Defining Rules for Credit Freezing

In music release coordination, metadata—the data about the music, such as song titles, dates, and credits—often arrives in fragments. A common failure happens when a system is configured to treat the most recent file as the 'source of truth.' If your AI agent simply updates the record every time a new file lands, a late-arriving document can silently overwrite a previously agreed-upon fact.

Consider this scenario: On Monday, you receive a producer's split sheet confirming their share is 25%. You log this as the current state. On Tuesday, a different file arrives—perhaps an old draft or a disputed version from a publisher—claiming the producer only gets 10%. If the agent is set to 'last-in wins,' it automatically replaces the 25% with 10% without alerting you. Suddenly, the wrong percentage is sent to the distributor, leading to payment disputes and legal friction.

To prevent this, you must move away from chronological truth. You will configure your agents to recognize that a file does not win a dispute simply by arriving later. Instead, any contradiction must trigger a freeze, keeping the credit flagged as 'open' until you personally resolve it.

At its heart, a release record is more than a spreadsheet; it is a formal act of attribution. As songwriter Tom Gray emphasizes, attribution is the public, named statement that particular human beings made the work. This ensures that the people who contributed their craft are correctly identified, credited, and paid. When you treat the record as an act of attribution, the goal shifts from 'processing files' to 'protecting people.'

For example, imagine a track that started as a voice memo with a hummed bassline. During the recording session, a professional session bassist is brought in and plays the part far better than the original hum. That bassist is a human contributor whose work is now part of the song's identity. If a late-arriving project file from the original demo is uploaded and the AI agent blindly replaces the credits with the demo's notes, the bassist disappears from the record. By prioritizing attribution over file chronology, you ensure the bassist remains on the final record regardless of when the files arrived. Your role is to use the agent as a guardrail that protects these human contributions from being erased by automated updates.

To stop disputed credits from becoming approved facts, you must first define the difference between 'approved facts' and 'pending metadata.' In an AI-driven release workflow, not all data is equal.

Approved facts are pieces of information that have been human-verified and locked. For example, if you as the manager review a songwriter's name and confirm it matches the legal contract, that name becomes an approved fact. Once a piece of data reaches this status, it is no longer subject to automatic updates from arriving files.

Pending metadata, by contrast, is any information found in a file that has not yet been vetted by a human. Consider a new CSV upload from a publishing house that lists a different spelling of that songwriter's name. The AI agent should treat this as pending metadata—a suggestion that needs review—rather than a correction to be applied immediately.

By maintaining this strict boundary, you ensure that your agent doesn't accidentally promote an unverified suggestion to a permanent record. You will instruct your agents to treat all incoming data as pending until a named owner converts it into an approved fact.

To prevent disputed credits from silently reaching a distributor, you must implement a system status called 'Open.' In a standard automated workflow, data is often binary: it is either missing or it is present. However, for a release coordinator, there is a critical middle state. A credit that is being argued over is not 'correct' or 'incorrect' yet—it is simply unresolved. By flagging these items as 'Open,' you create a digital blockade that prevents that specific piece of metadata from moving downstream to the delivery stage.

For example, if you are using a proactive agent like Dots, you can set a specific review rule: "If an incoming split sheet contradicts any previously logged percentage, do not update the record. Instead, tag the field as 'Open: Disputed'."

When this rule is active, the agent stops treating the data as a simple update and starts treating it as a flag. Instead of the record saying 'Producer: 10%' (which might be wrong), the record says 'Producer: OPEN.' This ensures that no matter how many files arrive, the distributor never receives an unverified number. You then monitor the 'Open' queue to resolve these conflicts manually.

In many automated systems, the default logic is 'last-in wins,' meaning the most recent upload is treated as the current truth. In release coordination, this is a dangerous assumption. To protect your release record, you must implement a hard rule for your AI agents: Arrival time does not equal truth. A document's timestamp is a record of when it was sent, not a verification of its accuracy. If a new file contradicts an existing approved fact, the agent must treat the new data as a conflict to be investigated, regardless of how 'fresh' the file is.

For example, imagine you are using a Claude Duet agent to manage credits. You have an approved email from the artist stating the songwriter split is 50/50. Two days later, a PDF from the label arrives listing the split as 60/40. If the agent follows chronological logic, it would simply overwrite the 50/50 with 60/40 because the PDF is the newest file. Instead, you configure the agent to flag the discrepancy. The agent sees the conflict, ignores the PDF's arrival date, and freezes the field as 'Open.' You then step in to determine if the label's PDF is a mistake or a new agreement.

Once a credit is frozen as 'Open,' it must remain locked until a named human decision-maker intervenes. In a professional release workflow, an AI agent cannot 'guess' which side of a dispute is correct, nor can it resolve a conflict based on a new file's arrival. The only way to move a field from 'Open' back to 'Approved' is through a manual override by the person assigned as the owner of that data.

Consider a conflict over a publisher split. Your agent detects that a new document contradicts the agreed-upon shares, so it freezes the field and tags it as 'Open.' Instead of attempting to merge the data, the agent sends a targeted notification to the release coordinator. The coordinator then investigates the discrepancy—perhaps by calling the publisher or checking a signed contract. Only after verifying the correct number does the coordinator manually enter the final percentage into the system. By doing this, the human provides the 'resolution trigger' that unlocks the record. You will implement this by ensuring your agents are configured to stop the workflow entirely for any 'Open' item, requiring a human signature before the metadata can proceed to the distributor.

The final step in establishing your lock is to verify that the agent actually stops the workflow when it encounters unresolved metadata. It is not enough to write the rules; you must prove that the AI will not 'hallucinate' a resolution or default to a last-in-wins logic under pressure. This is done through a stress test using a known conflict.

To run this test, upload a sample file—such as a CSV or PDF—that contains a credit explicitly designed to contradict an already approved fact in your release record. For example, if your approved record lists a producer as having a 20% split, upload a test file that claims they have 15%.

As the learner, you are looking for three specific outcomes: first, the agent must not update the percentage to 15%. Second, the record must automatically shift from 'Approved' to 'Open: Disputed.' Third, the agent must trigger an immediate notification to the manager. If the record updates silently, the lock has failed. By confirming these triggers, you ensure that no disputed credit can ever accidentally reach the distributor.