From b0a26663ace37bef1c6f7c1f5be9eca9f4ad013c Mon Sep 17 00:00:00 2001
From: relh <relh@softmax.com>
Date: Sun, 4 Oct 2026 08:39:11 -1000
Subject: [PATCH] Keep game-owned scoring evidence distinct from sampled tokens

Co-authored-by: GPT-6 Codex <noreply@openai.com>
---
 src/bitworld/decision_trajectory.nim | 11 +++++++
 tests/test_decision_trajectory.nim   | 44 ++++++++++++++++++++++++++++
 2 files changed, 55 insertions(+)

diff --git a/src/bitworld/decision_trajectory.nim b/src/bitworld/decision_trajectory.nim
index 918ccb3..8c5329c 100644
--- a/src/bitworld/decision_trajectory.nim
+++ b/src/bitworld/decision_trajectory.nim
@@ -22,6 +22,7 @@ type
     model*: Option[string]
     origin*: AttemptOrigin
     inferenceMode*: InferenceMode
+    actionEvidence*: Option[JsonNode] ## Game-owned scoring provenance; never player-supplied.
     prompt*, request*, response*, rawResponse*, parsedAction*, decoder*: JsonNode
     accepted*: bool
     platformCallId*, rejectionReason*, modelIdentity*, tokenizerIdentity*: Option[string]
@@ -244,9 +245,19 @@ proc recordDecision*(trajectory: DecisionTrajectory, decisionId, seat: string,
       selectedPrompt = attempt.prompt
       if not attempt.accepted or attempt.parsedAction != executedAction:
         raise newException(ValueError, "selected proposal differs from executed action")
+    if attempt.actionEvidence.isSome:
+      let action = attempt.actionEvidence.get()
+      if action.kind != JObject or action["protocol"].kind != JString or
+          action["protocol"].getStr().len == 0:
+        raise newException(ValueError, "game action evidence requires a protocol-tagged object")
+      if attempt.inferenceMode != imCandidate or attempt.response.kind != JNull or
+          attempt.sampledTokenIds.isSome or attempt.behaviorLogprobs.isSome:
+        raise newException(ValueError, "scoring evidence cannot claim generated text or sampled tokens")
     encoded.add(%*{
       "attempt_id": attempt.attemptId, "policy": attempt.policy,
       "inference_mode": $attempt.inferenceMode,
+      "action_evidence": (if attempt.actionEvidence.isSome:
+        copy(attempt.actionEvidence.get()) else: newJNull()),
       "origin": originName(attempt.origin), "model": jsonOption(attempt.model),
       "prompt": attempt.prompt, "request": attempt.request,
       "response": attempt.response, "raw_response": attempt.rawResponse,
diff --git a/tests/test_decision_trajectory.nim b/tests/test_decision_trajectory.nim
index ce317fc..559dfdb 100644
--- a/tests/test_decision_trajectory.nim
+++ b/tests/test_decision_trajectory.nim
@@ -285,3 +285,47 @@ suite "private authoritative decision trajectories":
     check readAttemptEvidence(wire).responseReaderJoined.isNone
     wire["response_reader_joined"] = %"joined"
     expect ValueError: discard readAttemptEvidence(wire)
+
+  test "game-owned scoring survives original failure and selected action joins":
+    let record = episode()
+    var rejected = newDecisionAttempt("rank-rejected", "policy", aoModel, imCandidate)
+    rejected.actionEvidence = some(%*{"protocol": "fixture.score.v1",
+      "scoring_result": {"kind": "rejected", "fault": "provider_error"}})
+    rejected.rejectionReason = some("original provider failure")
+    var selected = newDecisionAttempt("rank-scored", "policy", aoModel, imCandidate)
+    selected.actionEvidence = some(%*{"protocol": "fixture.score.v1",
+      "scoring_result": {"kind": "scored", "selected_index": 0}})
+    selected.parsedAction = %*{"move": 1}
+    selected.accepted = true
+    record.recordDecision("rank", "0", %*{}, @[rejected, selected],
+      some("rank-scored"), selected.parsedAction, asAccepted)
+    selected.actionEvidence.get()["scoring_result"]["selected_index"] = %9
+    record.finish(esCompleted, %*{}, %*[1])
+    let event = parseJson(record.eventsJsonl().splitLines()[0])
+    check event["attempts"].len == 2
+    check event["attempts"][0]["action_evidence"]["scoring_result"]["kind"].getStr() == "rejected"
+    check event["attempts"][1]["action_evidence"]["scoring_result"]["selected_index"].getInt() == 0
+    check event["attempts"][1]["response"].kind == JNull
+    check event["executed_action"] == event["attempts"][1]["parsed_action"]
+
+  test "scoring cannot acquire generation or sampled token provenance":
+    for fault in ["mode", "text", "tokens", "probabilities"]:
+      var attempt = newDecisionAttempt("rank", "policy", aoModel, imCandidate)
+      attempt.actionEvidence = some(%*{"protocol": "fixture.score.v1"})
+      case fault
+      of "mode": attempt.inferenceMode = imTextAction
+      of "text": attempt.response = %"generated-looking text"
+      of "tokens": attempt.sampledTokenIds = some(@[1])
+      else: attempt.behaviorLogprobs = some(@[-0.1])
+      expect ValueError:
+        episode().recordDecision("rank", "0", %*{}, @[attempt], none(string),
+          %*{"move": 1}, asFallback, fallbackOrigin = some("actual-scripted"))
+
+  test "private players cannot inject game-owned scoring evidence":
+    var attempt = newDecisionAttempt("rank", "policy", aoModel, imCandidate)
+    attempt.actionEvidence = some(%*{"protocol": "fixture.score.v1"})
+    let wire = attempt.attemptEvidenceJson()
+    check not wire.hasKey("action_evidence")
+    wire["action_evidence"] = %*{"protocol": "fixture.score.v1"}
+    expect ValueError:
+      discard readAttemptEvidence(wire)