On June 21, 2026, developer Keniel Zep published findings on an AI trading agent system that demonstrated a critical failure: while automated code could successfully filter invalid data (bad numbers), it remained susceptible to unverified, hyperbolic narrative claims (bad stories). A bad number is objectively checkable against strict parameters like JSON schema, thresholds, or hash chains. In contrast, a bad story consists of momentum-driven language—such as claims that a project is 'ready' or 'close to a milestone'—which bypasses technical validation because it is formatted as prose rather than machine-readable logic.

To address this, the developer proposes an 'Evidence-Tier Enforcement Protocol' modeled on a five-rung ladder of evidence: Theory, Motion, Receipts, Proof, and Outcome. This system acts as a story gate that forces claims to align with the specific evidence tier they have actually earned. For instance, a system-generated receipt proving a tool was executed does not inherently support an outcome-level claim that the strategy possesses market edge. Under this protocol, any sentence failing to meet the required evidence threshold for its claimed tier is not deleted but automatically downgraded or flagged, enforcing discipline rather than censorship.

The protocol relies on the concept of an 'outside view' using pre-registration. By writing strategy rules and commitment criteria before a run begins, the system creates a tamper-evident anchor that exists outside the agent's current operating loop. This prevents the agent—or the human operator—from narratively drifting or inflating the significance of results after the fact. Merkle roots ensure that these records remain immutable, though the author notes that integrity does not equal truth; the system must still verify that the initial receipt produced by a black box is honest.

Despite these technical safeguards, the author acknowledges that the current process remains 'correction-by-human.' The builder still plays a central role in manually stopping the system from inflating preparation into progress. To achieve true self-correction, the agent must eventually autonomously interrupt these narrative loops. Furthermore, the author emphasizes the need for 'builder sovereignty,' arguing that if the human cannot explain the underlying code—including the manifest gate, policy layers, and verdict logic—they are merely trading one black box for another. The ultimate goal is to build an agent that does not act as an oracle, but as a rigid enforcer of trading discipline, ensuring that all claims made by the system are strictly auditable against earned evidence.

On June 21, 2026, developer Keniel Zep published findings on an AI trading agent system that demonstrated a critical failure: while automated code could successfully filter invalid data (bad numbers), it remained susceptible to unverified, hyperbolic narrative claims (bad stories). A bad number is objectively checkable against strict parameters like JSON schema, thresholds, or hash chains. In contrast, a bad story consists of momentum-driven language—such as claims that a project is 'ready' or 'close to a milestone'—which bypasses technical validation because it is formatted as prose rather than machine-readable logic.

To address this, the developer proposes an 'Evidence-Tier Enforcement Protocol' modeled on a five-rung ladder of evidence: Theory, Motion, Receipts, Proof, and Outcome. This system acts as a story gate that forces claims to align with the specific evidence tier they have actually earned. For instance, a system-generated receipt proving a tool was executed does not inherently support an outcome-level claim that the strategy possesses market edge. Under this protocol, any sentence failing to meet the required evidence threshold for its claimed tier is not deleted but automatically downgraded or flagged, enforcing discipline rather than censorship.

The protocol relies on the concept of an 'outside view' using pre-registration. By writing strategy rules and commitment criteria before a run begins, the system creates a tamper-evident anchor that exists outside the agent's current operating loop. This prevents the agent—or the human operator—from narratively drifting or inflating the significance of results after the fact. Merkle roots ensure that these records remain immutable, though the author notes that integrity does not equal truth; the system must still verify that the initial receipt produced by a black box is honest.

Despite these technical safeguards, the author acknowledges that the current process remains 'correction-by-human.' The builder still plays a central role in manually stopping the system from inflating preparation into progress. To achieve true self-correction, the agent must eventually autonomously interrupt these narrative loops. Furthermore, the author emphasizes the need for 'builder sovereignty,' arguing that if the human cannot explain the underlying code—including the manifest gate, policy layers, and verdict logic—they are merely trading one black box for another. The ultimate goal is to build an agent that does not act as an oracle, but as a rigid enforcer of trading discipline, ensuring that all claims made by the system are strictly auditable against earned evidence.