thumbgate 1.27.20 → 1.28.1

package/scripts/agent-reward-model.js

@@ -11,6 +11,7 @@
 
 const path = require('node:path');
 const { loadEpisodes } = require('./session-episode-store');
+const { requireLearnedModelsEntitlement } = require('./entitlement');
 
 const HIGH_RISK_TAGS = new Set([
   'billing',
@@ -163,6 +164,10 @@ function isRlTuple(value) {
 }
 
 function buildPreferencePairs(episodes = [], options = {}) {
+  requireLearnedModelsEntitlement({
+    ...(options.entitlement || {}),
+    label: 'agent reward preference-pair generation',
+  });
   const tuples = episodes
     .map((episode) => episodeToRlTuple(episode, options))
     .sort((a, b) => a.reward.total - b.reward.total);
@@ -181,6 +186,10 @@ function buildPreferencePairs(episodes = [], options = {}) {
 }
 
 function rankGateCandidatesByReward(episodes = [], options = {}) {
+  requireLearnedModelsEntitlement({
+    ...(options.entitlement || {}),
+    label: 'agent reward gate ranking',
+  });
   const minOccurrences = Math.max(1, Number(options.minOccurrences || 2));
   const buckets = new Map();
 
@@ -305,6 +314,10 @@ function allocateTestTimeCompute(action = {}) {
 }
 
 function buildRewardReport(episodes = [], options = {}) {
+  requireLearnedModelsEntitlement({
+    ...(options.entitlement || {}),
+    label: 'agent reward model report',
+  });
   const tuples = episodes.map((episode) => episodeToRlTuple(episode, options));
   const rewards = tuples.map((tuple) => tuple.reward.total);
   const averageReward = rewards.length ? round(rewards.reduce((sum, value) => sum + value, 0) / rewards.length) : 0;

package/scripts/bayes-optimal-gate.js

@@ -32,6 +32,7 @@
 
 const fs = require('node:fs');
 const path = require('node:path');
+const { requireLearnedModelsEntitlement } = require('./entitlement');
 
 // Baseline loss matrix. `default` applies when no tag-specific override
 // matches. Higher = more expensive. The asymmetry below reflects the
@@ -181,7 +182,11 @@ function computeBayesPosterior({ tags, riskByTag, baseRate, modelProbability } =
  *
  * This reduces to the usual Bayes classifier when both costs are equal.
  */
-function bayesOptimalDecision(posterior, tags, lossMatrix = DEFAULT_LOSS_MATRIX) {
+function bayesOptimalDecision(posterior, tags, lossMatrix = DEFAULT_LOSS_MATRIX, options = {}) {
+  requireLearnedModelsEntitlement({
+    ...(options.entitlement || {}),
+    label: 'Bayes-optimal gate decisioning',
+  });
   const pHarmful = clip(Number(posterior?.pHarmful), 0, 1);
   const pSafe = clip(Number(posterior?.pSafe ?? 1 - pHarmful), 0, 1);
   const cFalseAllow = resolveCost(lossMatrix?.falseAllow || {}, tags);