted across items, in train and in test, and the map trains again. | 0.4521 (L 0.4021) |
| Test-only shuffle | The link between state and item at test time only. The trained map gets the states of other items. | 0.5322 (L 0.4021) |
| Strings only | The values. The receiver gets the candidate answers with no numbers. | 0.4132 (T4 0.3817) |
| Question only | The answer and the values. The receiver gets the question only. | 0.4571 |
| Probe vectors | The full sender state. The soft vectors carry only the probe values of T4. | 0.3617 (L 0.4021, T4 0.3817) |
| Base rate | All item information. Each item gets the train mean of the label. | 0.5312 |