The Multi-Step Record Format

A bank denies a credit-card application in 2.3 seconds. Three weeks later, the applicant's attorney asks how.

This is the request a model card cannot answer.

The application went to an agent. The agent pulled credit-bureau data. It called an income-verification API. It checked the applicant against a fraud table. It branched on a thin-file rule. It chose between two scoring sub-models. It returned a denial.

What the bank logged. The input. The output. The timestamp.

What the bank did not record. The decision graph the agent walked between input and output. The trajectory.

Five primitives are what the trajectory has to contain.

One. The tool-call ledger.

Every call the agent made, in order. Inputs. Outputs. Timestamps. Return codes. Hash-chained, so any post-hoc edit produces a verifiable break.

For the loan denial. The credit-bureau pull. The income-verification API. The fraud-table check. The scoring sub-model invocation. Four calls in sequence, signed at the moment each one happened.

If a call is in the ledger, it happened. If a call is not in the ledger, it did not happen. The ledger is the spine.

Most logging today captures the call. It does not capture the order, the hash, or the binding to the run.

Two. The state vector.

What the agent knew at the moment of each tool call, before the call and after the call. Scratchpad contents. Retrieved context. Prior tool results. The intermediate conclusion the model had reached.

For the loan denial. At the moment the agent invoked the thin-file branch, the state vector showed three things. The credit bureau had returned three records. The income API had returned a 401 retry. The model had already labeled the applicant as moderate-risk.

The branch decision in primitive three was made against that specific state. Without the state vector, the branch has no anchor. With it, the next reviewer reads the same context the model read.

Three. The branch trace.

The path that was taken and the paths that were not.

For the loan denial. The agent chose the thin-file scoring model. The standard model was the alternative considered. The reason the agent gave for the choice, if it gave one, is in the trace.

This is the primitive that answers the attorney's question. Could this run have gone differently. The trace is the answer.

Four. The model-identity binding.

A cryptographic signature binding the trajectory to a specific model at the moment of the decision. Weights hash. Version. Adapter identifiers. Decoding parameters.

For the loan denial. Three weeks later, the attorney asks which model decided. Without the binding, the answer is, we believe it was version 3.2, but the production runtime has retrained twice since then. With the binding, the answer is a hash, signed at the moment of the act.

The binding is what makes primitive five possible.

Five. The replay primitive.

A bundle that re-executes the run against the frozen weights and verifies the trajectory was produced by the model the binding names.

For the loan denial. Replay against the version 3.2 weights with the same inputs, the same state vector, and the same tool returns. Confirm the trajectory the bank produced is the trajectory the model produces under those conditions.

Without replay, the trajectory is a story the system told about itself. With replay, the trajectory is an artifact a second party can verify.

What makes the five admissible.

Four properties hold across the primitives. Tamper evidence lives in the hash chain. Independent verifiability lives in the replay. Time stamping lives in the signature on each ledger entry. Model attribution lives in the binding.

A record with the primitives but no replay is a log. A record with the primitives and the replay is an artifact. The difference is what the 2029 examiner will ask for.

What we are building.

Wayfinder Systems Group is building one answer. A runtime governance layer that records every tool call, every state transition, every branch decision, and binds each one to a model-identity signature at the moment it happens. The schema is the format. The chain is the integrity primitive. The replay is the admissibility test. Patents held in The Wayfinder Trust. We call her Velma.