Add gates.json — canonical release-gate thresholds for /lab model-compare (protoContent#399)

gates.json

@@ -0,0 +1,40 @@
+{
+  "version": "1.0",
+  "updated": "2026-07-05",
+  "description": "Release-gate thresholds for protoLabs model cards and the /lab model-compare delta semantics. A quant PASSES its gate when its regression vs baseline stays within max_regression. Delta rendering is gate-aware, NOT red-on-any-negative: beats_base (green), within_gate (muted chip -- the parity claim every model card ships), over_gate (red -- a real regression). Naive red-on-negative would contradict every card we publish.",
+  "direction": "higher_is_better",
+  "delta_convention": "delta = value - baseline. A row is within gate when (baseline - value) <= max_regression, i.e. delta >= -max_regression. beats_base when delta > 0; within_gate when -max_regression <= delta <= 0; over_gate when delta < -max_regression.",
+  "baseline_fields": ["baseline_bf16", "baseline_fp8_official"],
+  "baseline_note": "Gate rows carry ONE baseline field naming the reference the quant was gated against (bf16 for most; fp8_official when the shipped baseline was the official FP8, e.g. Agents-A1). Frontend: use whichever baseline_* key is present on the row.",
+  "gates": {
+    "claw": {
+      "max_regression": 0.03,
+      "unit": "score",
+      "metric": "mean_task_score",
+      "label": "claw (agentic)",
+      "suites": ["claw_paired", "claw_paired29", "claw", "claw_business", "claw_coding"]
+    },
+    "function_call": {
+      "max_regression": 0.02,
+      "unit": "rate",
+      "metric": "pass_rate",
+      "label": "function-call",
+      "suites": ["function_call"],
+      "note": "Published claim is '<= 2 percentage points'. pass_rate is stored as a 0-1 rate, so the threshold in row units is 0.02."
+    },
+    "reasoning": {
+      "max_regression": 0.05,
+      "unit": "score",
+      "metric": "mean_task_score",
+      "label": "reasoning",
+      "suites": ["reasoning_v2", "reasoning_v2_16k", "reasoning"]
+    },
+    "code": {
+      "max_regression": 0.05,
+      "unit": "score",
+      "metric": "mean_task_score",
+      "label": "code",
+      "suites": ["code_exec_v2", "code_spec_delta", "coding", "code_exec"]
+    }
+  }
+}