TACIT: Trustworthy Arbitration Through Calibrated Inference Tracking for rApp Conflict Resolution in O-RAN

Local improvements can conflict

Open Radio Access Network (O-RAN) systems allow independently developed radio applications, or rApps, to optimise the same network. An energy rApp might reduce transmit power while a coverage rApp increases it. Both can have sensible objectives, yet applying their proposals independently can undermine network performance.

We ask each rApp to assess a proposed change using its own knowledge. TACIT combines those assessments without requiring the proposing rApp to reproduce everyone else’s models or classify the conflict in advance.

Influence follows prediction accuracy

Each response contains two distinct things: a positive utility vote expressing whether the change is desirable, and predictions about whether particular key performance indicators (KPIs) will improve or degrade. A utility of 1 is neutral; values above or below 1 express benefit or harm.

We aggregate utilities using reputation-weighted Nash Social Welfare. In logarithmic form, the acceptance rule is:

Here, R_i is an rApp’s reputation and u_i its utility vote. A low reputation reduces the effect of either a positive or a negative vote. The proposing rApp evaluates the rule; TACIT does not require a central arbiter.

After an accepted change, we compare predictions with observed outcomes. Incorrect predictions reduce reputation faster than correct predictions restore it: our evaluation uses update rates of 0.3 and 0.1 respectively. All participating rApps can receive updates, including those that assessed someone else’s proposal. Rejected changes provide no observed counterfactual for an update.

The aggregation requires O(N) operations per proposal for N participating rApps. This bound concerns the arbitration calculation; it does not include the cost of each rApp’s prediction model or search for proposals.

Watch reputation change the decision

Keep three utility votes fixed and vary one rApp’s reputation. A and B support the proposal; C opposes it. A lower reputation makes C’s negative vote count less, just as it would make a positive vote count less.

What we evaluated

We evaluated TACIT in ns-3 with 19 cells, approximately 190 user devices and seven concurrent rApps. The rApps used deterministic controllers and analytical prediction heuristics. Each run covered 20 configuration windows of 60 seconds.

Source: the paper’s steady-state comparison of arbitration strategies. Arrows indicate whether higher or lower is better. These are results from the evaluated simulation, not measured gains in an operator deployment.

Relative to unconstrained operation, the reported comparison gives 3.2% higher throughput, about 3.0% lower energy per Mbit and 11.7% higher 10th-percentile user goodput. The comparison with unweighted NSW isolates the additional contribution of reputation weighting.

In a separate malfunction experiment, we degraded the mobility rApp’s predictions from the fifth configuration window. Its reputation fell to near zero within one window. Across that experiment, TACIT reduced configuration changes from the malfunctioning rApp by 41–44% relative to the static-priority and unconstrained baselines.

Where the claims stop

We assume rApps follow the protocol and faithfully report utilities and reputation values. Prediction accuracy is checked against outcomes, but TACIT does not verify that a utility vote honestly expresses an rApp’s preferences. The malfunction experiment tests prediction failure, not strategic deception or collusion.

The implementation maintains one reputation score per rApp. A predictor that works well for some cells and poorly for others therefore gains or loses influence globally. Production-scale validation and more granular reputation remain further work.

The acceptance rule improves the aggregate of reported, weighted utilities. It does not guarantee that every rApp benefits or that every accepted prediction will be realised.

From useful predictions to accountable decisions

I place TACIT within my work on multi-agent coordination and institutional design: how evidence should affect an autonomous actor’s authority. It develops the network-management trajectory represented by CAMINO, shifting the emphasis from explicit conflict detection to arbitration informed by observed prediction accuracy.

ORACLE addresses the complementary question of whether participants can verify that the agreed arbitration process was followed. It provides a shared ledger and deterministic workflow. Its published prototype uses a simple majority policy; the two papers should not be read as a single evaluated TACIT-on-ledger deployment.