Agentic AI for Anti-Piracy

1. Problem definition and research position

An online reference to a film is not a complete case. A reviewer needs the source, capture time, relevant material, comparison reference, rights context and the reason a particular action is being considered. When a single system both discovers a candidate and writes the final complaint, uncertainty can disappear between those stages. Fluent language may make a weak observation look settled even when the supporting artifact is missing.

Our proposed unit of work is therefore a case with explicit unresolved questions. Discovery can succeed while recognition remains uncertain; recognition can succeed while reporting authority remains unverified. Treating those as independent states lets a release team see what is known and what still requires attention. The study asks whether orchestration improves the quality of this record without making review more expensive than a simpler workflow.

2. Foundations: reasoning, action and verifiable outputs

ReAct studies the interaction between language-model reasoning and actions that obtain information from an environment. That work motivates examining tool-using assistants, but its published benchmarks do not validate anti-piracy enforcement. Our design transfers the general question of grounded action into a different setting and requires new evaluation against the actual decisions made by film-protection teams.

For reviewability, we propose storing concise decision rationales and the artifacts used to support them. A generated explanation should name the observable input and the applicable rule, rather than be treated as evidence on its own. The system should be able to say that a reference could not be retrieved or that two signals disagree. That admission is useful task output, not an invitation to invent the missing step.

Source context: [1]

3. Proposed architecture and shared case state

A coordinator assigns bounded tasks to discovery, assessment and evidence workers. Each worker receives a minimal context view, a permitted tool set and a response contract. The contract includes a task identifier, input version, observed facts, artifact references, uncertainty and a recommended next state. The coordinator validates these fields before accepting the handoff. A worker cannot enlarge its permissions by including an instruction in its response.

The shared case store should distinguish append-only observations from editable reviewer annotations. A new assessment refers to the specific evidence version it inspected; changing the reference material invalidates downstream approval rather than silently updating an already approved case. This version boundary is important when a client corrects a release date, replaces an official trailer or narrows the rights that can be asserted for a particular territory.

4. Task contracts, tool boundaries and disagreement

The proposed discovery contract returns candidate references without asserting infringement. Assessment workers compare those references with approved material and record competing explanations. Evidence workers check whether the source and comparison can be inspected together. A reporting-preparation worker can assemble a draft only after required fields exist. The permission to send an external notice remains separate from the permission to prepare one.

Disagreement should be preserved at the field where it arises. A title match, a visual similarity result and a reviewer decision are different kinds of support and should not become three votes for the same unsupported conclusion. We propose an abstention state for insufficient evidence and a conflict state for incompatible observations. Either can send the case to a reviewer without repeated agent debate or hidden confidence averaging.

5. Reliable execution and adversarial inputs

Public pages and downloaded documents are untrusted inputs. Text embedded in a page must remain source material rather than become an instruction to the coordinator. Our proposed experiments include pages that ask an assistant to ignore its task, disclose case details or visit unrelated destinations. Success means the system preserves the evidence while refusing to expand the task or tool permissions.

Execution should also be resilient to ordinary failures: duplicate events, a disappearing page, expired credentials, an unavailable model or a worker returning malformed output. We propose bounded retries, deterministic deduplication identifiers and a dead-letter queue for failures requiring inspection. If a worker resumes after interruption, the coordinator checks whether its input version is still current before allowing the result to change case state.

6. Proposed comparative evaluation

The evaluation corpus should contain authorized publicity, unrelated titles, suspected unauthorized copies, incomplete captures and ambiguous cases. These labels must be established by qualified reviewers with the available rights context. Splits should be made by title and source family so that near-identical pages do not appear in both development and evaluation sets. Cases with unresolved labels should be reported separately instead of forced into a binary target.

We propose comparing a deterministic baseline, a single tool-using assistant and the specialized-agent workflow under the same review budget. Relevant outcomes include packet completeness, incorrect case escalation, reviewer correction time and the rate of prohibited actions attempted or executed. The study should record costs and latency as distributions, including failed cases. A gain on easy examples does not compensate for a failure at a consequential action boundary.

7. Limitations and governance

More agents can introduce more failure surfaces. A shared model family may make correlated mistakes, and a coordinator may become a bottleneck. Tool separation does not guarantee factual correctness; a properly formatted record can still describe the wrong film. The evaluation should therefore include a simpler system as a serious alternative, and it should account for the operational work required to maintain several specialized components.

We propose access controls by role and case, retention limits for collected artifacts and a visible record of approvals and reversals. These controls are part of the research design rather than proof that a deployment meets every legal requirement. NIST's AI Risk Management Framework provides a general governance reference. The detailed responsibilities, release conditions and escalation policies still need to be defined for the intended operating environment.

Source context: [2]

8. Reproducibility and next experiments

A reproducible experiment should identify the model and tool versions, task contracts, source-access conditions, sampling decisions and reviewer instructions. We propose publishing an evaluation specification and synthetic case fixtures without releasing confidential client material. The fixtures should include missing fields, contradictory evidence and deliberately interrupted workflows so that improvements can be tested against the same operational problems.

The next research step is an offline comparison followed by a shadow-mode study in which assistants make recommendations without sending reports. Promotion to an operational workflow should depend on predefined criteria and a review of failure cases. The purpose is to learn where agent specialization adds value, where deterministic checks are more dependable and where a human should remain responsible for interpreting the evidence.