@tangle-network/agent-bench 0.4.0 → 0.4.3
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +15 -0
- package/dist/adapters.d.ts +4 -11
- package/dist/adapters.js +78 -41
- package/dist/adapters.js.map +1 -1
- package/dist/benchmarks/_harness.d.ts +51 -66
- package/dist/benchmarks/_harness.js +329 -31
- package/dist/benchmarks/_harness.js.map +1 -1
- package/dist/benchmarks/aec-bench.d.ts +4 -25
- package/dist/benchmarks/aec-bench.js +242 -7
- package/dist/benchmarks/aec-bench.js.map +1 -1
- package/dist/benchmarks/agentbench.d.ts +5 -13
- package/dist/benchmarks/agentbench.js +114 -9
- package/dist/benchmarks/agentbench.js.map +1 -1
- package/dist/benchmarks/appworld.d.ts +9 -29
- package/dist/benchmarks/appworld.js +317 -12
- package/dist/benchmarks/appworld.js.map +1 -1
- package/dist/benchmarks/bfcl.d.ts +5 -15
- package/dist/benchmarks/bfcl.js +264 -9
- package/dist/benchmarks/bfcl.js.map +1 -1
- package/dist/benchmarks/cad-design.d.ts +16 -41
- package/dist/benchmarks/cad-design.js +512 -6
- package/dist/benchmarks/cad-design.js.map +1 -1
- package/dist/benchmarks/cadbench.d.ts +4 -17
- package/dist/benchmarks/cadbench.js +2 -8
- package/dist/benchmarks/cadgenbench.d.ts +4 -20
- package/dist/benchmarks/cadgenbench.js +2 -8
- package/dist/benchmarks/commit0.d.ts +5 -27
- package/dist/benchmarks/commit0.js +187 -9
- package/dist/benchmarks/commit0.js.map +1 -1
- package/dist/benchmarks/crag.d.ts +4 -12
- package/dist/benchmarks/crag.js +110 -8
- package/dist/benchmarks/crag.js.map +1 -1
- package/dist/benchmarks/dabstep.d.ts +5 -15
- package/dist/benchmarks/dabstep.js +177 -9
- package/dist/benchmarks/dabstep.js.map +1 -1
- package/dist/benchmarks/enterpriseops-gym.d.ts +5 -34
- package/dist/benchmarks/enterpriseops-gym.js +236 -9
- package/dist/benchmarks/enterpriseops-gym.js.map +1 -1
- package/dist/benchmarks/finresearchbench.d.ts +4 -13
- package/dist/benchmarks/finresearchbench.js +218 -7
- package/dist/benchmarks/finresearchbench.js.map +1 -1
- package/dist/benchmarks/finsearchcomp.d.ts +8 -39
- package/dist/benchmarks/finsearchcomp.js +267 -6
- package/dist/benchmarks/finsearchcomp.js.map +1 -1
- package/dist/benchmarks/frames.d.ts +15 -37
- package/dist/benchmarks/frames.js +408 -11
- package/dist/benchmarks/frames.js.map +1 -1
- package/dist/benchmarks/hotpotqa.d.ts +4 -24
- package/dist/benchmarks/hotpotqa.js +250 -14
- package/dist/benchmarks/hotpotqa.js.map +1 -1
- package/dist/benchmarks/humaneval.d.ts +19 -37
- package/dist/benchmarks/humaneval.js +279 -16
- package/dist/benchmarks/humaneval.js.map +1 -1
- package/dist/benchmarks/mind2web.d.ts +7 -34
- package/dist/benchmarks/mind2web.js +257 -8
- package/dist/benchmarks/mind2web.js.map +1 -1
- package/dist/benchmarks/nomiracl.d.ts +4 -13
- package/dist/benchmarks/nomiracl.js +146 -8
- package/dist/benchmarks/nomiracl.js.map +1 -1
- package/dist/benchmarks/open-rag-bench.d.ts +4 -12
- package/dist/benchmarks/open-rag-bench.js +110 -8
- package/dist/benchmarks/open-rag-bench.js.map +1 -1
- package/dist/benchmarks/programbench.d.ts +5 -29
- package/dist/benchmarks/programbench.js +161 -9
- package/dist/benchmarks/programbench.js.map +1 -1
- package/dist/benchmarks/rag-shared.d.ts +21 -20
- package/dist/benchmarks/rag-shared.js +245 -38
- package/dist/benchmarks/rag-shared.js.map +1 -1
- package/dist/benchmarks/ragbench.d.ts +4 -14
- package/dist/benchmarks/ragbench.js +127 -8
- package/dist/benchmarks/ragbench.js.map +1 -1
- package/dist/benchmarks/simpleqa.d.ts +17 -44
- package/dist/benchmarks/simpleqa.js +293 -10
- package/dist/benchmarks/simpleqa.js.map +1 -1
- package/dist/benchmarks/swe-bench.d.ts +23 -36
- package/dist/benchmarks/swe-bench.js +234 -13
- package/dist/benchmarks/swe-bench.js.map +1 -1
- package/dist/benchmarks/t2-ragbench.d.ts +4 -12
- package/dist/benchmarks/t2-ragbench.js +118 -8
- package/dist/benchmarks/t2-ragbench.js.map +1 -1
- package/dist/benchmarks/tau-bench-shared.d.ts +14 -22
- package/dist/benchmarks/tau-bench-shared.js +169 -9
- package/dist/benchmarks/tau-bench-shared.js.map +1 -1
- package/dist/benchmarks/tau2-bench.d.ts +5 -5
- package/dist/benchmarks/tau2-bench.js +28 -10
- package/dist/benchmarks/tau2-bench.js.map +1 -1
- package/dist/benchmarks/tau3-banking.d.ts +4 -13
- package/dist/benchmarks/tau3-banking.js +28 -8
- package/dist/benchmarks/tau3-banking.js.map +1 -1
- package/dist/benchmarks/terminal-bench.d.ts +4 -22
- package/dist/benchmarks/terminal-bench.js +140 -7
- package/dist/benchmarks/terminal-bench.js.map +1 -1
- package/dist/benchmarks/toollm.d.ts +5 -13
- package/dist/benchmarks/toollm.js +184 -9
- package/dist/benchmarks/toollm.js.map +1 -1
- package/dist/benchmarks/trata-hedge.d.ts +4 -30
- package/dist/benchmarks/trata-hedge.js +336 -6
- package/dist/benchmarks/trata-hedge.js.map +1 -1
- package/dist/benchmarks/types.d.ts +85 -94
- package/dist/benchmarks/types.js +1 -1
- package/dist/benchmarks/webarena-verified.d.ts +5 -13
- package/dist/benchmarks/webarena-verified.js +152 -9
- package/dist/benchmarks/webarena-verified.js.map +1 -1
- package/dist/cadbench-DpQWZHp4.js +285 -0
- package/dist/cadbench-DpQWZHp4.js.map +1 -0
- package/dist/cadgenbench-DRhczfsG.js +151 -0
- package/dist/cadgenbench-DRhczfsG.js.map +1 -0
- package/dist/index.d.ts +245 -293
- package/dist/index.js +1669 -1791
- package/dist/index.js.map +1 -1
- package/package.json +17 -14
- package/pier_agents/candidate_contract.py +238 -24
- package/pier_agents/tangle_candidate.py +75 -5
- package/scripts/verify-packed-consumer.mjs +84 -17
- package/scripts/verify-pier-agent.mts +6 -4
- package/src/benchmarks/_harness.test.mts +16 -1
- package/src/benchmarks/_harness.ts +9 -2
- package/src/benchmarks/terminal-bench.ts +2 -1
- package/src/corpus.test.mts +13 -0
- package/src/corpus.ts +4 -0
- package/src/profile-coordinates.ts +2 -2
- package/src/rollout-ledger/backfill-swe-arena.mts +5 -1
- package/src/rollout-ledger/settle-capture.mts +7 -1
- package/src/search-bench/profiles.ts +1 -1
- package/src/skill-sandbox-smoke.mts +2 -1
- package/src/swe-arena/gepa-seat.mts +1 -1
- package/src/swe-arena/ledger-orphans.test.mts +36 -34
- package/dist/benchmarks/cadbench.js.map +0 -1
- package/dist/benchmarks/cadgenbench.js.map +0 -1
- package/dist/benchmarks/types.js.map +0 -1
- package/dist/chunk-3U5TXJZS.js +0 -251
- package/dist/chunk-3U5TXJZS.js.map +0 -1
- package/dist/chunk-53UPUNBZ.js +0 -325
- package/dist/chunk-53UPUNBZ.js.map +0 -1
- package/dist/chunk-5H5XV76F.js +0 -240
- package/dist/chunk-5H5XV76F.js.map +0 -1
- package/dist/chunk-7GRVHU22.js +0 -208
- package/dist/chunk-7GRVHU22.js.map +0 -1
- package/dist/chunk-C7T7WEK2.js +0 -103
- package/dist/chunk-C7T7WEK2.js.map +0 -1
- package/dist/chunk-HWST3SED.js +0 -162
- package/dist/chunk-HWST3SED.js.map +0 -1
- package/dist/chunk-IA2FBTWC.js +0 -318
- package/dist/chunk-IA2FBTWC.js.map +0 -1
- package/dist/chunk-IFVINJ4B.js +0 -142
- package/dist/chunk-IFVINJ4B.js.map +0 -1
- package/dist/chunk-INNOYXCP.js +0 -387
- package/dist/chunk-INNOYXCP.js.map +0 -1
- package/dist/chunk-IZ5M6OAC.js +0 -169
- package/dist/chunk-IZ5M6OAC.js.map +0 -1
- package/dist/chunk-JTHWEDEW.js +0 -32
- package/dist/chunk-JTHWEDEW.js.map +0 -1
- package/dist/chunk-K3BQGZCT.js +0 -221
- package/dist/chunk-K3BQGZCT.js.map +0 -1
- package/dist/chunk-KP5KD6EN.js +0 -276
- package/dist/chunk-KP5KD6EN.js.map +0 -1
- package/dist/chunk-MQMRLGOG.js +0 -136
- package/dist/chunk-MQMRLGOG.js.map +0 -1
- package/dist/chunk-NQG5XDSB.js +0 -147
- package/dist/chunk-NQG5XDSB.js.map +0 -1
- package/dist/chunk-PA2ZKHJC.js +0 -230
- package/dist/chunk-PA2ZKHJC.js.map +0 -1
- package/dist/chunk-PB64GYIG.js +0 -118
- package/dist/chunk-PB64GYIG.js.map +0 -1
- package/dist/chunk-PUIRNYI7.js +0 -189
- package/dist/chunk-PUIRNYI7.js.map +0 -1
- package/dist/chunk-RCYQEFNX.js +0 -30
- package/dist/chunk-RCYQEFNX.js.map +0 -1
- package/dist/chunk-RH5F53JT.js +0 -182
- package/dist/chunk-RH5F53JT.js.map +0 -1
- package/dist/chunk-SEVJPLZC.js +0 -260
- package/dist/chunk-SEVJPLZC.js.map +0 -1
- package/dist/chunk-SFLA7OH3.js +0 -27
- package/dist/chunk-SFLA7OH3.js.map +0 -1
- package/dist/chunk-SHM6MRRF.js +0 -130
- package/dist/chunk-SHM6MRRF.js.map +0 -1
- package/dist/chunk-SHYIRB7I.js +0 -120
- package/dist/chunk-SHYIRB7I.js.map +0 -1
- package/dist/chunk-SVR2LKYI.js +0 -116
- package/dist/chunk-SVR2LKYI.js.map +0 -1
- package/dist/chunk-TBKU5XQI.js +0 -228
- package/dist/chunk-TBKU5XQI.js.map +0 -1
- package/dist/chunk-UPAMRDX4.js +0 -233
- package/dist/chunk-UPAMRDX4.js.map +0 -1
- package/dist/chunk-V7AEBY6U.js +0 -144
- package/dist/chunk-V7AEBY6U.js.map +0 -1
- package/dist/chunk-VQRS7VUC.js +0 -342
- package/dist/chunk-VQRS7VUC.js.map +0 -1
- package/dist/chunk-WSKWVEQB.js +0 -317
- package/dist/chunk-WSKWVEQB.js.map +0 -1
- package/dist/chunk-X3BTXCJ4.js +0 -262
- package/dist/chunk-X3BTXCJ4.js.map +0 -1
- package/dist/chunk-XKEFIFIC.js +0 -197
- package/dist/chunk-XKEFIFIC.js.map +0 -1
- package/dist/chunk-XYA4XSNU.js +0 -148
- package/dist/chunk-XYA4XSNU.js.map +0 -1
- package/dist/chunk-YSMEKBTD.js +0 -211
- package/dist/chunk-YSMEKBTD.js.map +0 -1
- package/dist/chunk-Z4TZ76N7.js +0 -170
- package/dist/chunk-Z4TZ76N7.js.map +0 -1
package/dist/chunk-VQRS7VUC.js
DELETED
|
@@ -1,342 +0,0 @@
|
|
|
1
|
-
// src/benchmarks/trata-hedge.ts
|
|
2
|
-
import { readdirSync, readFileSync, statSync } from "fs";
|
|
3
|
-
import { join } from "path";
|
|
4
|
-
var DEFAULT_BENCH_ROOT = "/tmp/trata-hedge-bench";
|
|
5
|
-
var MAX_FILE_BYTES = 8e3;
|
|
6
|
-
var MAX_DATA_CHARS = 85e3;
|
|
7
|
-
var WORKER_SUFFIX = [
|
|
8
|
-
"",
|
|
9
|
-
"The data files above are your only source. Cite each claim with the filename that supports it.",
|
|
10
|
-
'Write your complete analysis to the text block that starts with "ANALYSIS:" on a line by itself.',
|
|
11
|
-
"Take a clear position on the topic. Every factual claim must name the file it comes from."
|
|
12
|
-
].join("\n");
|
|
13
|
-
var JUDGE_MODEL = "gemini-2.5-pro";
|
|
14
|
-
function readText(p, maxBytes = MAX_FILE_BYTES) {
|
|
15
|
-
try {
|
|
16
|
-
const buf = readFileSync(p);
|
|
17
|
-
if (buf.length <= maxBytes) return buf.toString("utf8");
|
|
18
|
-
return buf.subarray(0, maxBytes).toString("utf8") + `
|
|
19
|
-
[\u2026 truncated at ${maxBytes} bytes \u2026]`;
|
|
20
|
-
} catch {
|
|
21
|
-
return `[unreadable: ${p}]`;
|
|
22
|
-
}
|
|
23
|
-
}
|
|
24
|
-
function listFilesRec(dir) {
|
|
25
|
-
const out = [];
|
|
26
|
-
try {
|
|
27
|
-
for (const name of readdirSync(dir)) {
|
|
28
|
-
const full = join(dir, name);
|
|
29
|
-
try {
|
|
30
|
-
if (statSync(full).isDirectory()) out.push(...listFilesRec(full));
|
|
31
|
-
else out.push(full);
|
|
32
|
-
} catch {
|
|
33
|
-
}
|
|
34
|
-
}
|
|
35
|
-
} catch {
|
|
36
|
-
}
|
|
37
|
-
return out.sort();
|
|
38
|
-
}
|
|
39
|
-
function buildDataBlock(dataDir) {
|
|
40
|
-
const files = listFilesRec(dataDir);
|
|
41
|
-
if (files.length === 0) return "(no data files found)";
|
|
42
|
-
const priority = (f) => {
|
|
43
|
-
if (f.includes("company_profiles")) return 0;
|
|
44
|
-
if (f.includes("earnings_call")) return 1;
|
|
45
|
-
if (/financials\/(income|cash_flow|balance)/.test(f) && !/bam_|bx_|kkr_|ares_|cg_|owl_|fsk_|cat_|de_|tdg_|car_/.test(f)) return 2;
|
|
46
|
-
if (f.includes("sec_filings")) return 3;
|
|
47
|
-
if (f.includes("investor_pres")) return 4;
|
|
48
|
-
if (f.includes("press_release")) return 5;
|
|
49
|
-
return 6;
|
|
50
|
-
};
|
|
51
|
-
const sorted = [...files].sort((a, b) => priority(a) - priority(b) || a.localeCompare(b));
|
|
52
|
-
const blocks = [];
|
|
53
|
-
let total = 0;
|
|
54
|
-
for (const f of sorted) {
|
|
55
|
-
const rel = f.slice(dataDir.length + 1);
|
|
56
|
-
const content = readText(f);
|
|
57
|
-
const chunk = `
|
|
58
|
-
=== FILE: data/${rel} ===
|
|
59
|
-
${content}`;
|
|
60
|
-
if (total + chunk.length > MAX_DATA_CHARS) {
|
|
61
|
-
blocks.push(`
|
|
62
|
-
[\u2026 ${sorted.length - blocks.length} more files omitted to stay within context limit \u2026]`);
|
|
63
|
-
break;
|
|
64
|
-
}
|
|
65
|
-
blocks.push(chunk);
|
|
66
|
-
total += chunk.length;
|
|
67
|
-
}
|
|
68
|
-
return blocks.join("");
|
|
69
|
-
}
|
|
70
|
-
function buildCitedDataBlock(answerText, dataDir) {
|
|
71
|
-
const files = listFilesRec(dataDir);
|
|
72
|
-
const cited = [];
|
|
73
|
-
for (const f of files) {
|
|
74
|
-
const rel = f.slice(dataDir.length + 1);
|
|
75
|
-
const name = rel.split("/").pop() ?? "";
|
|
76
|
-
if (answerText.includes(rel) || answerText.includes(name)) cited.push(f);
|
|
77
|
-
}
|
|
78
|
-
if (cited.length === 0) {
|
|
79
|
-
return "(The agent did not cite any data files, or no cited file could be located.)";
|
|
80
|
-
}
|
|
81
|
-
const blocks = [];
|
|
82
|
-
for (const f of cited) {
|
|
83
|
-
const rel = f.slice(dataDir.length + 1);
|
|
84
|
-
blocks.push(`
|
|
85
|
-
=== FILE: data/${rel} ===
|
|
86
|
-
${readText(f)}`);
|
|
87
|
-
}
|
|
88
|
-
return blocks.join("");
|
|
89
|
-
}
|
|
90
|
-
function parseThemeMoveCounts(groundTruth) {
|
|
91
|
-
const lines = groundTruth.split("\n");
|
|
92
|
-
const themeStarts = lines.reduce((acc, l, i) => {
|
|
93
|
-
if (/^\d+\./.test(l.trim())) acc.push(i);
|
|
94
|
-
return acc;
|
|
95
|
-
}, []);
|
|
96
|
-
return themeStarts.map((start, k) => {
|
|
97
|
-
const end = k + 1 < themeStarts.length ? themeStarts[k + 1] : lines.length;
|
|
98
|
-
const block = lines.slice(start, end).join("\n");
|
|
99
|
-
const label = lines[start].replace(/^\d+\.\s*/, "").trim();
|
|
100
|
-
const nMoves = (block.match(/^\s*\[[a-z]\]\s+/gm) ?? []).length;
|
|
101
|
-
return { label, nMoves };
|
|
102
|
-
});
|
|
103
|
-
}
|
|
104
|
-
function coverageThreshold(nMoves) {
|
|
105
|
-
return Math.max(1, Math.min(nMoves - 1, 3));
|
|
106
|
-
}
|
|
107
|
-
function judgeRouter() {
|
|
108
|
-
const key = process.env.TANGLE_API_KEY;
|
|
109
|
-
if (!key) throw new Error("TANGLE_API_KEY is required for the Trata hedge-bench judge");
|
|
110
|
-
const model = process.env.JUDGE_MODEL ?? JUDGE_MODEL;
|
|
111
|
-
const baseUrl = process.env.ROUTER_BASE ?? "https://router.tangle.tools/v1";
|
|
112
|
-
return { baseUrl, key, model };
|
|
113
|
-
}
|
|
114
|
-
function parseJsonFallback(raw) {
|
|
115
|
-
let s = raw.trim();
|
|
116
|
-
if (s.startsWith("```")) s = s.split("\n").slice(1).join("\n");
|
|
117
|
-
if (s.endsWith("```")) s = s.slice(0, s.lastIndexOf("```"));
|
|
118
|
-
s = s.replace(/,(\s*[}\]])/g, "$1").trim();
|
|
119
|
-
try {
|
|
120
|
-
return JSON.parse(s);
|
|
121
|
-
} catch {
|
|
122
|
-
try {
|
|
123
|
-
const m = s.match(/\{[\s\S]*\}/);
|
|
124
|
-
if (m) return JSON.parse(m[0]);
|
|
125
|
-
} catch {
|
|
126
|
-
}
|
|
127
|
-
}
|
|
128
|
-
return null;
|
|
129
|
-
}
|
|
130
|
-
async function callJudge(router, prompt, maxAttempts = 2) {
|
|
131
|
-
for (let i = 0; i < maxAttempts; i++) {
|
|
132
|
-
const res = await fetch(`${router.baseUrl}/chat/completions`, {
|
|
133
|
-
method: "POST",
|
|
134
|
-
headers: { "content-type": "application/json", authorization: `Bearer ${router.key}` },
|
|
135
|
-
body: JSON.stringify({
|
|
136
|
-
model: router.model,
|
|
137
|
-
temperature: 0,
|
|
138
|
-
max_tokens: 16384,
|
|
139
|
-
messages: [{ role: "user", content: prompt }]
|
|
140
|
-
})
|
|
141
|
-
});
|
|
142
|
-
if (!res.ok) {
|
|
143
|
-
if (i < maxAttempts - 1) continue;
|
|
144
|
-
throw new Error(`Trata judge HTTP ${res.status}: ${(await res.text()).slice(0, 300)}`);
|
|
145
|
-
}
|
|
146
|
-
const body = await res.json();
|
|
147
|
-
const content = body.choices?.[0]?.message?.content;
|
|
148
|
-
if (typeof content !== "string") continue;
|
|
149
|
-
const parsed = parseJsonFallback(content);
|
|
150
|
-
if (parsed !== null) return parsed;
|
|
151
|
-
}
|
|
152
|
-
return null;
|
|
153
|
-
}
|
|
154
|
-
async function task1HallucinationCheck(meta, answer, router) {
|
|
155
|
-
const citedBlock = buildCitedDataBlock(answer, meta.dataDir);
|
|
156
|
-
const dataContext = `The block below contains the contents of the data files the agent cited in their answer. Treat this as the ground truth for factual verification.
|
|
157
|
-
|
|
158
|
-
<cited_data>
|
|
159
|
-
${citedBlock}
|
|
160
|
-
</cited_data>`;
|
|
161
|
-
const prompt = meta.gradingTask1.replace("{data_context}", dataContext).replace("{agent_answer}", answer);
|
|
162
|
-
const result = await callJudge(router, prompt);
|
|
163
|
-
return result ?? { hallucinations_detected: false, unverifiable_claims: [] };
|
|
164
|
-
}
|
|
165
|
-
async function task2PerMoveCheck(meta, answer, flaggedClaims, router) {
|
|
166
|
-
const claimsBlock = flaggedClaims && flaggedClaims.length > 0 ? JSON.stringify(flaggedClaims, null, 2) : "(No claims were flagged as hallucinations.)";
|
|
167
|
-
const prompt = meta.gradingTask2.replace("{ground_truth}", meta.groundTruth).replace("{agent_answer}", answer).replace("{flagged_claims_block}", claimsBlock).replace("{num_themes}", String(meta.themeMoveCounts.length));
|
|
168
|
-
const result = await callJudge(router, prompt);
|
|
169
|
-
return result ?? { themes: [] };
|
|
170
|
-
}
|
|
171
|
-
async function task3SynthesisCheck(meta, answer, router) {
|
|
172
|
-
const prompt = meta.gradingTask3.replace("{ground_truth}", meta.groundTruth).replace("{agent_answer}", answer);
|
|
173
|
-
const result = await callJudge(router, prompt);
|
|
174
|
-
return result ?? { synthesis_found: null };
|
|
175
|
-
}
|
|
176
|
-
async function runTriataJudge(meta, answer, router) {
|
|
177
|
-
if (!answer.trim()) {
|
|
178
|
-
return { resolved: false, score: 0, detail: JSON.stringify({ error: "empty answer" }) };
|
|
179
|
-
}
|
|
180
|
-
const t1 = await task1HallucinationCheck(meta, answer, router);
|
|
181
|
-
const flaggedClaims = t1.unverifiable_claims ?? [];
|
|
182
|
-
const [t2, t3] = await Promise.all([
|
|
183
|
-
task2PerMoveCheck(meta, answer, flaggedClaims, router),
|
|
184
|
-
task3SynthesisCheck(meta, answer, router)
|
|
185
|
-
]);
|
|
186
|
-
const judgeThemes = t2.themes ?? [];
|
|
187
|
-
const themesHit = [];
|
|
188
|
-
const themesMissed = [];
|
|
189
|
-
for (let i = 0; i < meta.themeMoveCounts.length; i++) {
|
|
190
|
-
const { label, nMoves } = meta.themeMoveCounts[i];
|
|
191
|
-
const judge = judgeThemes[i] ?? {};
|
|
192
|
-
const movesHit = (judge.moves_hit ?? []).map((m) => m.replace(/[\[\]]/g, ""));
|
|
193
|
-
const tainted = new Set((judge.moves_tainted ?? []).map((m) => m.replace(/[\[\]]/g, "")));
|
|
194
|
-
const validHit = movesHit.filter((m) => !tainted.has(m));
|
|
195
|
-
const threshold = coverageThreshold(nMoves);
|
|
196
|
-
if (validHit.length >= threshold) themesHit.push(label);
|
|
197
|
-
else themesMissed.push(label);
|
|
198
|
-
}
|
|
199
|
-
const numThemes = meta.themeMoveCounts.length;
|
|
200
|
-
const synth = t3.synthesis_found;
|
|
201
|
-
const hasSynthesis = typeof synth === "string" && synth.length > 0 && synth !== "null";
|
|
202
|
-
let rawScore;
|
|
203
|
-
if (numThemes === 0) rawScore = 0;
|
|
204
|
-
else if (themesHit.length === numThemes && hasSynthesis) rawScore = 4;
|
|
205
|
-
else if (themesHit.length === numThemes) rawScore = 3;
|
|
206
|
-
else if (themesHit.length >= 2) rawScore = 2;
|
|
207
|
-
else if (themesHit.length >= 1) rawScore = 1;
|
|
208
|
-
else rawScore = 0;
|
|
209
|
-
const hallRaw = t1.hallucinations_detected;
|
|
210
|
-
const hallucinations = hallRaw === true || hallRaw === "true";
|
|
211
|
-
return {
|
|
212
|
-
resolved: rawScore === 4,
|
|
213
|
-
score: numThemes > 0 ? rawScore / 4 : 0,
|
|
214
|
-
detail: JSON.stringify({
|
|
215
|
-
rawScore,
|
|
216
|
-
themesHit,
|
|
217
|
-
themesMissed,
|
|
218
|
-
hallucinations,
|
|
219
|
-
unverifiableClaims: flaggedClaims.length,
|
|
220
|
-
synthesis: synth ?? null,
|
|
221
|
-
judgeModel: router.model
|
|
222
|
-
})
|
|
223
|
-
};
|
|
224
|
-
}
|
|
225
|
-
function benchRoot() {
|
|
226
|
-
return process.env.TRATA_BENCH_ROOT ?? DEFAULT_BENCH_ROOT;
|
|
227
|
-
}
|
|
228
|
-
function loadTask(taskId) {
|
|
229
|
-
const taskDir = join(benchRoot(), "environments", taskId);
|
|
230
|
-
const dataDir = join(taskDir, "environment", "data");
|
|
231
|
-
const testsDir = join(taskDir, "tests");
|
|
232
|
-
const instruction = readText(join(taskDir, "instruction.md"), Infinity);
|
|
233
|
-
const dataBlock = buildDataBlock(dataDir);
|
|
234
|
-
const groundTruth = readText(join(testsDir, "ground_truth.txt"), Infinity);
|
|
235
|
-
const gradingTask1 = readText(join(testsDir, "grading_prompt_task1.md"), Infinity);
|
|
236
|
-
const gradingTask2 = readText(join(testsDir, "grading_prompt_task2.md"), Infinity);
|
|
237
|
-
const gradingTask3 = readText(join(testsDir, "grading_prompt_task3.md"), Infinity);
|
|
238
|
-
const themeMoveCounts = parseThemeMoveCounts(groundTruth);
|
|
239
|
-
const prompt = [
|
|
240
|
-
instruction.replace(
|
|
241
|
-
"You are a financial analyst with access to the data in `/app/data/`.",
|
|
242
|
-
"You are a financial analyst. The relevant financial data files are provided inline below."
|
|
243
|
-
).replace(/Write your full analysis to `\/app\/answer\.txt`\.?/, "Write your full analysis below.").replace(/`\/app\/data\/`/g, "the data files below"),
|
|
244
|
-
"",
|
|
245
|
-
"## Data files (inline)",
|
|
246
|
-
dataBlock,
|
|
247
|
-
WORKER_SUFFIX
|
|
248
|
-
].join("\n");
|
|
249
|
-
const meta = {
|
|
250
|
-
taskDir,
|
|
251
|
-
taskId,
|
|
252
|
-
dataDir,
|
|
253
|
-
groundTruth,
|
|
254
|
-
gradingTask1,
|
|
255
|
-
gradingTask2,
|
|
256
|
-
gradingTask3,
|
|
257
|
-
themeMoveCounts
|
|
258
|
-
};
|
|
259
|
-
const task = {
|
|
260
|
-
id: taskId,
|
|
261
|
-
prompt,
|
|
262
|
-
metadata: meta
|
|
263
|
-
};
|
|
264
|
-
return { task, meta };
|
|
265
|
-
}
|
|
266
|
-
function readMeta(task) {
|
|
267
|
-
const md = task.metadata;
|
|
268
|
-
if (!md || typeof md.taskDir !== "string" || typeof md.groundTruth !== "string") {
|
|
269
|
-
throw new Error(`Trata task ${task.id} missing judge metadata \u2014 was it loaded by this adapter?`);
|
|
270
|
-
}
|
|
271
|
-
return md;
|
|
272
|
-
}
|
|
273
|
-
function listTaskIds(root) {
|
|
274
|
-
try {
|
|
275
|
-
return readdirSync(join(root, "environments")).filter((name) => {
|
|
276
|
-
try {
|
|
277
|
-
return statSync(join(root, "environments", name)).isDirectory();
|
|
278
|
-
} catch {
|
|
279
|
-
return false;
|
|
280
|
-
}
|
|
281
|
-
}).sort();
|
|
282
|
-
} catch (err) {
|
|
283
|
-
throw new Error(
|
|
284
|
-
`Trata benchmark environments not found at ${join(root, "environments")}: ${err instanceof Error ? err.message : err}.
|
|
285
|
-
Clone the repo: git clone --depth 1 https://github.com/Trata-Inc/trata-hedge-bench ${root}`
|
|
286
|
-
);
|
|
287
|
-
}
|
|
288
|
-
}
|
|
289
|
-
function extractAnalysis(artifact) {
|
|
290
|
-
if (artifact.startsWith("ANALYSIS:")) return artifact.slice("ANALYSIS:".length).trim();
|
|
291
|
-
const marker = artifact.indexOf("\nANALYSIS:");
|
|
292
|
-
if (marker !== -1) return artifact.slice(marker + "\nANALYSIS:".length).trim();
|
|
293
|
-
return "";
|
|
294
|
-
}
|
|
295
|
-
function createTrataHedgeAdapter() {
|
|
296
|
-
return {
|
|
297
|
-
name: "trata-hedge",
|
|
298
|
-
async preflight() {
|
|
299
|
-
const root = benchRoot();
|
|
300
|
-
const envDir = join(root, "environments");
|
|
301
|
-
try {
|
|
302
|
-
const stat = statSync(envDir);
|
|
303
|
-
if (!stat.isDirectory()) throw new Error("not a directory");
|
|
304
|
-
} catch {
|
|
305
|
-
throw new Error(
|
|
306
|
-
`Trata hedge-bench not found at ${envDir}.
|
|
307
|
-
Clone: git clone --depth 1 https://github.com/Trata-Inc/trata-hedge-bench ${root}`
|
|
308
|
-
);
|
|
309
|
-
}
|
|
310
|
-
judgeRouter();
|
|
311
|
-
const ids = listTaskIds(root);
|
|
312
|
-
if (ids.length === 0) throw new Error(`No task directories found under ${envDir}`);
|
|
313
|
-
console.warn(
|
|
314
|
-
`[trata-hedge] ${ids.length} tasks loaded from ${root}. Worker model reads embedded data inline (Harbor/Docker not required). Judge uses ${process.env.JUDGE_MODEL ?? JUDGE_MODEL} via Tangle router.`
|
|
315
|
-
);
|
|
316
|
-
},
|
|
317
|
-
async loadTasks(opts = {}) {
|
|
318
|
-
const root = benchRoot();
|
|
319
|
-
let ids = listTaskIds(root);
|
|
320
|
-
if (opts.ids) {
|
|
321
|
-
const want = new Set(opts.ids);
|
|
322
|
-
ids = ids.filter((id) => want.has(id));
|
|
323
|
-
} else if (opts.limit !== void 0) {
|
|
324
|
-
ids = ids.slice(0, opts.limit);
|
|
325
|
-
}
|
|
326
|
-
return ids.map((id) => loadTask(id).task);
|
|
327
|
-
},
|
|
328
|
-
async goldArtifact(_task) {
|
|
329
|
-
return void 0;
|
|
330
|
-
},
|
|
331
|
-
async judge(task, artifact) {
|
|
332
|
-
const meta = readMeta(task);
|
|
333
|
-
const answer = extractAnalysis(artifact);
|
|
334
|
-
return runTriataJudge(meta, answer, judgeRouter());
|
|
335
|
-
}
|
|
336
|
-
};
|
|
337
|
-
}
|
|
338
|
-
|
|
339
|
-
export {
|
|
340
|
-
createTrataHedgeAdapter
|
|
341
|
-
};
|
|
342
|
-
//# sourceMappingURL=chunk-VQRS7VUC.js.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"sources":["../src/benchmarks/trata-hedge.ts"],"sourcesContent":["/**\n * Trata Hedge-Bench adapter (Trata-Inc/trata-hedge-bench). 102 financial\n * analysis tasks across 6 domains (private equity, managed care, industrials,\n * vertical SaaS, REITs, insurance). Each task bundles real earnings-call\n * transcripts, financial statements, press releases, and SEC filings under\n * `environment/data/`; the agent must produce a grounded analysis citing those\n * files.\n *\n * Harbor architecture: the original benchmark runs each task in a Docker\n * container with file-read tools. This adapter skips Harbor and embeds the data\n * files directly in the worker prompt so any router backend can score it. Large\n * files (> 30 KB) are truncated to fit within model context limits. The judge\n * reads the actual files from disk for citation verification.\n *\n * Judge: a 3-stage LLM cascade faithful to grade.py —\n * Task 1: hallucination check (cited-file context + agent answer)\n * Task 2: per-move coverage check (ground truth + agent answer + flagged claims)\n * Task 3: synthesis check (ground truth + agent answer)\n * Score 0–4: themes_covered == num_themes && synthesis → 4, all themes → 3,\n * ≥2 themes → 2, ≥1 theme → 1, else → 0. Normalized to 0..1 for BenchScore;\n * resolved = score 4 (sparse reward).\n *\n * Requires TANGLE_API_KEY and TRATA_BENCH_ROOT=/tmp/trata-hedge-bench.\n * Clone: git clone --depth 1 https://github.com/Trata-Inc/trata-hedge-bench /tmp/trata-hedge-bench\n */\n\nimport { readdirSync, readFileSync, statSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'\n\nconst DEFAULT_BENCH_ROOT = '/tmp/trata-hedge-bench'\nconst MAX_FILE_BYTES = 8_000\n// Sandbox API rejects messages > 100K chars; instruction + suffix adds ~500 chars overhead.\nconst MAX_DATA_CHARS = 85_000\nconst WORKER_SUFFIX = [\n '',\n 'The data files above are your only source. Cite each claim with the filename that supports it.',\n 'Write your complete analysis to the text block that starts with \"ANALYSIS:\" on a line by itself.',\n 'Take a clear position on the topic. Every factual claim must name the file it comes from.',\n].join('\\n')\n\n// Judge model: Gemini 2.5 Pro via router (matches gemini-3.1-pro-preview tier used by grade.py).\nconst JUDGE_MODEL = 'gemini-2.5-pro'\n\ninterface TrataTaskMeta {\n taskDir: string\n taskId: string\n dataDir: string\n groundTruth: string\n gradingTask1: string\n gradingTask2: string\n gradingTask3: string\n /** [(label, n_moves)] in rubric order. */\n themeMoveCounts: Array<{ label: string; nMoves: number }>\n}\n\n// ── file reading helpers ──────────────────────────────────────────────────────\n\nfunction readText(p: string, maxBytes = MAX_FILE_BYTES): string {\n try {\n const buf = readFileSync(p)\n if (buf.length <= maxBytes) return buf.toString('utf8')\n return buf.subarray(0, maxBytes).toString('utf8') + `\\n[… truncated at ${maxBytes} bytes …]`\n } catch {\n return `[unreadable: ${p}]`\n }\n}\n\nfunction listFilesRec(dir: string): string[] {\n const out: string[] = []\n try {\n for (const name of readdirSync(dir)) {\n const full = join(dir, name)\n try {\n if (statSync(full).isDirectory()) out.push(...listFilesRec(full))\n else out.push(full)\n } catch {\n // skip unreadable entries\n }\n }\n } catch {\n // dir absent\n }\n return out.sort()\n}\n\nfunction buildDataBlock(dataDir: string): string {\n const files = listFilesRec(dataDir)\n if (files.length === 0) return '(no data files found)'\n // Prioritise: company profiles → earnings calls → primary-ticker financials →\n // sec filings → press releases → peers. Stop once total chars hit MAX_DATA_CHARS.\n const priority = (f: string) => {\n if (f.includes('company_profiles')) return 0\n if (f.includes('earnings_call')) return 1\n if (/financials\\/(income|cash_flow|balance)/.test(f) && !/bam_|bx_|kkr_|ares_|cg_|owl_|fsk_|cat_|de_|tdg_|car_/.test(f)) return 2\n if (f.includes('sec_filings')) return 3\n if (f.includes('investor_pres')) return 4\n if (f.includes('press_release')) return 5\n return 6\n }\n const sorted = [...files].sort((a, b) => priority(a) - priority(b) || a.localeCompare(b))\n const blocks: string[] = []\n let total = 0\n for (const f of sorted) {\n const rel = f.slice(dataDir.length + 1)\n const content = readText(f)\n const chunk = `\\n=== FILE: data/${rel} ===\\n${content}`\n if (total + chunk.length > MAX_DATA_CHARS) {\n blocks.push(`\\n[… ${sorted.length - blocks.length} more files omitted to stay within context limit …]`)\n break\n }\n blocks.push(chunk)\n total += chunk.length\n }\n return blocks.join('')\n}\n\nfunction buildCitedDataBlock(answerText: string, dataDir: string): string {\n const files = listFilesRec(dataDir)\n const cited: string[] = []\n for (const f of files) {\n const rel = f.slice(dataDir.length + 1)\n const name = rel.split('/').pop() ?? ''\n if (answerText.includes(rel) || answerText.includes(name)) cited.push(f)\n }\n if (cited.length === 0) {\n return '(The agent did not cite any data files, or no cited file could be located.)'\n }\n const blocks: string[] = []\n for (const f of cited) {\n const rel = f.slice(dataDir.length + 1)\n blocks.push(`\\n=== FILE: data/${rel} ===\\n${readText(f)}`)\n }\n return blocks.join('')\n}\n\n// ── ground-truth parser ───────────────────────────────────────────────────────\n\nfunction parseThemeMoveCounts(groundTruth: string): Array<{ label: string; nMoves: number }> {\n const lines = groundTruth.split('\\n')\n const themeStarts = lines.reduce<number[]>((acc, l, i) => {\n if (/^\\d+\\./.test(l.trim())) acc.push(i)\n return acc\n }, [])\n return themeStarts.map((start, k) => {\n const end = k + 1 < themeStarts.length ? themeStarts[k + 1] : lines.length\n const block = lines.slice(start, end).join('\\n')\n const label = lines[start].replace(/^\\d+\\.\\s*/, '').trim()\n const nMoves = (block.match(/^\\s*\\[[a-z]\\]\\s+/gm) ?? []).length\n return { label, nMoves }\n })\n}\n\nfunction coverageThreshold(nMoves: number): number {\n return Math.max(1, Math.min(nMoves - 1, 3))\n}\n\n// ── router judge calls ────────────────────────────────────────────────────────\n\ninterface JudgeRouter {\n baseUrl: string\n key: string\n model: string\n}\n\nfunction judgeRouter(): JudgeRouter {\n const key = process.env.TANGLE_API_KEY\n if (!key) throw new Error('TANGLE_API_KEY is required for the Trata hedge-bench judge')\n const model = process.env.JUDGE_MODEL ?? JUDGE_MODEL\n const baseUrl = process.env.ROUTER_BASE ?? 'https://router.tangle.tools/v1'\n return { baseUrl, key, model }\n}\n\nfunction parseJsonFallback(raw: string): unknown {\n let s = raw.trim()\n if (s.startsWith('```')) s = s.split('\\n').slice(1).join('\\n')\n if (s.endsWith('```')) s = s.slice(0, s.lastIndexOf('```'))\n s = s.replace(/,(\\s*[}\\]])/g, '$1').trim()\n try {\n return JSON.parse(s)\n } catch {\n try {\n const m = s.match(/\\{[\\s\\S]*\\}/)\n if (m) return JSON.parse(m[0])\n } catch {\n // fallthrough\n }\n }\n return null\n}\n\nasync function callJudge(router: JudgeRouter, prompt: string, maxAttempts = 2): Promise<unknown> {\n for (let i = 0; i < maxAttempts; i++) {\n const res = await fetch(`${router.baseUrl}/chat/completions`, {\n method: 'POST',\n headers: { 'content-type': 'application/json', authorization: `Bearer ${router.key}` },\n body: JSON.stringify({\n model: router.model,\n temperature: 0,\n max_tokens: 16384,\n messages: [{ role: 'user', content: prompt }],\n }),\n })\n if (!res.ok) {\n if (i < maxAttempts - 1) continue\n throw new Error(`Trata judge HTTP ${res.status}: ${(await res.text()).slice(0, 300)}`)\n }\n const body = (await res.json()) as { choices?: Array<{ message?: { content?: string } }> }\n const content = body.choices?.[0]?.message?.content\n if (typeof content !== 'string') continue\n const parsed = parseJsonFallback(content)\n if (parsed !== null) return parsed\n }\n return null\n}\n\ninterface Task1Result {\n hallucinations_detected?: boolean | string\n unverifiable_claims?: Array<{ claim: string; cited_file?: string; issue?: string; evidence_check?: string }>\n}\n\ninterface Task2ThemeResult {\n label?: string\n moves_hit?: string[]\n moves_missed?: string[]\n moves_tainted?: string[]\n move_reasoning?: Record<string, string>\n}\n\ninterface Task2Result {\n themes?: Task2ThemeResult[]\n}\n\ninterface Task3Result {\n synthesis_found?: string | null\n}\n\nasync function task1HallucinationCheck(\n meta: TrataTaskMeta,\n answer: string,\n router: JudgeRouter,\n): Promise<Task1Result> {\n const citedBlock = buildCitedDataBlock(answer, meta.dataDir)\n const dataContext =\n 'The block below contains the contents of the data files the agent cited in their answer. ' +\n 'Treat this as the ground truth for factual verification.\\n\\n' +\n `<cited_data>\\n${citedBlock}\\n</cited_data>`\n const prompt = meta.gradingTask1\n .replace('{data_context}', dataContext)\n .replace('{agent_answer}', answer)\n const result = await callJudge(router, prompt)\n return (result as Task1Result) ?? { hallucinations_detected: false, unverifiable_claims: [] }\n}\n\nasync function task2PerMoveCheck(\n meta: TrataTaskMeta,\n answer: string,\n flaggedClaims: Task1Result['unverifiable_claims'],\n router: JudgeRouter,\n): Promise<Task2Result> {\n const claimsBlock =\n flaggedClaims && flaggedClaims.length > 0\n ? JSON.stringify(flaggedClaims, null, 2)\n : '(No claims were flagged as hallucinations.)'\n const prompt = meta.gradingTask2\n .replace('{ground_truth}', meta.groundTruth)\n .replace('{agent_answer}', answer)\n .replace('{flagged_claims_block}', claimsBlock)\n .replace('{num_themes}', String(meta.themeMoveCounts.length))\n const result = await callJudge(router, prompt)\n return (result as Task2Result) ?? { themes: [] }\n}\n\nasync function task3SynthesisCheck(\n meta: TrataTaskMeta,\n answer: string,\n router: JudgeRouter,\n): Promise<Task3Result> {\n const prompt = meta.gradingTask3\n .replace('{ground_truth}', meta.groundTruth)\n .replace('{agent_answer}', answer)\n const result = await callJudge(router, prompt)\n return (result as Task3Result) ?? { synthesis_found: null }\n}\n\nasync function runTriataJudge(meta: TrataTaskMeta, answer: string, router: JudgeRouter): Promise<BenchScore> {\n if (!answer.trim()) {\n return { resolved: false, score: 0, detail: JSON.stringify({ error: 'empty answer' }) }\n }\n\n // Stage 1: hallucination check (serial — feeds into stage 2).\n const t1 = await task1HallucinationCheck(meta, answer, router)\n const flaggedClaims = t1.unverifiable_claims ?? []\n\n // Stages 2 + 3 in parallel.\n const [t2, t3] = await Promise.all([\n task2PerMoveCheck(meta, answer, flaggedClaims, router),\n task3SynthesisCheck(meta, answer, router),\n ])\n\n const judgeThemes = t2.themes ?? []\n const themesHit: string[] = []\n const themesMissed: string[] = []\n\n for (let i = 0; i < meta.themeMoveCounts.length; i++) {\n const { label, nMoves } = meta.themeMoveCounts[i]\n const judge = judgeThemes[i] ?? {}\n const movesHit = (judge.moves_hit ?? []).map((m: string) => m.replace(/[\\[\\]]/g, ''))\n const tainted = new Set((judge.moves_tainted ?? []).map((m: string) => m.replace(/[\\[\\]]/g, '')))\n const validHit = movesHit.filter((m: string) => !tainted.has(m))\n const threshold = coverageThreshold(nMoves)\n if (validHit.length >= threshold) themesHit.push(label)\n else themesMissed.push(label)\n }\n\n const numThemes = meta.themeMoveCounts.length\n const synth = t3.synthesis_found\n const hasSynthesis = typeof synth === 'string' && synth.length > 0 && synth !== 'null'\n\n let rawScore: number\n if (numThemes === 0) rawScore = 0\n else if (themesHit.length === numThemes && hasSynthesis) rawScore = 4\n else if (themesHit.length === numThemes) rawScore = 3\n else if (themesHit.length >= 2) rawScore = 2\n else if (themesHit.length >= 1) rawScore = 1\n else rawScore = 0\n\n const hallRaw = t1.hallucinations_detected\n const hallucinations = hallRaw === true || hallRaw === 'true'\n\n return {\n resolved: rawScore === 4,\n score: numThemes > 0 ? rawScore / 4 : 0,\n detail: JSON.stringify({\n rawScore,\n themesHit,\n themesMissed,\n hallucinations,\n unverifiableClaims: flaggedClaims.length,\n synthesis: synth ?? null,\n judgeModel: router.model,\n }),\n }\n}\n\n// ── task loading ──────────────────────────────────────────────────────────────\n\nfunction benchRoot(): string {\n return process.env.TRATA_BENCH_ROOT ?? DEFAULT_BENCH_ROOT\n}\n\nfunction loadTask(taskId: string): { task: BenchTask; meta: TrataTaskMeta } {\n const taskDir = join(benchRoot(), 'environments', taskId)\n const dataDir = join(taskDir, 'environment', 'data')\n const testsDir = join(taskDir, 'tests')\n\n const instruction = readText(join(taskDir, 'instruction.md'), Infinity)\n const dataBlock = buildDataBlock(dataDir)\n const groundTruth = readText(join(testsDir, 'ground_truth.txt'), Infinity)\n const gradingTask1 = readText(join(testsDir, 'grading_prompt_task1.md'), Infinity)\n const gradingTask2 = readText(join(testsDir, 'grading_prompt_task2.md'), Infinity)\n const gradingTask3 = readText(join(testsDir, 'grading_prompt_task3.md'), Infinity)\n\n const themeMoveCounts = parseThemeMoveCounts(groundTruth)\n\n // Rewrite the instruction so the agent knows data is inline (no /app/data/ FS).\n const prompt = [\n instruction\n .replace(\n \"You are a financial analyst with access to the data in `/app/data/`.\",\n 'You are a financial analyst. The relevant financial data files are provided inline below.',\n )\n .replace(/Write your full analysis to `\\/app\\/answer\\.txt`\\.?/, 'Write your full analysis below.')\n .replace(/`\\/app\\/data\\/`/g, 'the data files below'),\n '',\n '## Data files (inline)',\n dataBlock,\n WORKER_SUFFIX,\n ].join('\\n')\n\n const meta: TrataTaskMeta = {\n taskDir,\n taskId,\n dataDir,\n groundTruth,\n gradingTask1,\n gradingTask2,\n gradingTask3,\n themeMoveCounts,\n }\n\n // Store meta in the BenchTask so the judge can access it without re-reading.\n const task: BenchTask = {\n id: taskId,\n prompt,\n metadata: meta as unknown as Record<string, unknown>,\n }\n\n return { task, meta }\n}\n\nfunction readMeta(task: BenchTask): TrataTaskMeta {\n const md = task.metadata\n if (!md || typeof md.taskDir !== 'string' || typeof md.groundTruth !== 'string') {\n throw new Error(`Trata task ${task.id} missing judge metadata — was it loaded by this adapter?`)\n }\n return md as unknown as TrataTaskMeta\n}\n\n/** List all task ids from the environments directory. */\nfunction listTaskIds(root: string): string[] {\n try {\n return readdirSync(join(root, 'environments'))\n .filter((name) => {\n try {\n return statSync(join(root, 'environments', name)).isDirectory()\n } catch {\n return false\n }\n })\n .sort()\n } catch (err) {\n throw new Error(\n `Trata benchmark environments not found at ${join(root, 'environments')}: ` +\n `${err instanceof Error ? err.message : err}.\\n` +\n `Clone the repo: git clone --depth 1 https://github.com/Trata-Inc/trata-hedge-bench ${root}`,\n )\n }\n}\n\n// ── extract agent analysis from artifact ─────────────────────────────────────\n\nfunction extractAnalysis(artifact: string): string {\n // Model may write \"ANALYSIS:\" as its very first token (no leading newline) or\n // after some preamble. Both are valid; fail-closed on absence.\n if (artifact.startsWith('ANALYSIS:')) return artifact.slice('ANALYSIS:'.length).trim()\n const marker = artifact.indexOf('\\nANALYSIS:')\n if (marker !== -1) return artifact.slice(marker + '\\nANALYSIS:'.length).trim()\n return ''\n}\n\n// ── adapter factory ───────────────────────────────────────────────────────────\n\nexport function createTrataHedgeAdapter(): BenchmarkAdapter {\n return {\n name: 'trata-hedge',\n\n async preflight() {\n const root = benchRoot()\n const envDir = join(root, 'environments')\n try {\n const stat = statSync(envDir)\n if (!stat.isDirectory()) throw new Error('not a directory')\n } catch {\n throw new Error(\n `Trata hedge-bench not found at ${envDir}.\\n` +\n `Clone: git clone --depth 1 https://github.com/Trata-Inc/trata-hedge-bench ${root}`,\n )\n }\n // Validate judge router connectivity.\n judgeRouter()\n const ids = listTaskIds(root)\n if (ids.length === 0) throw new Error(`No task directories found under ${envDir}`)\n console.warn(\n `[trata-hedge] ${ids.length} tasks loaded from ${root}. ` +\n `Worker model reads embedded data inline (Harbor/Docker not required). ` +\n `Judge uses ${process.env.JUDGE_MODEL ?? JUDGE_MODEL} via Tangle router.`,\n )\n },\n\n async loadTasks(opts: LoadOptions = {}): Promise<BenchTask[]> {\n const root = benchRoot()\n let ids = listTaskIds(root)\n if (opts.ids) {\n const want = new Set(opts.ids)\n ids = ids.filter((id) => want.has(id))\n } else if (opts.limit !== undefined) {\n ids = ids.slice(0, opts.limit)\n }\n return ids.map((id) => loadTask(id).task)\n },\n\n async goldArtifact(_task: BenchTask): Promise<string | undefined> {\n // No single gold answer: the score comes from rubric theme coverage, not\n // string-match against a reference response. Return undefined — the bench\n // harness will skip judge self-verification for this adapter.\n return undefined\n },\n\n async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n const meta = readMeta(task)\n const answer = extractAnalysis(artifact)\n return runTriataJudge(meta, answer, judgeRouter())\n },\n }\n}\n"],"mappings":";AA0BA,SAAS,aAAa,cAAc,gBAAgB;AACpD,SAAS,YAAY;AAGrB,IAAM,qBAAqB;AAC3B,IAAM,iBAAiB;AAEvB,IAAM,iBAAiB;AACvB,IAAM,gBAAgB;AAAA,EACpB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,EAAE,KAAK,IAAI;AAGX,IAAM,cAAc;AAgBpB,SAAS,SAAS,GAAW,WAAW,gBAAwB;AAC9D,MAAI;AACF,UAAM,MAAM,aAAa,CAAC;AAC1B,QAAI,IAAI,UAAU,SAAU,QAAO,IAAI,SAAS,MAAM;AACtD,WAAO,IAAI,SAAS,GAAG,QAAQ,EAAE,SAAS,MAAM,IAAI;AAAA,uBAAqB,QAAQ;AAAA,EACnF,QAAQ;AACN,WAAO,gBAAgB,CAAC;AAAA,EAC1B;AACF;AAEA,SAAS,aAAa,KAAuB;AAC3C,QAAM,MAAgB,CAAC;AACvB,MAAI;AACF,eAAW,QAAQ,YAAY,GAAG,GAAG;AACnC,YAAM,OAAO,KAAK,KAAK,IAAI;AAC3B,UAAI;AACF,YAAI,SAAS,IAAI,EAAE,YAAY,EAAG,KAAI,KAAK,GAAG,aAAa,IAAI,CAAC;AAAA,YAC3D,KAAI,KAAK,IAAI;AAAA,MACpB,QAAQ;AAAA,MAER;AAAA,IACF;AAAA,EACF,QAAQ;AAAA,EAER;AACA,SAAO,IAAI,KAAK;AAClB;AAEA,SAAS,eAAe,SAAyB;AAC/C,QAAM,QAAQ,aAAa,OAAO;AAClC,MAAI,MAAM,WAAW,EAAG,QAAO;AAG/B,QAAM,WAAW,CAAC,MAAc;AAC9B,QAAI,EAAE,SAAS,kBAAkB,EAAG,QAAO;AAC3C,QAAI,EAAE,SAAS,eAAe,EAAG,QAAO;AACxC,QAAI,yCAAyC,KAAK,CAAC,KAAK,CAAC,uDAAuD,KAAK,CAAC,EAAG,QAAO;AAChI,QAAI,EAAE,SAAS,aAAa,EAAG,QAAO;AACtC,QAAI,EAAE,SAAS,eAAe,EAAG,QAAO;AACxC,QAAI,EAAE,SAAS,eAAe,EAAG,QAAO;AACxC,WAAO;AAAA,EACT;AACA,QAAM,SAAS,CAAC,GAAG,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,SAAS,CAAC,IAAI,SAAS,CAAC,KAAK,EAAE,cAAc,CAAC,CAAC;AACxF,QAAM,SAAmB,CAAC;AAC1B,MAAI,QAAQ;AACZ,aAAW,KAAK,QAAQ;AACtB,UAAM,MAAM,EAAE,MAAM,QAAQ,SAAS,CAAC;AACtC,UAAM,UAAU,SAAS,CAAC;AAC1B,UAAM,QAAQ;AAAA,iBAAoB,GAAG;AAAA,EAAS,OAAO;AACrD,QAAI,QAAQ,MAAM,SAAS,gBAAgB;AACzC,aAAO,KAAK;AAAA,UAAQ,OAAO,SAAS,OAAO,MAAM,0DAAqD;AACtG;AAAA,IACF;AACA,WAAO,KAAK,KAAK;AACjB,aAAS,MAAM;AAAA,EACjB;AACA,SAAO,OAAO,KAAK,EAAE;AACvB;AAEA,SAAS,oBAAoB,YAAoB,SAAyB;AACxE,QAAM,QAAQ,aAAa,OAAO;AAClC,QAAM,QAAkB,CAAC;AACzB,aAAW,KAAK,OAAO;AACrB,UAAM,MAAM,EAAE,MAAM,QAAQ,SAAS,CAAC;AACtC,UAAM,OAAO,IAAI,MAAM,GAAG,EAAE,IAAI,KAAK;AACrC,QAAI,WAAW,SAAS,GAAG,KAAK,WAAW,SAAS,IAAI,EAAG,OAAM,KAAK,CAAC;AAAA,EACzE;AACA,MAAI,MAAM,WAAW,GAAG;AACtB,WAAO;AAAA,EACT;AACA,QAAM,SAAmB,CAAC;AAC1B,aAAW,KAAK,OAAO;AACrB,UAAM,MAAM,EAAE,MAAM,QAAQ,SAAS,CAAC;AACtC,WAAO,KAAK;AAAA,iBAAoB,GAAG;AAAA,EAAS,SAAS,CAAC,CAAC,EAAE;AAAA,EAC3D;AACA,SAAO,OAAO,KAAK,EAAE;AACvB;AAIA,SAAS,qBAAqB,aAA+D;AAC3F,QAAM,QAAQ,YAAY,MAAM,IAAI;AACpC,QAAM,cAAc,MAAM,OAAiB,CAAC,KAAK,GAAG,MAAM;AACxD,QAAI,SAAS,KAAK,EAAE,KAAK,CAAC,EAAG,KAAI,KAAK,CAAC;AACvC,WAAO;AAAA,EACT,GAAG,CAAC,CAAC;AACL,SAAO,YAAY,IAAI,CAAC,OAAO,MAAM;AACnC,UAAM,MAAM,IAAI,IAAI,YAAY,SAAS,YAAY,IAAI,CAAC,IAAI,MAAM;AACpE,UAAM,QAAQ,MAAM,MAAM,OAAO,GAAG,EAAE,KAAK,IAAI;AAC/C,UAAM,QAAQ,MAAM,KAAK,EAAE,QAAQ,aAAa,EAAE,EAAE,KAAK;AACzD,UAAM,UAAU,MAAM,MAAM,oBAAoB,KAAK,CAAC,GAAG;AACzD,WAAO,EAAE,OAAO,OAAO;AAAA,EACzB,CAAC;AACH;AAEA,SAAS,kBAAkB,QAAwB;AACjD,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,SAAS,GAAG,CAAC,CAAC;AAC5C;AAUA,SAAS,cAA2B;AAClC,QAAM,MAAM,QAAQ,IAAI;AACxB,MAAI,CAAC,IAAK,OAAM,IAAI,MAAM,4DAA4D;AACtF,QAAM,QAAQ,QAAQ,IAAI,eAAe;AACzC,QAAM,UAAU,QAAQ,IAAI,eAAe;AAC3C,SAAO,EAAE,SAAS,KAAK,MAAM;AAC/B;AAEA,SAAS,kBAAkB,KAAsB;AAC/C,MAAI,IAAI,IAAI,KAAK;AACjB,MAAI,EAAE,WAAW,KAAK,EAAG,KAAI,EAAE,MAAM,IAAI,EAAE,MAAM,CAAC,EAAE,KAAK,IAAI;AAC7D,MAAI,EAAE,SAAS,KAAK,EAAG,KAAI,EAAE,MAAM,GAAG,EAAE,YAAY,KAAK,CAAC;AAC1D,MAAI,EAAE,QAAQ,gBAAgB,IAAI,EAAE,KAAK;AACzC,MAAI;AACF,WAAO,KAAK,MAAM,CAAC;AAAA,EACrB,QAAQ;AACN,QAAI;AACF,YAAM,IAAI,EAAE,MAAM,aAAa;AAC/B,UAAI,EAAG,QAAO,KAAK,MAAM,EAAE,CAAC,CAAC;AAAA,IAC/B,QAAQ;AAAA,IAER;AAAA,EACF;AACA,SAAO;AACT;AAEA,eAAe,UAAU,QAAqB,QAAgB,cAAc,GAAqB;AAC/F,WAAS,IAAI,GAAG,IAAI,aAAa,KAAK;AACpC,UAAM,MAAM,MAAM,MAAM,GAAG,OAAO,OAAO,qBAAqB;AAAA,MAC5D,QAAQ;AAAA,MACR,SAAS,EAAE,gBAAgB,oBAAoB,eAAe,UAAU,OAAO,GAAG,GAAG;AAAA,MACrF,MAAM,KAAK,UAAU;AAAA,QACnB,OAAO,OAAO;AAAA,QACd,aAAa;AAAA,QACb,YAAY;AAAA,QACZ,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,OAAO,CAAC;AAAA,MAC9C,CAAC;AAAA,IACH,CAAC;AACD,QAAI,CAAC,IAAI,IAAI;AACX,UAAI,IAAI,cAAc,EAAG;AACzB,YAAM,IAAI,MAAM,oBAAoB,IAAI,MAAM,MAAM,MAAM,IAAI,KAAK,GAAG,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,IACvF;AACA,UAAM,OAAQ,MAAM,IAAI,KAAK;AAC7B,UAAM,UAAU,KAAK,UAAU,CAAC,GAAG,SAAS;AAC5C,QAAI,OAAO,YAAY,SAAU;AACjC,UAAM,SAAS,kBAAkB,OAAO;AACxC,QAAI,WAAW,KAAM,QAAO;AAAA,EAC9B;AACA,SAAO;AACT;AAuBA,eAAe,wBACb,MACA,QACA,QACsB;AACtB,QAAM,aAAa,oBAAoB,QAAQ,KAAK,OAAO;AAC3D,QAAM,cACJ;AAAA;AAAA;AAAA,EAEiB,UAAU;AAAA;AAC7B,QAAM,SAAS,KAAK,aACjB,QAAQ,kBAAkB,WAAW,EACrC,QAAQ,kBAAkB,MAAM;AACnC,QAAM,SAAS,MAAM,UAAU,QAAQ,MAAM;AAC7C,SAAQ,UAA0B,EAAE,yBAAyB,OAAO,qBAAqB,CAAC,EAAE;AAC9F;AAEA,eAAe,kBACb,MACA,QACA,eACA,QACsB;AACtB,QAAM,cACJ,iBAAiB,cAAc,SAAS,IACpC,KAAK,UAAU,eAAe,MAAM,CAAC,IACrC;AACN,QAAM,SAAS,KAAK,aACjB,QAAQ,kBAAkB,KAAK,WAAW,EAC1C,QAAQ,kBAAkB,MAAM,EAChC,QAAQ,0BAA0B,WAAW,EAC7C,QAAQ,gBAAgB,OAAO,KAAK,gBAAgB,MAAM,CAAC;AAC9D,QAAM,SAAS,MAAM,UAAU,QAAQ,MAAM;AAC7C,SAAQ,UAA0B,EAAE,QAAQ,CAAC,EAAE;AACjD;AAEA,eAAe,oBACb,MACA,QACA,QACsB;AACtB,QAAM,SAAS,KAAK,aACjB,QAAQ,kBAAkB,KAAK,WAAW,EAC1C,QAAQ,kBAAkB,MAAM;AACnC,QAAM,SAAS,MAAM,UAAU,QAAQ,MAAM;AAC7C,SAAQ,UAA0B,EAAE,iBAAiB,KAAK;AAC5D;AAEA,eAAe,eAAe,MAAqB,QAAgB,QAA0C;AAC3G,MAAI,CAAC,OAAO,KAAK,GAAG;AAClB,WAAO,EAAE,UAAU,OAAO,OAAO,GAAG,QAAQ,KAAK,UAAU,EAAE,OAAO,eAAe,CAAC,EAAE;AAAA,EACxF;AAGA,QAAM,KAAK,MAAM,wBAAwB,MAAM,QAAQ,MAAM;AAC7D,QAAM,gBAAgB,GAAG,uBAAuB,CAAC;AAGjD,QAAM,CAAC,IAAI,EAAE,IAAI,MAAM,QAAQ,IAAI;AAAA,IACjC,kBAAkB,MAAM,QAAQ,eAAe,MAAM;AAAA,IACrD,oBAAoB,MAAM,QAAQ,MAAM;AAAA,EAC1C,CAAC;AAED,QAAM,cAAc,GAAG,UAAU,CAAC;AAClC,QAAM,YAAsB,CAAC;AAC7B,QAAM,eAAyB,CAAC;AAEhC,WAAS,IAAI,GAAG,IAAI,KAAK,gBAAgB,QAAQ,KAAK;AACpD,UAAM,EAAE,OAAO,OAAO,IAAI,KAAK,gBAAgB,CAAC;AAChD,UAAM,QAAQ,YAAY,CAAC,KAAK,CAAC;AACjC,UAAM,YAAY,MAAM,aAAa,CAAC,GAAG,IAAI,CAAC,MAAc,EAAE,QAAQ,WAAW,EAAE,CAAC;AACpF,UAAM,UAAU,IAAI,KAAK,MAAM,iBAAiB,CAAC,GAAG,IAAI,CAAC,MAAc,EAAE,QAAQ,WAAW,EAAE,CAAC,CAAC;AAChG,UAAM,WAAW,SAAS,OAAO,CAAC,MAAc,CAAC,QAAQ,IAAI,CAAC,CAAC;AAC/D,UAAM,YAAY,kBAAkB,MAAM;AAC1C,QAAI,SAAS,UAAU,UAAW,WAAU,KAAK,KAAK;AAAA,QACjD,cAAa,KAAK,KAAK;AAAA,EAC9B;AAEA,QAAM,YAAY,KAAK,gBAAgB;AACvC,QAAM,QAAQ,GAAG;AACjB,QAAM,eAAe,OAAO,UAAU,YAAY,MAAM,SAAS,KAAK,UAAU;AAEhF,MAAI;AACJ,MAAI,cAAc,EAAG,YAAW;AAAA,WACvB,UAAU,WAAW,aAAa,aAAc,YAAW;AAAA,WAC3D,UAAU,WAAW,UAAW,YAAW;AAAA,WAC3C,UAAU,UAAU,EAAG,YAAW;AAAA,WAClC,UAAU,UAAU,EAAG,YAAW;AAAA,MACtC,YAAW;AAEhB,QAAM,UAAU,GAAG;AACnB,QAAM,iBAAiB,YAAY,QAAQ,YAAY;AAEvD,SAAO;AAAA,IACL,UAAU,aAAa;AAAA,IACvB,OAAO,YAAY,IAAI,WAAW,IAAI;AAAA,IACtC,QAAQ,KAAK,UAAU;AAAA,MACrB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,oBAAoB,cAAc;AAAA,MAClC,WAAW,SAAS;AAAA,MACpB,YAAY,OAAO;AAAA,IACrB,CAAC;AAAA,EACH;AACF;AAIA,SAAS,YAAoB;AAC3B,SAAO,QAAQ,IAAI,oBAAoB;AACzC;AAEA,SAAS,SAAS,QAA0D;AAC1E,QAAM,UAAU,KAAK,UAAU,GAAG,gBAAgB,MAAM;AACxD,QAAM,UAAU,KAAK,SAAS,eAAe,MAAM;AACnD,QAAM,WAAW,KAAK,SAAS,OAAO;AAEtC,QAAM,cAAc,SAAS,KAAK,SAAS,gBAAgB,GAAG,QAAQ;AACtE,QAAM,YAAY,eAAe,OAAO;AACxC,QAAM,cAAc,SAAS,KAAK,UAAU,kBAAkB,GAAG,QAAQ;AACzE,QAAM,eAAe,SAAS,KAAK,UAAU,yBAAyB,GAAG,QAAQ;AACjF,QAAM,eAAe,SAAS,KAAK,UAAU,yBAAyB,GAAG,QAAQ;AACjF,QAAM,eAAe,SAAS,KAAK,UAAU,yBAAyB,GAAG,QAAQ;AAEjF,QAAM,kBAAkB,qBAAqB,WAAW;AAGxD,QAAM,SAAS;AAAA,IACb,YACG;AAAA,MACC;AAAA,MACA;AAAA,IACF,EACC,QAAQ,uDAAuD,iCAAiC,EAChG,QAAQ,oBAAoB,sBAAsB;AAAA,IACrD;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AAEX,QAAM,OAAsB;AAAA,IAC1B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAGA,QAAM,OAAkB;AAAA,IACtB,IAAI;AAAA,IACJ;AAAA,IACA,UAAU;AAAA,EACZ;AAEA,SAAO,EAAE,MAAM,KAAK;AACtB;AAEA,SAAS,SAAS,MAAgC;AAChD,QAAM,KAAK,KAAK;AAChB,MAAI,CAAC,MAAM,OAAO,GAAG,YAAY,YAAY,OAAO,GAAG,gBAAgB,UAAU;AAC/E,UAAM,IAAI,MAAM,cAAc,KAAK,EAAE,+DAA0D;AAAA,EACjG;AACA,SAAO;AACT;AAGA,SAAS,YAAY,MAAwB;AAC3C,MAAI;AACF,WAAO,YAAY,KAAK,MAAM,cAAc,CAAC,EAC1C,OAAO,CAAC,SAAS;AAChB,UAAI;AACF,eAAO,SAAS,KAAK,MAAM,gBAAgB,IAAI,CAAC,EAAE,YAAY;AAAA,MAChE,QAAQ;AACN,eAAO;AAAA,MACT;AAAA,IACF,CAAC,EACA,KAAK;AAAA,EACV,SAAS,KAAK;AACZ,UAAM,IAAI;AAAA,MACR,6CAA6C,KAAK,MAAM,cAAc,CAAC,KAClE,eAAe,QAAQ,IAAI,UAAU,GAAG;AAAA,qFAC2C,IAAI;AAAA,IAC9F;AAAA,EACF;AACF;AAIA,SAAS,gBAAgB,UAA0B;AAGjD,MAAI,SAAS,WAAW,WAAW,EAAG,QAAO,SAAS,MAAM,YAAY,MAAM,EAAE,KAAK;AACrF,QAAM,SAAS,SAAS,QAAQ,aAAa;AAC7C,MAAI,WAAW,GAAI,QAAO,SAAS,MAAM,SAAS,cAAc,MAAM,EAAE,KAAK;AAC7E,SAAO;AACT;AAIO,SAAS,0BAA4C;AAC1D,SAAO;AAAA,IACL,MAAM;AAAA,IAEN,MAAM,YAAY;AAChB,YAAM,OAAO,UAAU;AACvB,YAAM,SAAS,KAAK,MAAM,cAAc;AACxC,UAAI;AACF,cAAM,OAAO,SAAS,MAAM;AAC5B,YAAI,CAAC,KAAK,YAAY,EAAG,OAAM,IAAI,MAAM,iBAAiB;AAAA,MAC5D,QAAQ;AACN,cAAM,IAAI;AAAA,UACR,kCAAkC,MAAM;AAAA,4EACuC,IAAI;AAAA,QACrF;AAAA,MACF;AAEA,kBAAY;AACZ,YAAM,MAAM,YAAY,IAAI;AAC5B,UAAI,IAAI,WAAW,EAAG,OAAM,IAAI,MAAM,mCAAmC,MAAM,EAAE;AACjF,cAAQ;AAAA,QACN,iBAAiB,IAAI,MAAM,sBAAsB,IAAI,sFAErC,QAAQ,IAAI,eAAe,WAAW;AAAA,MACxD;AAAA,IACF;AAAA,IAEA,MAAM,UAAU,OAAoB,CAAC,GAAyB;AAC5D,YAAM,OAAO,UAAU;AACvB,UAAI,MAAM,YAAY,IAAI;AAC1B,UAAI,KAAK,KAAK;AACZ,cAAM,OAAO,IAAI,IAAI,KAAK,GAAG;AAC7B,cAAM,IAAI,OAAO,CAAC,OAAO,KAAK,IAAI,EAAE,CAAC;AAAA,MACvC,WAAW,KAAK,UAAU,QAAW;AACnC,cAAM,IAAI,MAAM,GAAG,KAAK,KAAK;AAAA,MAC/B;AACA,aAAO,IAAI,IAAI,CAAC,OAAO,SAAS,EAAE,EAAE,IAAI;AAAA,IAC1C;AAAA,IAEA,MAAM,aAAa,OAA+C;AAIhE,aAAO;AAAA,IACT;AAAA,IAEA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,OAAO,SAAS,IAAI;AAC1B,YAAM,SAAS,gBAAgB,QAAQ;AACvC,aAAO,eAAe,MAAM,QAAQ,YAAY,CAAC;AAAA,IACnD;AAAA,EACF;AACF;","names":[]}
|