Evidence Orchestration Agent

What makes it an agent?

It receives a bounded goal, observes structured state, generates and ranks valid actions, selects one registered tool, consumes a step and cost budget, updates memory and evidence state, calculates progress reward, and applies explicit continuation or stop rules. The trace shows structured decision evidence—not hidden chain-of-thought.

Registered tools

• inspect_evidence_window
• inspect_visual_observation
• inspect_text_observation
• inspect_audio_event
• inspect_sensor_event
• update_evidence_graph
• score_evidence_utility
• detect_contradiction
• run_counterfactual
• compare_modalities
• request_next_modality
• generate_grounded_answer
• stop_with_answer
• stop_inconclusive

There is no browsing, shell, external API, arbitrary code, or user-defined tool access.

Plan–act–observe loop

Candidate actions are scored from uncertainty reduction, evidence utility, contradiction value, provenance, novelty, redundancy, cost, latency, remaining budget, and answer completeness. Each step records candidates, selected tool, inputs, results, citations, state deltas, reward components, and stop checks.

Agent Progress Reward

Positive terms include information gain, grounded claims, provenance, contradiction resolution, corroboration, coverage, graph connectivity, and a cited stop. Penalties include redundancy, unsupported output, repetition, cost, latency, budget use, unresolved contradiction, premature stop, and tool failure.

Progress reward is an evaluation heuristic—not truth, legal weight, or evidence validity.

Fixed pipeline versus agent

The fixed baseline follows authored evidence order. The utility-driven policy can skip repeated frames, request a useful modality, select lower-cost sources, or test a contradiction. Results are scenario-specific: the contradiction case requires more agent steps than some simple cases and ends inconclusive rather than forcing an answer.

Boundaries and stop rules

The agent can act only through registered tools and is bounded by 14 steps, scenario cost, available evidence, structured memory, provenance checks, independent-source requirements, and explicit stop conditions. Missing evidence and unresolved conflict can end in an honest inconclusive state.