PolicyEngine · MaxGhenis · Aug 23, 2026 · Aug 22, 2026 · Aug 22, 2026 · Aug 22, 2026
diff --git a/annotations/us_full_run_20260612_policyengine_4_16_1_populace/us_audit_row_annotations.csv b/annotations/us_full_run_20260612_policyengine_4_16_1_populace/us_audit_row_annotations.csv
diff --git a/app/src/components/Hero.tsx b/app/src/components/Hero.tsx
@@ -2,7 +2,7 @@ import { DEFAULT_VERSION_ID } from "../lib/dataVersionsRuntime";
 import type { BenchData, CountryCode } from "../types";
 import SiteHeader, { type HeaderNavItem } from "./SiteHeader";
 
-const SNAPSHOT_DATE_LABEL = "Snapshot 2026-08-17";
+const SNAPSHOT_DATE_LABEL = "Snapshot 2026-08-22";
 
 export default function Hero({
   selectedView,

diff --git a/app/src/data.artifact.json b/app/src/data.artifact.json
@@ -1,9 +1,9 @@
 {
   "version": 1,
   "repo": "PolicyEngine/policybench",
-  "tag": "dashboard-data-20260817",
+  "tag": "dashboard-data-20260822",
   "asset": "dashboard-data.json",
-  "url": "https://github.com/PolicyEngine/policybench/releases/download/dashboard-data-20260817/dashboard-data.json",
-  "sha256": "a71de04ab9b3aa6d20e99fb8c7f90ec60ce4fdb708e2546bfb68ddbd968409fe",
-  "bytes": 80150275
+  "url": "https://github.com/PolicyEngine/policybench/releases/download/dashboard-data-20260822/dashboard-data.json",
+  "sha256": "b883ec669d510ea29c9c18f18c30030c5bbd29f770bcd90d257779940929a895",
+  "bytes": 85042876
 }
diff --git a/app/src/data.versions.json b/app/src/data.versions.json
@@ -5,7 +5,7 @@
     {
       "id": "1.1",
       "label": "1.1",
-      "description": "Corrected ground truth (policyengine-us 1.755.4); identical forced-tool request for every model - 30 models",
+      "description": "Corrected ground truth (policyengine-us 1.755.4); identical forced-tool request for every model - 32 models",
       "snapshotLabel": null,
       "artifact": {
         "pointer": "live"

diff --git a/app/src/modelMeta.ts b/app/src/modelMeta.ts
@@ -13,6 +13,8 @@ export const MODEL_ORDER = [
   "claude-haiku-4.5",
   "grok-4.3",
   "grok-4.5",
+  "grok-4.6",
+  "ox-alpha",
   "grok-build-0.1",
   "gpt-5.6-sol",
   "gpt-5.6-terra",
@@ -47,6 +49,8 @@ export const MODEL_LABELS: Record<string, string> = {
   "claude-sonnet-4.6": "Claude Sonnet 4.6",
   "grok-4.3": "Grok 4.3",
   "grok-4.5": "Grok 4.5",
+  "grok-4.6": "Grok 4.6",
+  "ox-alpha": "Ox Alpha (preview)",
   "grok-build-0.1": "Grok Build 0.1",
   "gpt-5.6-sol": "GPT-5.6 Sol",
   "gpt-5.6-terra": "GPT-5.6 Terra",
@@ -108,6 +112,8 @@ export const MODEL_RELEASE_DATES: Record<string, string> = {
   "gpt-5.6-luna": "2026-07-09",
   "grok-4.3": "2026-04-17",
   "grok-4.5": "2026-07-08",
+  "grok-4.6": "2026-08-12",
+  "ox-alpha": "2026-08-21",
   "grok-build-0.1": "2026-05-29",
   "deepseek-v4-pro": "2026-04-24",
   "deepseek-v4-flash": "2026-04-24",

diff --git a/paper/snapshot/20260501/manifest.json b/paper/snapshot/20260501/manifest.json
@@ -1,7 +1,7 @@
 {
   "audit_annotation_artifacts": {
     "files": {
-      "us_audit_row_annotations.csv": "9225b168d63ab233a374823587d128efd810ff66b17fe9984538aade2f09ee40",
+      "us_audit_row_annotations.csv": "f4b89cdb075d1725e4f1808a73e36c3eb79c72bc851d2553083f6688da2f5b7c",
       "us_case_notes.csv": "7e340f1dc6e22e2bbae18fd92579eab1d886918725377d3bf28be63f08ef5030",
       "us_case_reference_explanations.csv": "a3fc7504dcf3d8abe11e98c2228c786a5171a825640b980835e1e210d3afa198"
     },
@@ -22,11 +22,11 @@
   ],
   "live_dashboard_artifact": {
     "asset": "dashboard-data.json",
-    "bytes": 80150275,
+    "bytes": 85042876,
     "derivation": "At freeze time the live artifact equals the frozen published_dashboard_artifact: the combined export of the source-run data.json listed under source_run_artifacts. Annotation-class republishes may advance this entry ahead of the frozen pin without changing any score.",
-    "sha256": "a71de04ab9b3aa6d20e99fb8c7f90ec60ce4fdb708e2546bfb68ddbd968409fe",
-    "tag": "dashboard-data-20260817",
-    "url": "https://github.com/PolicyEngine/policybench/releases/download/dashboard-data-20260817/dashboard-data.json"
+    "sha256": "b883ec669d510ea29c9c18f18c30030c5bbd29f770bcd90d257779940929a895",
+    "tag": "dashboard-data-20260822",
+    "url": "https://github.com/PolicyEngine/policybench/releases/download/dashboard-data-20260822/dashboard-data.json"
   },
   "live_dashboard_note": "The live dashboard payload is a published release asset; the committed pointer app/src/data.artifact.json must reference the artifact pinned under live_dashboard_artifact. The separate published_dashboard_artifact freezes the combined export of the source run data.json files listed under source_run_artifacts. A later publication may advance the live entry without changing the frozen pin.",
   "model_response_date": "2026-06-12 to 2026-08-17",
@@ -135,4 +135,4 @@
   "source_run_labels": {
     "us": "us_full_run_20260612_policyengine_4_16_1_populace"
   }
-}
+}
diff --git a/policybench/config.py b/policybench/config.py
@@ -43,6 +43,7 @@
     "claude-haiku-4.5": "claude-haiku-4-5-20251001",
     "grok-4.3": "xai/grok-4.3",
     "grok-4.5": "xai/grok-4.5",
+    "grok-4.6": "xai/grok-4.6",
     "grok-build-0.1": "xai/grok-build-0.1",
     **GPT_56_MODELS,
     "gpt-5.5": "gpt-5.5",
@@ -52,6 +53,9 @@
     "gemini-3.5-flash": "gemini/gemini-3.5-flash",
     "gemini-3.6-flash": "gemini/gemini-3.6-flash",
     "gemini-3.7-flash": "gemini/gemini-3.7-flash",
+    # Cloaked OpenRouter preview listed 2026-08-21 (free window); maker
+    # unconfirmed. On the board as a labeled preview row.
+    "ox-alpha": "openrouter/stealth/ox-alpha",
     "gemini-3-flash-preview": "gemini/gemini-3-flash-preview",
     "gemini-3.1-flash-lite-preview": "gemini/gemini-3.1-flash-lite-preview",
     "deepseek-v4-pro": "deepseek/deepseek-v4-pro",

diff --git a/policybench/model_cards.py b/policybench/model_cards.py
@@ -159,6 +159,20 @@ class ModelCard:
             "tail scenarios."
         ),
     ),
+    "xai/grok-4.6": ModelCard(
+        litellm_id="xai/grok-4.6",
+        answer_contract="tool",
+        request_timeout_seconds=600,
+        thinking_budget=True,
+        expected_cost_per_scenario_usd=0.07,
+        notes=(
+            "Onboarded 2026-08-20: forced tool contract passed 3/3 and "
+            "16/16 whole-scenario. Heavy reasoner — the 3-variable probe "
+            "spent 26,260 tokens over 391s, so the timeout is 600s and the "
+            "thinking-class budget applies. Released 2026-08-12 at $2/$6 "
+            "per 1M (litellm map). The 100-scenario run cost $8.70."
+        ),
+    ),
     "deepseek/deepseek-v4-pro": ModelCard(
         litellm_id="deepseek/deepseek-v4-pro",
         answer_contract="json",
@@ -211,6 +225,21 @@ class ModelCard:
             "2027-01-01."
         ),
     ),
+    "openrouter/stealth/ox-alpha": ModelCard(
+        litellm_id="openrouter/stealth/ox-alpha",
+        answer_contract="tool",
+        request_timeout_seconds=600,
+        thinking_budget=True,
+        provider_max_completion_tokens=131_072,
+        expected_cost_per_scenario_usd=0.0,
+        notes=(
+            "Cloaked OpenRouter preview listed 2026-08-21 (free window); "
+            "maker unconfirmed. Board row labeled as a preview: public and "
+            "callable under the identical request, dated by its listing. "
+            "Tool contract passed the 2026-08-20 gauntlet 3/3 and 16/16 "
+            "after the thinking-class budget."
+        ),
+    ),
     "openrouter/moonshotai/kimi-k3": ModelCard(
         litellm_id="openrouter/moonshotai/kimi-k3",
         answer_contract="json",

diff --git a/sensitivity/ox-alpha-2026-08.md b/sensitivity/ox-alpha-2026-08.md
@@ -0,0 +1,73 @@
+# Ox Alpha preview run (August 2026)
+
+`stealth/ox-alpha` is a cloaked model OpenRouter listed on 2026-08-21 with
+no maker named, a 1M-token context, and a free window of about a week.
+Community fingerprinting points to a Zhipu GLM lineage. Nothing is
+confirmed. We ran it through the full benchmark because free previews of
+unreleased frontier models are worth measuring while they exist.
+
+It sits on the board as a labeled preview row. It is publicly callable
+under the identical request, so it meets the board's inclusion rule; what
+it lacks is a named maker and durability. The row is dated by its
+OpenRouter listing (2026-08-21). If the model later ships under its own
+name, that model gets onboarded normally, with this run as its early read.
+
+## Result
+
+Canonical v1 condition: identical whole-scenario prompt, forced answer
+tool call, no reasoning parameters, 100 households, 18 outputs.
+
+| | exact | would rank | within 1% | parsed |
+|---|---|---|---|---|
+| ox-alpha | **84.2** | #4 of 32 | 86.4 | 1,984 / 1,984 |
+
+Only GPT-5.6 Sol (88.7), Kimi K3 (86.2), and GPT-5.6 Luna (84.5) score
+higher on the board. Inference cost was $0 in the free window; the
+tool contract passed the gauntlet 3/3 and 16/16 after a thinking-class
+completion budget (the 384-token probe budget starved it, the known
+reasoning-model trap).
+
+## Per program
+
+Within-$1 rates per output. GLM-5.2, the latest released model in the
+suspected lineage, appears for reference only.
+
+| program | Ox Alpha | GLM-5.2 |
+|---|---|---|
+| local_income_tax | 100.0 | 94.0 |
+| person_early_head_start_eligible | 100.0 | 89.5 |
+| person_head_start_eligible | 100.0 | 89.5 |
+| person_wic_eligible | 100.0 | 92.7 |
+| person_medicare_eligible | 99.4 | 89.3 |
+| reduced_price_school_meals_eligible | 99.0 | 94.0 |
+| self_employment_tax | 99.0 | 91.0 |
+| tanf | 99.0 | 94.0 |
+| free_school_meals_eligible | 98.0 | 93.0 |
+| ssi | 98.0 | 90.0 |
+| person_medicaid_eligible | 93.8 | 85.9 |
+| person_chip_eligible | 92.7 | 93.2 |
+| payroll_tax | 87.0 | 73.0 |
+| federal_refundable_credits | 84.0 | 79.0 |
+| snap | 79.0 | 76.0 |
+| state_refundable_credits | 79.0 | 72.0 |
+| federal_income_tax_before_refundable_credits | 63.0 | 43.0 |
+| state_income_tax_before_refundable_credits | 55.0 | 45.0 |
+
+On the 20 SNAP cases where the household is owed benefits, Ox Alpha
+answers exactly $0 seven times and gets none exact — the same failure
+shape as the rest of the board.
+
+## Reproducing
+
+```
+python -m policybench.cli run \
+  --model ox-alpha \
+  --scenario-manifest paper/snapshot/20260501/us_scenarios.csv \
+  --run-dir results/local/oxalpha/run \
+  --budget-usd 5 --max-workers 4
+```
+
+The model card (`openrouter/stealth/ox-alpha`) pins the tool contract, a
+600s timeout, and the thinking-class budget. Predictions are attached to
+the `dashboard-data-20260817` release as
+`preview-ox-alpha-predictions.csv.gz`.
diff --git a/tests/test_model_cards.py b/tests/test_model_cards.py
@@ -41,6 +41,8 @@
     "gpt-5.4-nano": ("tool", None, 20, 4_096),
     "xai/grok-4.3": ("tool", None, 420, 4_096),
     "xai/grok-4.5": ("tool", None, 420, 16_384),
+    "xai/grok-4.6": ("tool", None, 600, 16_384),
+    "openrouter/stealth/ox-alpha": ("tool", None, 600, 16_384),
     "xai/grok-build-0.1": ("tool", None, 420, 4_096),
     "deepseek/deepseek-v4-pro": ("json", None, 300, 16_384),
     "deepseek/deepseek-v4-flash": ("json", None, 300, 16_384),