Route feedback to the first broken contract

A user sees an incorrect breakout explanation and reports that the model is wrong. That description identifies the experience. Assigning the repair requires finding where the evidence first diverged from the expected contract. We propose a routing procedure that follows the saved observation through data, tool output, harness assembly and model response.

The distinction has a current DXAP anchor. The October 3 releases describe completed-candle statistics, preceding breakout baselines and explicit data coverage. They also describe a fix for string-encoded symbol arrays produced by some models. Similar visible failures can therefore arise from candle content, argument handling or interpretation of an otherwise valid result.

Follow one observation through four checkpoints

Consider an illustrative report: “The answer says the latest close broke the preceding high, but the table shows otherwise.” Preserve the selected symbol, timeframe and observation cutoff. Then inspect four checkpoints using the same saved input.

First, compare the source candles with their declared coverage and completion status. If a provisional candle entered a completed-candle window, the datafeed contract failed. Next, recompute the statistic from the accepted candles. If those candles are correct but the returned baseline includes the comparison candle, the numerical tool owns the mismatch.

Third, inspect what reached the model. A correct tool result attached to the wrong symbol or combined with an older observation time points toward harness assembly. Finally, compare the answer with the exact result delivered. If the model received the correct close and baseline but reversed their relationship, the response is the candidate model failure.

These are diagnostic categories for the proposed procedure, rather than a claim about DXAP's internal team structure. “First broken contract” means the earliest independently verified mismatch along this path. A later component may need a defensive check even when it did not originate the defect.

Measure the handoff, including returns

Start the routing clock when a report contains enough information to attempt reproduction. Stop it when the responsible engineer accepts a reproducible fixture and the disputed contract. Record the earlier intake delay separately: a fast technical handoff can conceal days spent obtaining the selected turn.

For an illustrative pilot, twelve complete reports yield eight accepted first assignments, three returned assignments and one unresolved case. First-assignment acceptance is eight of twelve, or about 67%. The unresolved report stays in the denominator. Report return reasons and time to accepted ownership alongside that percentage; a team can inflate apparent accuracy by leaving hard cases unassigned.

Review activity supplies the user-facing sequence from selected turn to action details and execution views. The routing procedure would add internal contract checks to that documented inspection path. It would preserve an unresolved designation whenever required evidence is absent.

The repair owner's closing artifact should be a replayable input with the expected boundary behavior. Model training becomes one available destination after diagnosis. A tool-parser fix can instead become a tool regression case, while a stale-input defect becomes a harness timing test. The user contributes the mismatch; the investigation determines which component can remove it.