Epistemic status: This appendix is exploratory. The model extension (Section F.1) is a formulation — a representational choice extending TFT's single-agent framework. The communication gain (F.2) is a hypothesis — an extension of TF-06's uncertainty ratio to social channels that is structurally motivated but not derived. The distributed tempo and persistence conditions (F.3) are derived from the model extension, conditional on the communication gain hypothesis. The topology analysis (F.4) and game-theoretic integration (F.5) are discussion — qualitative consequences and connection points with existing theory, not formal results. Throughout, the single-agent TFT results are assumed; this appendix adds a coupling layer.
TFT's core theory treats a single agent coupled to an environment. But many of the most important adaptive systems involve multiple agents whose feedback loops interact: organizations, teams, adversarial competitors, multi-agent AI systems, immune networks, scientific communities, ecosystems. TF-08 and TF-11 already address the two-agent adversarial case. This appendix extends the framework to
The extension principle: each agent runs a local TFT loop; inter-agent interaction enters as additional observation and disturbance channels. This preserves all single-agent results and adds a coupling layer.
For agents
-
$M_i$ : local model (TF-03) -
$\delta_i$ : local mismatch (TF-05) -
$\eta_i^*$ ,$\mathcal{T}_i$ ,$\alpha_i$ ,$R_i$ : local gain, tempo, sector bound, model class capacity (TF-06, TF-11, App. A)
[Formulation (multi-agent-update)]
where:
- First term: standard self-update from direct observation (single-agent TFT, unchanged).
-
Second term: model transfer from neighbors
$\mathcal{N}(i)$ — a weighted sum of corrections based on model differences. -
$w_{ji} \in [0, 1]$ : trust weight that agent$i$ assigns to agent$j$ . -
$\eta_{ji}$ : communication-channel gain (how much$i$ updates from$j$ 's signal; defined in F.2). -
$h_{ji}(M_j^-, M_i^-)$ : the communication transform — maps the difference (or message) between$j$ 's model and$i$ 's model into an update direction for$i$ . This is the social analog of TF-06's mismatch transform$g$ .
What counts as "communication." The second term is general: it covers explicit message exchange (language, data transfer), observation of another agent's actions (social learning, imitation), access to another agent's outputs (ensemble aggregation, consulting an expert), and even indirect signals (market prices, published results). In each case, agent
Relation to query actions (TF-08). Query actions are the mechanism by which the communication term is activated: agent
The optimal gain for incorporating agent
[Hypothesis (communication-gain)]
where:
-
$U_{M_i}$ : agent$i$ 's model uncertainty (same as TF-06) -
$U_{o,ji}$ : communication channel noise — latency, ambiguity, compression loss, bandwidth limitations of the channel between$j$ and$i$ -
$U_{\text{src},j}$ : source quality uncertainty —$i$ 's uncertainty about$j$ 's model calibration and domain competence. Low when$j$ has demonstrated reliable predictions; high when$j$ 's track record is unknown or poor -
$U_{\text{align},ji}$ : alignment uncertainty —$i$ 's uncertainty about whether$j$ 's communications are optimized to improve$i$ 's model or to serve$j$ 's potentially conflicting objectives. Zero for fully aligned teammates; high for competitors or unknown agents; potentially very high for adversaries
Normalization requirement. All four uncertainty terms must be expressed in a common predictive-dispersion scale before summation — the same units as
Additivity as heuristic. The additive denominator treats all uncertainty sources as independent, zero-mean noise — a structural heuristic, not a strict variance derivation. This is appropriate for
When all additional uncertainty terms are zero (perfect channel, perfectly calibrated and fully aligned source):
When any term is large (noisy channel, unreliable or misaligned source):
The denominator sums are not all equivalent in nature:
-
$U_{o,ji}$ is a property of the channel — potentially improvable by infrastructure -
$U_{\text{src},j}$ is a property of the source — improvable by$j$ improving its own model, or estimable by$i$ through calibration tracking -
$U_{\text{align},ji}$ is a property of the relationship — the game-theoretic variable (F.5)
When
When
Agent
A minimal Bayesian trust update: track per-source calibration by comparing
When agent
Bayesian transitive prior. Model
[Hypothesis (transitive-trust)]
where
Three-phase trust formation:
-
Prior:
$P_i(\theta_k)$ — default trust for unknown agents (could be context-dependent: higher in a cooperative team, lower in an adversarial environment). -
Transitive update: Incorporate
$j$ 's recommendation via the mixture likelihood above. Multiple intermediaries contribute additively (each with their own$r_{ji}$ ), and corroborating signals concentrate the posterior faster. -
Direct experience: As
$i$ interacts with$k$ directly, each interaction provides a ground-truth observation that updates the posterior normally. After sufficient direct experience, the intermediary-derived prior is swamped.
Risk-asymmetric trust (HILP/LIHP). The Bayesian posterior gives the best estimate of
- HILP (High Impact, Low Probability): Trusting a deceptive agent can cause catastrophic model corruption (Corollary A.3.1's effects spiral), possibly irreversible. The cost of misplaced trust in this regime is disproportionately large.
- LIHP (Low Impact, High Probability): Mild miscalibration of trust toward a reliable source causes small ongoing inefficiency — slightly suboptimal gain, minor tempo loss.
The risk-adjusted communication gain replaces the posterior mean of
[Discussion (risk-adjusted-trust)]
where
The severity of the quantile
Connection to TF-06 gain dynamics. This risk adjustment is the trust analog of TF-06's observation that an agent with high
Agent
[Definition (distributed-tempo)] $$\mathcal{T}i = \underbrace{\sum_k \nu_i^{(k)} \eta_i^{(k)*}}{\text{direct observation tempo}} + \underbrace{\sum_{j \in \mathcal{N}(i)} \nu_{ji}^{\text{comm}} , \eta_{ji}^*}_{\text{communication tempo}}$$
where
The effective disturbance rate for agent
[Formulation (disturbance-decomposition)] $$\rho_i = \rho_{i,\text{env}} + \sum_{j \in \mathcal{A}i} \gamma{j \to i}^{\text{adv}} , \mathcal{T}j - \sum{j \in \mathcal{C}i} \gamma{j \to i}^{\text{coop}} , \mathcal{T}_j$$
where
The cooperative term is negative. Allies reduce agent
Sign convention. The decomposition above can yield
[Definition (effective-disturbance)]
When
Applying Appendix A's sector-condition framework with
[Derived (team-persistence)]
Substituting the decomposition (the
This reveals three distinct levers for team persistence:
-
Increase
$\alpha_i$ (individual correction efficiency) — better models, better gain calibration -
Increase cooperative tempo (
$\gamma^{\text{coop}} \mathcal{T}_j$ ) — more reliable allies, faster communication -
Reduce adversarial coupling (
$\gamma^{\text{adv}} \mathcal{T}_j$ ) — better deception detection, reduced exposure
Communication channels have costs: time to compose/parse messages, bandwidth limitations, synchronization requirements, meeting overhead. These costs reduce the agent's effective tempo by diverting time and compute from direct adaptation. Let
The net benefit of adding agent
[Discussion — Coordination Threshold]
Both sides have units
The multi-agent framework instantiates differently depending on the coupling structure.
Agents with roughly equal capability and aligned objectives share information to reduce collective mismatch.
- Update structure: Weighted consensus or diffusion — each agent moves its model toward a trust-weighted average of neighbors' models, while also updating from its own observations.
-
Key metric: Disagreement energy —
$\sum_{(i,j)} w_{ij} | M_i - M_j |^2$ . High disagreement with low external mismatch often indicates miscalibrated trust (one agent is right, the other hasn't updated), not model inadequacy. Low disagreement with high mismatch indicates the team is wrong together — a shared structural limitation (TF-10). -
Failure mode: Groupthink — the network converges on a shared model that is confidently wrong, because internal consensus suppresses the mismatch signals from dissenting agents. In TFT terms:
$U_{\text{src},j} \to 0$ for all in-group sources (excessive trust), while$\eta_i^*$ for direct observations is depressed by the confidence of the group model. The remedy is to maintain sensitivity to direct-observation mismatch even when the group is confident — i.e., to never let communication tempo fully substitute for observation tempo.
Multiple agents maintain domain-specific models; an aggregator combines them.
- Update structure: The aggregator's policy weights experts by their uncertainty and historical calibration, not confidence alone. An expert that is confident and wrong should be downweighted; one that is uncertain but well-calibrated should be upweighted.
- Exploration policy: The aggregator should allocate queries (attention, resources) to experts whose models are most uncertain about the current situation — the CIY-maximizing strategy at the aggregation level.
- Structural adaptation trigger: If all experts fail similarly (correlated residuals), the model class needs expansion — the ensemble's diversity is insufficient. If only some fail, re-route or recalibrate rather than restructure.
Agents at different timescales coupled vertically, mapping TF-11's temporal nesting to organizational structure.
- Update structure: Upper layers receive summaries (quasi-steady-state output) from lower layers and issue directives (hyperparameter updates: goals, constraints, priors), not high-frequency control commands.
-
Key constraint: TF-11's convergence condition (
$\nu_{n+1} \ll \nu_n$ ) applies between organizational levels. When upper-layer intervention frequency approaches lower-layer correction timescale, the result is command-induced oscillation — the organizational analog of control-theoretic instability from gain mismatch. This is the formal content of "micromanagement degrades performance." -
Information asymmetry: Upper layers have broader context but lower resolution; lower layers have higher resolution but narrower scope. The communication gain
$\eta_{ji}^*$ should reflect this: upward communication compresses local detail into aggregate summaries ($U_{o,ji}$ high from compression), while downward communication translates broad directives into local constraints ($U_{\text{align},ji}$ encodes the mapping quality from strategic objectives to tactical actions).
TFT provides the state variables (mismatch, gain, tempo, reserve) and dynamics (the correction loop). Game theory provides the strategic analysis: when agents choose their coupling parameters (
TFT alone cannot answer:
-
When is honest communication incentive-compatible? — Under what conditions does agent
$j$ prefer to send signals that reduce$i$ 's mismatch rather than increase it? The cheap-talk and signaling-game literature addresses this. -
What trust levels are equilibrium? — In repeated interaction, what
$U_{\text{align},ji}$ values are self-consistent? Reputation games and mechanism design provide tools. - How do adversaries optimally allocate disruption? — An adversary facing multiple targets must choose how to distribute its tempo across opponents. This is a resource-allocation game.
Within each agent's loop, decision theory refines the policy:
-
$\lambda$ as value-of-information price (TF-08). The exploration weight$\lambda(M_t)$ in the unified policy objective is a decision-theoretic quantity: it prices the expected future value of information against the expected immediate value of exploitation. In environments with known dynamics,$\lambda$ is derivable (Gittins index, information-directed sampling). In general, it is a design parameter whose setting reflects the agent's time horizon and risk preference. -
Deliberation stopping (TF-09). Proposition 9.1's threshold is an optimal-stopping criterion: continue deliberating iff the marginal gain exceeds the marginal cost. This is a classic metareasoning problem from bounded rationality. The extension to multi-agent settings adds a dimension: deliberation may include consulting other agents, whose response time enters the deliberation cost.
-
Structural adaptation as switching under costs (TF-10). The decision to change model class is a sequential hypothesis test with switching cost. Decision theory gives the optimal switching rule: change when the expected future mismatch reduction from the new model class exceeds the transition cost (knowledge loss, temporary performance drop, coordination cost). In multi-agent settings, structural change may require coordinated switching — adding a consensus or signaling requirement.
TFT's value for game-theoretic analysis is providing precise, measurable state variables for strategic interaction:
-
Tempo as a strategic resource:
$\mathcal{T}$ is the asset agents compete over or cooperate to increase. -
Reserve as vulnerability:
$\Delta\rho^*$ is the target adversaries aim to exhaust and allies aim to preserve. -
Mismatch as the battlefield:
$|\delta|$ is what adversaries try to increase in opponents and reduce in themselves. -
Trust calibration as the meta-game: the accuracy of
$U_{\text{align},ji}$ estimates determines whether the agent can correctly price communication signals.
This gives game theory concrete operands rather than abstract "utilities" — the payoff structure is grounded in measurable adaptive quantities.
For implementing multi-agent TFT, the following quantities should be tracked per source
| Symbol | Meaning | Estimation |
|---|---|---|
| Reliability calibration of source |
Running Brier/log-loss score against subsequent ground truth | |
| Communication delay | Direct measurement of message latency | |
| Alignment score | Inferred from agreement on verifiable predictions; domain-specific | |
| Tempo-equivalent coordination cost | Direct-adaptation tempo lost to communication overhead |
|
| Network disagreement energy |
|
- High-stakes updates from low-reliability sources should require cross-source agreement (quorum) before integration.
- Track per-source calibration over time; auto-adjust
$w_{ji}$ based on verified prediction accuracy. - Separate query actions for information from actions on the world in policy logs — these have different CIY structures and different cost profiles.
- Trigger structural adaptation only when residual mismatch patterns persist after: (a) trust recalibration, (b) routing changes, (c) communication-channel upgrades.
If the multi-agent extension's core claims are wrong, the following observable failures should occur. These are testable in multi-agent simulations (distributed estimation, multi-agent RL, adversarial games) and, with instrumentation, in organizational and team settings.
-
Adaptive communication gain outperforms fixed trust. The communication gain hypothesis (F.2) predicts that
$\eta_{ji}^* = U_{M_i}/(U_{M_i} + U_{o,ji} + U_{\text{src},j} + U_{\text{align},ji})$ should yield lower steady-state mismatch than any fixed trust weight$w_{ji}$ when source reliability varies over time or across contexts. Falsified if: in a multi-agent estimation task with heterogeneous and time-varying source reliability, a well-tuned fixed trust weight consistently matches or outperforms the adaptive communication gain across a range of drift rates and source-failure scenarios. This would indicate that the additional uncertainty terms ($U_{\text{src}}$ ,$U_{\text{align}}$ ) either don't carry usable information or that the overhead of estimating them exceeds their value. -
Cooperative communication enables persistence beyond individual capacity. The team persistence condition (F.3) predicts that agents with $\mathcal{T}i < \rho_i / R_i$ individually (insufficient solo tempo) can persist when cooperative communication provides enough distributed tempo to close the gap. Specifically: adding a reliable ally with tempo $\mathcal{T}j$ and cooperative coupling $\gamma{j \to i}^{\text{coop}}$ should reduce $i$'s effective $\rho_i$ by $\gamma{j \to i}^{\text{coop}} \mathcal{T}_j$, bringing the persistence condition into range. Falsified if: in a controlled setting, adding cooperative communication channels with measured
$\gamma^{\text{coop}}$ and$\mathcal{T}_j$ does not reduce$i$ 's steady-state mismatch by the predicted amount. This would indicate the cooperative$\rho$ decomposition is structurally wrong — cooperation helps through a mechanism not captured by the additive disturbance model. -
Risk-adjusted transitive trust outperforms Bayesian mean trust under adversarial contamination. The HILP/LIHP trust adjustment (F.2) predicts that using a conservative quantile
$Q_p[\theta_k]$ rather than the posterior mean$\mathbb{E}[\theta_k]$ for setting$U_{\text{align}}$ should yield better worst-case mismatch in environments with occasional deceptive sources, at the cost of slightly worse average-case efficiency. Falsified if: in a multi-agent setting with a known fraction of adversarial sources, conservative-quantile trust does not improve worst-case mismatch relative to mean trust, or if the average-case efficiency cost is so large that total mismatch is higher. This would indicate that the risk-asymmetric adjustment overcorrects — that the HILP tail risk is already adequately handled by the Bayesian posterior without additional conservatism.
These predictions target the extension's three main hypotheses (communication gain structure, cooperative persistence, risk-adjusted trust). Failures in (1) or (2) would undermine the core multi-agent framework; failure in (3) would indicate the trust transitivity mechanism needs refinement while leaving the broader structure intact.
-
Emergence. When does the network exhibit qualitatively different adaptive behavior than any individual agent? TFT gives a partial answer: when the network's effective distributed tempo exceeds any individual's, the network can persist in environments where no individual can. But "emergence" in the richer sense — novel model classes or capabilities arising from interaction that no individual agent could discover alone — requires a theory of network-level structural adaptation (TF-10 at the collective level). This is beyond TFT's current scope.
-
Convergence in consensus dynamics. The peer-network update rule (F.1's second term) defines a dynamical system over model states. Under what conditions does this converge? Standard consensus theory gives conditions for convergence of scalar averages; convergence of structured models (posterior distributions, value functions, neural network weights) is a harder problem.
-
Adversarial equilibria. In the two-agent case (TF-11, Appendix A), the adversarial outcome is determined by tempo ratios and reserves. In
$N$ -agent networks, coalition formation, asymmetric information, and multi-front competition create richer dynamics. Mapping TFT's framework onto$N$ -player game solutions is a substantial extension. -
Trust transitivity — remaining questions. The three-phase model in F.2 (prior → transitive update → direct experience) with risk-asymmetric quantile adjustment provides a starting framework. Open questions remain: (a) How should the null distribution
$P_0$ be chosen? A uniform$P_0$ makes the mixture model equivalent to soft discounting; a structured$P_0$ (e.g., biased toward deception) models actively adversarial intermediaries. (b) How should the risk quantile$p$ be calibrated? Ideally from the loss function, but the loss from deception may be hard to estimate before it occurs. (c) When multiple intermediaries provide corroborating recommendations, how should correlation be handled? If$j_1$ and$j_2$ both recommend$k$ but both got their information from the same source, the corroboration is illusory. The trust network's dependency structure matters. -
Communication as structural adaptation. A well-formed query to an expert can transfer not just parametric information (reducing
$|\delta|$ ) but structural information (expanding$\mathcal{M}$ ). This is the "grafting" mechanism from TF-10: incorporating another agent's representational structure. A theory of when and how communication triggers structural vs. parametric adaptation in the receiver is needed.
| Concept | Single-agent TFT | Multi-agent extension |
|---|---|---|
| Update | + communication term |
|
| Gain | ||
| Tempo | + |
|
| Disturbance |
|
|
| Persistence | Same, with cooperative |
|
| New quantities | — |
|