The same 0.5 is not the same recoverable trust state
Two trust estimators display 0.5. After identical positive feedback, one rises to about 0.667 and the other to about 0.549. There is no computational contradiction: the displayed values agree, but the retained evidence masses differ. For a system that must resume updating after restart, saving a score is a different obligation from saving recoverable state.
This counterexample comes from our internal source-function experiment against a pinned TASMAN revision. It examines estimator state and prompt construction. It runs no model negotiations and measures no agreement benefit. The resulting judgment is narrower: estimator state, information shown to a decision maker and guidance for using that information should be distinct design and evaluation objects.[1][2]
One mean, two subsequent updates
The paper's Bayesian reputation method, BRS, retains positive and negative evidence before projecting it into a trust score. In the recorded mechanism, a model reports events and code normalizes the feedback and performs the update. The statement that only trust scores pass between episodes concerns what negotiating models receive; the estimator still retains evidence counts. It does not describe an entirely history-free system.[1, §3 and Appendix C]
Our internal experiment on October 26, 2026 called the unchanged functions at the pinned revision with two states, identical normalized positive feedback and a decay coefficient of 0.9.[2]
| Positive/negative evidence state | Initial score | Updated state | Updated score |
|---|---|---|---|
| (0, 0) | 0.5 | (1, 0) | 0.6666666667 |
| (4, 4) | 0.5 | (4.6, 3.6) | 0.5490196078 |
These are internal source-probe results, not the authors' negotiation measurements or an independent cross-team replication. The states share a current projection but respond differently. A saved value of 0.5 cannot identify which evidence mass to restore.
Our interpretation is that this displayed mean loses information needed for continued updating. That does not require every negotiating model to receive the full history. A system can retain full estimation state while exposing only a mean; its persistence boundary and presentation boundary serve different purposes. If a task merely replays frozen scores and never updates them, the full counts may be unnecessary on that replay path.
The score can stay fixed while guidance changes
A second internal control held the displayed value at 0.20 while constructing prompts that only display it, ask the model to consider it, explain its meaning or provide more prescriptive guidance. The number stayed constant while voter text or system guidance changed. This isolates differences in input construction, not changes in voting behavior.[2]
“Does trust help?” therefore conceals several treatments. Evidence formation determines estimator state; projection determines what the decision maker sees; instructions determine how the model is asked to use it. If all three change between methods, their difference cannot be assigned entirely to an update formula.
The paper also compares frozen, random and replay conditions. Replaying an existing score trajectory constrains information and helps examine use guidance; it does not isolate every behavioral feedback. Its parameter-sensitivity curves recompute estimates offline with signals and negotiation behavior held fixed. They are not results from negotiating again with new parameters.[1, §4 and Appendix E]
Reported benefits belong to particular conditions
For one model in the one-hard-bargainer/three-generalist setting, the paper reports agreement gains from replay over the original score condition of 30.5±6.8, 28.0±10.4 and 16.5±12.4 percentage points for its three trust methods, with paired 95% confidence intervals. These are the authors' grok-r results, not measurements from our probe.[1, Table 3]
Some gemini-r role settings show no significant improvement; replay intervals in the two-hard-bargainer setting also include zero. Accurate adversary identification does not guarantee agreement gains. Frozen and random controls also resist attributing every gain to generic guidance alone. Model identity, role composition and intervals must travel with the numbers rather than disappear behind the largest increase.
This defines the contribution of our evidence: the source counterexample supports insufficient recovery from this mean alone; the fixed-score controls show that guidance can change inputs separately. Neither establishes that retaining more state improves agreement.
What recovery should retain, and what remains untested
For trust systems that continue learning, our engineering interpretation is to retain evidence state and update-rule version as recovery data, recording decision-facing projections and guidance separately. This is an inference from the examined implementation, not a sufficiency theorem for every estimator. Another estimator requires its own account of what state is sufficient.
We made no paid model calls and did not repeat the full negotiation study. Repository defaults do not freeze every paper arm. Multiple or adaptive adversaries, changing identities and mixed-model conditions remain gaps. A next experiment would separately vary recovery behavior and use guidance while holding private utilities and other conditions fixed, then measure agreement and surplus. Those measurements remain future work rather than results of this article.
Evidence and sources
Source-reported findings here are the authors' mechanisms and outcomes; no independent negotiation experiment was examined. Internal experimental evidence means our same-day pinned source-function probe. Interpretations are bounded design inferences. Public code establishes inspectable implementation, while publication, citation and provenance do not establish independent validation or general effectiveness.
- TASMAN research team, research paper, October 2026 version, §§3–5, Table 3 and Appendices C/E: full paper. Mechanisms, study conditions and benefits remain attributed to the authors.
- TASMAN maintainers, public source code, revision
7e55fdb8791dd9ec03819534649420311ab8b374: reputation implementation. The internal probe retains source hashes, inputs, outputs and prompt controls; it measures no model behavior.
Cover: an editorial illustration of state projection, not an experimental result figure.