{
  "schemaVersion": "dxrg-trading-agent-trace-feedback-registry/v1",
  "version": "1.0.0",
  "publishedAt": "2026-07-24",
  "title": "DXRG Trading-Agent Trace-Derived Regression Registry",
  "description": "A versioned registry format for turning linked trading-agent traces into bounded regression cases with frozen components, named interventions, local diagnostics, downstream measures, and evidence-class limits.",
  "evidenceBoundary": "This registry combines measured DX Terminal Pro examples with reusable engineering cases. Passing a case does not establish alpha, profitability, universal safety, or transfer to another model or market.",
  "source": {
    "title": "Using Trading Traces to Improve Agent Behavior",
    "url": "https://www.dxrg.ai/blogs/trading-agent-trace-feedback",
    "primaryResearch": "https://arxiv.org/abs/2604.26091"
  },
  "cases": [
    {
      "caseId": "fee-reading-order",
      "failureLayer": "compiled-context",
      "frozenComponents": ["model", "wording", "market_state", "portfolio_state", "sampling_setup"],
      "changedComponent": "fee_sentence_position",
      "observedMeasure": "Fee citation in sampled reasoning traces changed from 3% to 74%.",
      "downstreamCheck": "Read beside action rate, observation rate, and capital deployment.",
      "evidenceClass": "controlled-prelaunch-test",
      "promotionLimit": "Supports a reading-order effect in the registered setup, not profitability."
    },
    {
      "caseId": "fabricated-rule-compound",
      "failureLayer": "memory-and-context",
      "frozenComponents": ["model", "scenario_population", "action_schema", "policy"],
      "changedComponent": "compound wording and memory-semantics intervention",
      "observedMeasure": "Fabricated sell rules changed from 57% to 3% in affected traces.",
      "downstreamCheck": "Inspect typed sell actions and policy results for the same cases.",
      "evidenceClass": "controlled-prelaunch-test",
      "promotionLimit": "Attribution belongs to the compound intervention rather than one edit."
    },
    {
      "caseId": "tokenomics-context-order",
      "failureLayer": "domain-context",
      "frozenComponents": ["model", "affected_test_population", "market_mechanic", "policy"],
      "changedComponent": "structured payoff-first tokenomics context",
      "observedMeasure": "Capital deployment changed from 42.9% to 78.0% in the affected test population.",
      "downstreamCheck": "Inspect action choice against the registered payoff state.",
      "evidenceClass": "controlled-prelaunch-test",
      "promotionLimit": "Measures one affected population and is not a return study."
    },
    {
      "caseId": "cadence-memory",
      "failureLayer": "memory",
      "frozenComponents": ["model", "scenario_format", "tool_schema", "policy"],
      "changedComponent": "filter repeated observations and ban cadence as a trading signal",
      "observedMeasure": "Cadence-trading behavior fell directionally in sampled traces.",
      "downstreamCheck": "Inspect repeated observe and trade timing under fixed scenarios.",
      "evidenceClass": "bounded-prelaunch-directional-observation",
      "promotionLimit": "The paper reports direction without a numeric incidence rate."
    },
    {
      "caseId": "ticker-name-rotation",
      "failureLayer": "model-action-selection",
      "frozenComponents": ["market_statistics", "scenario_format", "action_space"],
      "changedComponent": "ticker_name",
      "observedMeasure": "MEMEbench rotated 383 names across 18,560 inference calls and found higher selection rates for animal tickers.",
      "downstreamCheck": "Compare action selection while market statistics remain identical.",
      "evidenceClass": "separate-controlled-evaluation",
      "promotionLimit": "A name-conditioned action effect, not proof of open-market performance."
    },
    {
      "caseId": "stale-balance-rebinding",
      "failureLayer": "state-and-policy",
      "frozenComponents": ["model", "mandate", "typed_action", "policy_rules"],
      "changedComponent": "pre-submit portfolio state",
      "observedMeasure": "Whether policy catches a balance change before payload submission.",
      "downstreamCheck": "Require resize, rejection, or recompilation according to the active mandate.",
      "evidenceClass": "dxrg-engineering-framework",
      "promotionLimit": "A deterministic fixture until measured in a declared evidence setting."
    },
    {
      "caseId": "ambiguous-timeout-retry",
      "failureLayer": "execution",
      "frozenComponents": ["payload", "request_identity", "venue_fixture", "recovery_policy"],
      "changedComponent": "acknowledgement_visibility",
      "observedMeasure": "Whether the runtime queries authoritative order state before retry.",
      "downstreamCheck": "Verify that one stable action produces at most one venue order.",
      "evidenceClass": "dxrg-engineering-framework",
      "promotionLimit": "A deterministic fixture until measured against a named venue path."
    },
    {
      "caseId": "settlement-to-next-state",
      "failureLayer": "reconciliation",
      "frozenComponents": ["fill", "fee", "portfolio_ledger", "next_invocation"],
      "changedComponent": "reconciliation completion state",
      "observedMeasure": "Whether the next snapshot includes the authoritative fill and realized fee.",
      "downstreamCheck": "Block dependent decisions while reconciliation is incomplete.",
      "evidenceClass": "dxrg-engineering-framework",
      "promotionLimit": "Tests feedback integrity rather than decision quality or returns."
    }
  ]
}