Retain the Market Inputs Needed to Reconstruct an Agent Decision

By DX Research Group · · Market data

A deletion fixture shows why prompt summaries and hashes cannot substitute for missing source payloads.

Retain the exact input bundle a decision used, together with the source records needed to interpret it, for the declared research window. A later API request may return corrected values, a different universe, or no record at all. We would design retention around the questions the trace must answer and the permissions governing each dataset.

FRED's vintage documentation distinguishes today's knowledge of the past from information available in an earlier period. That source illustrates a reconstruction problem; our retention procedure applies separately to each collector and its source terms.

A digest survives while the evidence disappears

Assume a synthetic agent turn that receives a JSON bundle containing three quotes and one news excerpt. The trace stores the bundle digest and a rationale saying “depth supports entry.” Thirty days later, the raw bundle is deleted but its digest remains.

The digest can verify a candidate copy if someone still has it. It cannot reconstruct the three quotes or the excerpt. The rationale supplies the agent's assertion, with no recoverable depth calculation. A reviewer can confirm that a reference existed while the factual basis of the claim remains unavailable.

Retained artifactQuestion it can answer
Digest onlyDoes a supplied copy match?
Rendered input bundleWhat did the agent receive?
Source payload plus adapter versionHow was the feature derived?
Decision and settlement traceWhat action and outcome followed?

These layers support different kinds of review. Retaining one does not imply that all others are recoverable.

Choose retention by dependency

Start with a representative decision and walk backward from its material claims. Preserve the rendered input, feature values, and source references required to reproduce them. Include normalization settings when decimal handling or aggregation changes the result. Record deletion dates and whether an artifact can legally and operationally be reacquired.

Raw social content and proprietary feeds may have restrictions that differ from public aggregate data. Apply those terms explicitly. Where raw retention is restricted, document the resulting reconstruction limit and retain permitted metadata or derived fields. A proposed evaluation should respect that limit instead of silently assuming permanent source access.

Store a compact manifest per decision that identifies all dependencies and their retention status. A missing dependency should produce a partial reconstruction result with the affected claims named. Keep sensitive account records access-controlled and separate from public research extracts.

Test retrieval before a review depends on it

A proposed monthly local audit would sample saved turns, restore their bundles, and recompute a few material features. It would report complete, partial, and unavailable reconstructions, including failures caused by expired data. This is a suggested procedure, with no claim that such an audit has already run.

Our minimum trace schema defines the joins to preserve. Our continuous-record companion demonstrates why evidence scope belongs beside a result. The retention contribution here is a dependency-based recovery test. A successful recovery establishes inspectability of selected decisions; it leaves prediction quality, policy compliance, and return interpretation to separate analyses.

Sources

Related field notes