@smartmemory/compose 0.1.39-beta → 0.1.40-beta

package/contracts/judge-result.json

@@ -29,7 +29,7 @@
         },
         "stakes": {
           "enum": ["cheap", "default", "paranoid"],
-          "description": "Depth class set by caller. v1: cheap (T1 only), default (T1+T2). paranoid raises StakesNotAvailableError until v2."
+          "description": "Depth class set by caller. cheap (T1 only), default (T1+T2). paranoid (T1+T2, then a T3 adversary over every interpretive 'met'). STRAT-JUDGE-T3-READJAIL: the paranoid T3 is now a true cross-model jailed Codex adversary (macOS sandbox-exec read-jail confining reads to the staged turn tree) when available, else an honest probe-time degrade to the v1 in-process Claude cold-read. Per-predicate provenance is in predicates[].t3."
         },
         "predicates": {
           "type": "array",
@@ -50,7 +50,7 @@
               "id": { "type": "string" },
               "type": {
                 "enum": ["deterministic", "verified", "judged"],
-                "description": "Predicate type sets the tier ceiling. deterministic = T1 only. verified = T2. judged = T3 in v2, downgraded to T2 in v1."
+                "description": "Predicate type sets the tier ceiling. deterministic = T1 only. verified = T2. judged = T2, plus T3 cold-read adversary when stakes=paranoid (else T2-terminal; judge_kernel_meta.degraded_judged flags judged predicates that got no adversary)."
               },
               "statement": { "type": "string" },
               "verdict": {
@@ -97,13 +97,36 @@
                     "reason": { "type": "string" }
                   }
                 }
+              },
+              "t3": {
+                "type": ["object", "null"],
+                "description": "STRAT-JUDGE-T3-READJAIL. Populated ONLY when a T3 adversary was reached for this predicate (stakes='paranoid' AND T2='met'); null/absent means no adversary ran for this predicate (the absence is itself the honest signal). Descriptive of what actually ran, never intent. Authoritative T3 model identity is t3.model_id — top-level meta.model_id is the T1/T2 worker-lane identity only.",
+                "required": ["mode", "residual"],
+                "properties": {
+                  "mode": {
+                    "enum": ["codex_jailed", "codex_jailed_error", "claude_cold_fallback"],
+                    "description": "codex_jailed: true cross-model Codex adversary, OS read-jailed (macOS sandbox-exec) to the staged turn tree. claude_cold_fallback: probe-time degrade to in-process Claude cold-read (no OS jail; reasoning-isolation+ordering only). codex_jailed_error: jailed adversary failed post-launch — predicate resolved ambiguous, error surfaced, NEVER silently relabeled fallback."
+                  },
+                  "guarantee": {
+                    "type": ["string", "null"],
+                    "description": "os_read_jail+cross_model | reasoning_isolation+ordering | null (error)."
+                  },
+                  "model_id": {
+                    "type": ["string", "null"],
+                    "description": "The actual model that ran T3 (authoritative for the adversary lane)."
+                  },
+                  "residual": {
+                    "type": "string",
+                    "description": "The accepted residual for this mode, verbatim from the design Problem statement. Never rounded up to 'confined'."
+                  }
+                }
               }
             }
           }
         },
         "tier_disagreements": {
           "type": "array",
-          "description": "Cross-tier disagreements within a single judge call. Empty in v1 (single-tier-per-predicate after T1). Reserved for v2+ T2/T3 quorum resolution.",
+          "description": "Cross-tier disagreements within a single judge call. Populated when stakes=paranoid and T3 contradicts a T2 'met' (resolved to 'ambiguous'; T4 quorum deferred). Empty for cheap/default.",
           "items": {
             "type": "object",
             "required": ["predicate", "tiers", "resolution"],
@@ -142,7 +165,7 @@
             },
             "degraded_judged": {
               "type": "boolean",
-              "description": "True when a 'judged' predicate was downgraded to T2 because T3 is unavailable in v1."
+              "description": "True when a 'judged' predicate did not receive adversarial (T3) verification — i.e. stakes was cheap/default so T2 was terminal. False once T3 ran (stakes=paranoid)."
             }
           }
         }