Deduplicate Traces by Economic Intent Without Losing Retries

Trace deduplication should preserve one economic intention and every attempt to carry it out. We would use explicit intent lineage rather than identical text or timestamps. Two identical orders can be deliberate separate decisions; two differently formatted requests can be retries of the same decision.

One intention, three transport records

In an illustrative fixture, a decision authorizes buying one unit. The first submission receives no response, the retry receives an acknowledgement, and a reconciliation worker later records the fill. There are three transport events, one intended order and one economic fill. Counting the records as three decisions inflates activity and can distort both failure rates and turnover.

Give the decision an intent identifier. Attach attempt identifiers beneath it, and retain the venue order and fill identifiers where available. Preserve raw event records even after creating the analytical deduplicated view. The audit should be reversible: a reviewer can expand an intent and see every retry and receipt.

OpenTelemetry's tracing API provides trace and span identity concepts. Our proposed economic-intent layer supplies a trading meaning above those transport spans. A trace identifier alone may cover several deliberate order decisions within one agent turn.

Identical text can still mean a second action

Extend the fixture with a later decision to buy another unit at the same limit price. Its JSON body matches the earlier order, yet its intent identifier and portfolio precondition differ. Collapsing by request-body hash would remove legitimate additional exposure. Conversely, a retry that refreshes an authentication field can change the serialized bytes while retaining the same authorized intent.

For a worked count, four submissions represent two intentions: three attempts for the first and one for the second. One intention fills and one expires. Submission success, intent completion and economic fills therefore need separate denominators. The audit can report four attempts, two intentions and one filled intent without choosing an artificial single activity count.

The operating-layer controls companion motivates tracing the mandate through settlement. The continuous record companion depends on reliable joins among turns and fills. This proposal addresses record identity, while the published historical results retain their stated definitions.

We would test duplicate delivery, delayed acknowledgement and a deliberate second order. Expected outcomes include unchanged economic quantity under duplicate delivery and increased authorized quantity under the second intent. Unknown venue identity should stay unresolved until reconciliation provides evidence.

The resulting table includes an intent row with child attempts, terminal venue state and reconciliation status. It supports counting model decisions without hiding infrastructure work. A deduplicated analytical extract should disclose its grouping rule and retain a route back to raw events, so later corrections to lineage can be evaluated without rebuilding the story from summaries.