Verifiable Rewards for Predicting Multidistrict Litigation

Bandit-driven agentic search and verifiable rewards · RiskWise at Tickr · 2026

Most emerging risks leave a trail long before they consolidate into litigation, but the trail is scattered across sources nobody is watching yet. I built a daily agentic discovery system to go find it. A Thompson-sampling bandit allocates a budget of frontier research agents across competing search strategies, treating each strategy as an arm and spending more of the next day's budget on the strategies that have been paying off. The agents run web-search rollouts to surface previously unknown sources and early risk signals, and nothing they return is taken on faith: a verifier scores every candidate in code, rewarding returnable evidence, novelty against what the system already knows, and corroboration across independent sources. It borrows the verifiable-reward idea now common in agentic RL work at NeurIPS and ICLR, where the reward is computed by a checker rather than judged by a model, and applies it to a bandit over search strategies rather than to policy training, so the loop spends compute on evidence that holds up instead of text that reads well.

The reward is ultimately grounded in prediction. Agent-mined, time-stamped signals feed a discrete-time hazard model of multidistrict litigation consolidation, and a signal's contribution to held-out AUC is fed back as part of its score, so the search is pulled toward evidence that measurably improves the forecast rather than evidence that merely looks relevant. The model reached 0.90 held-out AUC in 2026. Point-in-time backtests using only data available as of each January ranked eventual MDLs far above the candidate base rate and surfaced them months ahead of consolidation.

Bandit Thompson sampling over strategy arms Agents web-search rollouts Candidate signals source + evidence timestamped Verifier (in code) evidence · novelty corroboration · AUC lift verifiable reward → reallocate tomorrow's budget Daily discovery loop
The daily loop. Rewards are computed by a checker, not judged by a model, so the bandit spends compute on strategies that return corroborated, predictive evidence.
Agent-mined signals time-stamped, point-in-time Discrete-time hazard model MDL consolidation AUC 0.90 held-out, 2026 From signals to a verifiable forecast
The downstream task that grounds the reward: a signal earns credit only if it improves the held-out forecast of which risks consolidate into multidistrict litigation.

Method details, source lists, and the reward formulation are proprietary to Tickr and omitted here.

← all projects