@holdyourvoice/hyv 3.2.0 → 3.3.0

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.
Files changed (50) hide show
  1. package/Readme.md +23 -10
  2. package/dist/ai-editor-rules.js +5 -2
  3. package/dist/ai-editor.js +52 -9
  4. package/dist/ai-editor.test.js +62 -10
  5. package/dist/approval-capability.js +111 -0
  6. package/dist/approval-capability.test.js +52 -0
  7. package/dist/approval-context.js +54 -0
  8. package/dist/approval-context.test.js +38 -0
  9. package/dist/benchmark.js +232 -0
  10. package/dist/benchmark.test.js +328 -0
  11. package/dist/canonical-json.js +123 -0
  12. package/dist/canonical-json.test.js +24 -0
  13. package/dist/cli.js +272 -19
  14. package/dist/cli.test.js +205 -8
  15. package/dist/hygiene.js +6 -0
  16. package/dist/hygiene.test.js +7 -1
  17. package/dist/judgment-task.js +171 -0
  18. package/dist/judgment-task.test.js +162 -0
  19. package/dist/learning.js +240 -100
  20. package/dist/learning.test.js +203 -3
  21. package/dist/lifecycle-adapter.js +75 -0
  22. package/dist/lifecycle-adapter.test.js +56 -0
  23. package/dist/mcp-tools.js +101 -7
  24. package/dist/mcp-tools.test.js +156 -6
  25. package/dist/mcp.js +213 -6
  26. package/dist/mcp.test.js +210 -11
  27. package/dist/pipeline.js +78 -14
  28. package/dist/pipeline.test.js +36 -2
  29. package/dist/preservation.js +89 -0
  30. package/dist/preservation.test.js +22 -0
  31. package/dist/profile.js +87 -0
  32. package/dist/profile.test.js +114 -0
  33. package/dist/rebuild-task.js +226 -0
  34. package/dist/rebuild-task.test.js +179 -0
  35. package/dist/release-audit.test.js +111 -2
  36. package/dist/rewrite-task.js +136 -16
  37. package/dist/rewrite-task.test.js +62 -7
  38. package/dist/rule-reconciliation.test.js +50 -0
  39. package/dist/semantic-review.js +176 -7
  40. package/dist/semantic-review.test.js +98 -14
  41. package/dist/stage1-dry-run.test.js +39 -0
  42. package/dist/stage1-evaluation.js +579 -0
  43. package/dist/stage1-evaluation.test.js +184 -0
  44. package/dist/stage1-human-packet.test.js +102 -0
  45. package/dist/stage1-schema-contract.test.js +95 -0
  46. package/dist/stage2-human-packet.test.js +81 -0
  47. package/dist/version.js +1 -1
  48. package/dist/voice-dna.js +53 -1
  49. package/dist/voice-dna.test.js +79 -1
  50. package/package.json +2 -2
@@ -0,0 +1,579 @@
1
+ import { createHash } from 'node:crypto';
2
+ import { canonicalJson } from './canonical-json.js';
3
+ export const BASELINE_COMMIT = '4e6269121d551c008a34db73077e1e4fea41b3f9';
4
+ export const STAGE1_COMMIT = '550ea24f652291dca13757fdbd2f0fa0b5e3f621';
5
+ export class EvaluationContractError extends Error {
6
+ code;
7
+ constructor(code) {
8
+ super(`Stage 1 evaluation rejected (${code}).`);
9
+ this.code = code;
10
+ this.name = 'EvaluationContractError';
11
+ }
12
+ }
13
+ function fail(code) { throw new EvaluationContractError(code); }
14
+ function record(value, code = 'invalid_record') {
15
+ if (typeof value !== 'object' || value === null || Array.isArray(value))
16
+ fail(code);
17
+ return value;
18
+ }
19
+ function text(value, code = 'invalid_string') {
20
+ if (typeof value !== 'string' || value.length === 0)
21
+ fail(code);
22
+ return value;
23
+ }
24
+ function digest(value, code = 'invalid_digest') {
25
+ const result = text(value, code);
26
+ if (!/^[a-f0-9]{64}$/.test(result))
27
+ fail(code);
28
+ return result;
29
+ }
30
+ function timestamp(value, code) {
31
+ const result = text(value, code);
32
+ if (!/^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(?:\.\d{3})?Z$/.test(result) || !Number.isFinite(Date.parse(result)))
33
+ fail(code);
34
+ const canonical = result.includes('.') ? result : result.replace('Z', '.000Z');
35
+ if (new Date(result).toISOString() !== canonical)
36
+ fail(code);
37
+ return result;
38
+ }
39
+ function exactKeys(value, keys) {
40
+ const allowed = new Set(keys);
41
+ if (Object.keys(value).some((key) => !allowed.has(key)))
42
+ fail('unexpected_field');
43
+ if (keys.some((key) => !(key in value)))
44
+ fail('missing_field');
45
+ }
46
+ function finiteNumber(value, minimum, maximum = Number.POSITIVE_INFINITY, code = 'invalid_number') {
47
+ if (typeof value !== 'number' || !Number.isFinite(value) || value < minimum || value > maximum)
48
+ fail(code);
49
+ return value;
50
+ }
51
+ function count(value, code = 'invalid_count') {
52
+ const result = finiteNumber(value, 0, Number.POSITIVE_INFINITY, code);
53
+ if (!Number.isInteger(result))
54
+ fail(code);
55
+ return result;
56
+ }
57
+ export function sha256Canonical(value) {
58
+ try {
59
+ return createHash('sha256').update(canonicalJson(value)).digest('hex');
60
+ }
61
+ catch {
62
+ return fail('non_canonical_value');
63
+ }
64
+ }
65
+ function sha256Text(value) { return createHash('sha256').update(value).digest('hex'); }
66
+ function wilson95(numerator, denominator) {
67
+ if (denominator === 0)
68
+ return [0, 0];
69
+ const rate = numerator / denominator;
70
+ const z = 1.96;
71
+ const center = (rate + z * z / (2 * denominator)) / (1 + z * z / denominator);
72
+ const spread = z * Math.sqrt((rate * (1 - rate) + z * z / (4 * denominator)) / denominator) / (1 + z * z / denominator);
73
+ return [Math.max(0, center - spread), Math.min(1, center + spread)];
74
+ }
75
+ function parseProtocol(input) {
76
+ const value = record(input);
77
+ exactKeys(value, ['kind', 'version', 'mode', 'baseline', 'stage1', 'benchmark', 'cases', 'intentToTreat', 'execution', 'randomization', 'reviewerRosterDigest', 'analysis', 'rights', 'releaseAuditContractDigest', 'registeredAt', 'measures', 'gates']);
78
+ if (value.kind !== 'hyv-stage1-preregistration' || value.version !== '1')
79
+ fail('protocol_version_invalid');
80
+ if (value.mode !== 'development-calibration' && value.mode !== 'locked-human')
81
+ fail('protocol_mode_invalid');
82
+ const baseline = record(value.baseline);
83
+ exactKeys(baseline, ['packageVersion', 'sourceCommit']);
84
+ if (baseline.packageVersion !== '3.2.0' || baseline.sourceCommit !== BASELINE_COMMIT)
85
+ fail('baseline_identity_mismatch');
86
+ const stage1 = record(value.stage1);
87
+ exactKeys(stage1, ['sourceCommit']);
88
+ if (stage1.sourceCommit !== STAGE1_COMMIT)
89
+ fail('stage1_identity_mismatch');
90
+ const benchmark = record(value.benchmark);
91
+ exactKeys(benchmark, ['manifestDigest', 'partitionDigest', 'partition', 'synthetic']);
92
+ digest(benchmark.manifestDigest);
93
+ digest(benchmark.partitionDigest);
94
+ text(benchmark.partition);
95
+ if (typeof benchmark.synthetic !== 'boolean')
96
+ fail('benchmark_synthetic_invalid');
97
+ if ((value.mode === 'locked-human') === benchmark.synthetic)
98
+ fail('evidence_class_mismatch');
99
+ if (value.mode === 'locked-human' && benchmark.partition !== 'locked-test')
100
+ fail('benchmark_partition_mismatch');
101
+ if (!Array.isArray(value.cases) || value.cases.length === 0)
102
+ fail('cases_invalid');
103
+ const caseIds = new Set();
104
+ for (const item of value.cases) {
105
+ const entry = record(item);
106
+ exactKeys(entry, ['caseId', 'caseDigest', 'provenanceDigest', 'rightsDigest', 'providerDisclosureDigest', 'reviewerDisclosureDigest', 'derivedRetentionDigest']);
107
+ const id = text(entry.caseId);
108
+ for (const key of ['caseDigest', 'provenanceDigest', 'rightsDigest', 'providerDisclosureDigest', 'reviewerDisclosureDigest', 'derivedRetentionDigest'])
109
+ digest(entry[key]);
110
+ const expectedRights = sha256Canonical({ caseId: entry.caseId, caseDigest: entry.caseDigest, provenanceDigest: entry.provenanceDigest, providerDisclosureDigest: entry.providerDisclosureDigest, reviewerDisclosureDigest: entry.reviewerDisclosureDigest, derivedRetentionDigest: entry.derivedRetentionDigest });
111
+ if (entry.rightsDigest !== expectedRights)
112
+ fail('case_rights_digest_mismatch');
113
+ if (caseIds.has(id))
114
+ fail('duplicate_case');
115
+ caseIds.add(id);
116
+ }
117
+ const itt = record(value.intentToTreat);
118
+ exactKeys(itt, ['expectedAssignments', 'expectedReviewers']);
119
+ if (!Number.isInteger(itt.expectedAssignments) || itt.expectedAssignments < 1 || !Number.isInteger(itt.expectedReviewers) || itt.expectedReviewers < 1)
120
+ fail('itt_invalid');
121
+ const execution = record(value.execution);
122
+ exactKeys(execution, ['provider', 'model', 'modelRevision', 'settingsDigest', 'taskContractDigest', 'rulesetDigest', 'rubricDigest']);
123
+ text(execution.provider);
124
+ text(execution.model);
125
+ text(execution.modelRevision);
126
+ for (const key of ['settingsDigest', 'taskContractDigest', 'rulesetDigest', 'rubricDigest'])
127
+ digest(execution[key]);
128
+ const randomization = record(value.randomization);
129
+ exactKeys(randomization, ['algorithm', 'commitment']);
130
+ text(randomization.algorithm);
131
+ digest(randomization.commitment);
132
+ digest(value.reviewerRosterDigest);
133
+ const analysis = record(value.analysis);
134
+ exactKeys(analysis, ['statistic', 'decisionRule', 'margin', 'minimumCases', 'minimumRatings', 'missingRatingPolicy', 'tiePolicy', 'retryPolicy', 'routingPolicy']);
135
+ if (analysis.statistic !== 'paired-preference-rate' || analysis.decisionRule !== 'preference-or-correction-with-workflow-non-regression')
136
+ fail('analysis_statistic_invalid');
137
+ if (typeof analysis.margin !== 'number' || analysis.margin < 0 || analysis.margin > 1)
138
+ fail('analysis_margin_invalid');
139
+ if (!Number.isInteger(analysis.minimumCases) || analysis.minimumCases < 1 || !Number.isInteger(analysis.minimumRatings) || analysis.minimumRatings < 1)
140
+ fail('analysis_minimum_invalid');
141
+ if (analysis.missingRatingPolicy !== 'count-as-non-preference' || analysis.tiePolicy !== 'count-as-half' || analysis.retryPolicy !== 'no-retry' || analysis.routingPolicy !== 'precommitted-blind-routing')
142
+ fail('analysis_policy_invalid');
143
+ const rights = record(value.rights);
144
+ exactKeys(rights, ['providerDisclosureDigest', 'reviewerDisclosureDigest', 'derivedRetentionDigest']);
145
+ digest(rights.providerDisclosureDigest);
146
+ digest(rights.reviewerDisclosureDigest);
147
+ digest(rights.derivedRetentionDigest);
148
+ digest(value.releaseAuditContractDigest);
149
+ if (value.cases.some((item) => { const entry = item; return entry.providerDisclosureDigest !== rights.providerDisclosureDigest || entry.reviewerDisclosureDigest !== rights.reviewerDisclosureDigest || entry.derivedRetentionDigest !== rights.derivedRetentionDigest; }))
150
+ fail('rights_scope_mismatch');
151
+ timestamp(value.registeredAt, 'registered_at_invalid');
152
+ if (JSON.stringify(value.measures) !== JSON.stringify(['writer_preference', 'correction_versus_confirm', 'workflow_completion', 'workflow_abandonment']))
153
+ fail('measures_invalid');
154
+ if (JSON.stringify(value.gates) !== JSON.stringify(['semantic', 'copy_spec', 'hygiene', 'preservation', 'cli_mcp_parity', 'backward_compatibility']))
155
+ fail('gates_invalid');
156
+ return value;
157
+ }
158
+ export function commitProtocol(input) {
159
+ const protocol = parseProtocol(input);
160
+ return { ...protocol, protocolDigest: sha256Canonical(protocol) };
161
+ }
162
+ function assertCommittedProtocol(input) {
163
+ const value = record(input);
164
+ const { protocolDigest, ...protocol } = value;
165
+ const parsed = parseProtocol(protocol);
166
+ if (digest(protocolDigest, 'protocol_digest_invalid') !== sha256Canonical(parsed))
167
+ fail('protocol_digest_mismatch');
168
+ return value;
169
+ }
170
+ function effectiveRunOutcome(run) {
171
+ if (run.outcome === 'completed' && run.hardGates && Object.values(run.hardGates).some((passed) => !passed))
172
+ return 'hard-gate-failed';
173
+ return run.outcome;
174
+ }
175
+ function candidateDigest(protocol) {
176
+ return sha256Canonical({ baseline: protocol.baseline, stage1: protocol.stage1 });
177
+ }
178
+ function parseRun(protocol, input) {
179
+ const value = record(input);
180
+ const required = ['kind', 'version', 'evidenceClass', 'protocolDigest', 'candidateDigest', 'assignmentId', 'participantId', 'caseId', 'caseDigest', 'rightsDigest', 'arm', 'ordinal', 'latencyMs', 'inputTokens', 'outputTokens', 'costMicrousd', 'occurredAt', 'outcome'];
181
+ exactKeys(value, value.outcome === 'completed' ? [...required, 'outputDigest', 'hardGates'] : required);
182
+ if (value.kind !== 'hyv-stage1-run-record' || value.version !== '1')
183
+ fail('run_version_invalid');
184
+ if (value.protocolDigest !== protocol.protocolDigest)
185
+ fail('protocol_digest_mismatch');
186
+ if (value.candidateDigest !== candidateDigest(protocol))
187
+ fail('candidate_digest_mismatch');
188
+ if (value.evidenceClass !== (protocol.mode === 'locked-human' ? 'human' : 'synthetic-dry-run'))
189
+ fail('evidence_class_mismatch');
190
+ text(value.assignmentId);
191
+ text(value.participantId);
192
+ const caseId = text(value.caseId);
193
+ const expected = protocol.cases.find((item) => item.caseId === caseId);
194
+ if (!expected)
195
+ fail('case_unknown');
196
+ if (value.caseDigest !== expected.caseDigest)
197
+ fail('case_digest_mismatch');
198
+ if (value.rightsDigest !== expected.rightsDigest)
199
+ fail('rights_digest_mismatch');
200
+ if (value.arm !== 'baseline' && value.arm !== 'stage1')
201
+ fail('arm_invalid');
202
+ if (!Number.isInteger(value.ordinal) || value.ordinal < 1)
203
+ fail('ordinal_invalid');
204
+ for (const key of ['latencyMs', 'inputTokens', 'outputTokens', 'costMicrousd'])
205
+ if (typeof value[key] !== 'number' || value[key] < 0 || !Number.isFinite(value[key]))
206
+ fail('run_measure_invalid');
207
+ timestamp(value.occurredAt, 'occurred_at_invalid');
208
+ if (!['completed', 'hard-failure', 'timeout', 'abandoned'].includes(String(value.outcome)))
209
+ fail('outcome_invalid');
210
+ if (value.outcome === 'completed')
211
+ digest(value.outputDigest, 'output_digest_invalid');
212
+ if (value.outcome === 'completed') {
213
+ const gates = record(value.hardGates);
214
+ exactKeys(gates, ['semantic', 'copySpec', 'hygiene', 'preservation', 'cliMcpParity', 'backwardCompatibility']);
215
+ if (Object.values(gates).some((item) => typeof item !== 'boolean'))
216
+ fail('hard_gates_invalid');
217
+ }
218
+ return value;
219
+ }
220
+ export function validateRuns(protocolInput, runInputs) {
221
+ const protocol = assertCommittedProtocol(protocolInput);
222
+ if (!Array.isArray(runInputs))
223
+ fail('runs_invalid');
224
+ const runs = runInputs.map((item) => parseRun(protocol, item));
225
+ if (runs.length !== protocol.intentToTreat.expectedAssignments)
226
+ fail('itt_denominator_mismatch');
227
+ if (new Set(runs.map((run) => run.assignmentId)).size !== runs.length)
228
+ fail('duplicate_assignment');
229
+ if (JSON.stringify(runs.map((run) => run.ordinal).sort((a, b) => a - b)) !== JSON.stringify(Array.from({ length: runs.length }, (_, index) => index + 1)))
230
+ fail('ordinal_sequence_invalid');
231
+ for (const item of protocol.cases) {
232
+ const pair = runs.filter((run) => run.caseId === item.caseId);
233
+ const arms = pair.map((run) => run.arm).sort();
234
+ if (JSON.stringify(arms) !== JSON.stringify(['baseline', 'stage1']))
235
+ fail('paired_assignment_mismatch');
236
+ if (new Set(pair.map((run) => run.participantId)).size !== 1)
237
+ fail('paired_participant_mismatch');
238
+ }
239
+ return { denominator: runs.length, completed: runs.filter((r) => effectiveRunOutcome(r) === 'completed').length, hardFailures: runs.filter((r) => ['hard-failure', 'hard-gate-failed'].includes(effectiveRunOutcome(r))).length, timeouts: runs.filter((r) => r.outcome === 'timeout').length, abandonments: runs.filter((r) => r.outcome === 'abandoned').length, runsDigest: sha256Canonical(runs) };
240
+ }
241
+ function parsePacket(input) {
242
+ const packet = record(input);
243
+ exactKeys(packet, ['kind', 'version', 'protocolDigest', 'candidateDigest', 'runsDigest', 'mappingDigest', 'contentDigest', 'nonReviewableDigest', 'reviewableCount', 'nonReviewableCount', 'encryptedAtRest', 'approvedStorage', 'packetDigest']);
244
+ if (packet.kind !== 'hyv-stage1-blind-packet' || packet.version !== '1' || packet.encryptedAtRest !== true || packet.approvedStorage !== true)
245
+ fail('blind_packet_invalid');
246
+ for (const key of ['protocolDigest', 'candidateDigest', 'runsDigest', 'mappingDigest', 'contentDigest', 'nonReviewableDigest', 'packetDigest'])
247
+ digest(packet[key]);
248
+ if (!Number.isInteger(packet.reviewableCount) || !Number.isInteger(packet.nonReviewableCount) || packet.reviewableCount < 0 || packet.nonReviewableCount < 0)
249
+ fail('blind_packet_invalid');
250
+ const { packetDigest, ...base } = packet;
251
+ if (packetDigest !== sha256Canonical(base))
252
+ fail('packet_digest_mismatch');
253
+ return packet;
254
+ }
255
+ function parseMapping(input) {
256
+ const mapping = record(input);
257
+ exactKeys(mapping, ['kind', 'version', 'nonce', 'labels', 'custodianId', 'unblindingAccess', 'attestation']);
258
+ if (mapping.kind !== 'hyv-stage1-blind-mapping' || mapping.version !== '1')
259
+ fail('blind_mapping_invalid');
260
+ const labels = record(mapping.labels);
261
+ exactKeys(labels, ['A', 'B']);
262
+ if (!['baseline', 'stage1'].includes(String(labels.A)) || !['baseline', 'stage1'].includes(String(labels.B)) || labels.A === labels.B)
263
+ fail('blind_mapping_invalid');
264
+ if (!/^[A-Za-z0-9_-]{22,}$/.test(text(mapping.nonce)))
265
+ fail('randomization_nonce_invalid');
266
+ text(mapping.custodianId);
267
+ if (!Array.isArray(mapping.unblindingAccess) || mapping.unblindingAccess.length === 0 || new Set(mapping.unblindingAccess).size !== mapping.unblindingAccess.length || !mapping.unblindingAccess.every((item) => typeof item === 'string' && item.length > 0))
268
+ fail('unblinding_access_invalid');
269
+ if (mapping.attestation !== null)
270
+ validateAttestation(mapping.attestation);
271
+ return mapping;
272
+ }
273
+ function mappingCore(mapping) { const { attestation: _attestation, ...core } = mapping; return core; }
274
+ function validateMappingOpening(protocol, mapping) {
275
+ if (protocol.randomization.algorithm !== 'sha256-counter-v1' || protocol.randomization.commitment !== sha256Canonical({ algorithm: protocol.randomization.algorithm, nonce: mapping.nonce, labels: mapping.labels }))
276
+ fail('randomization_commitment_mismatch');
277
+ }
278
+ export function freezeBlind(protocolInput, runInputs, mappingInput, contentsInput, nonReviewableInput = []) {
279
+ const protocol = assertCommittedProtocol(protocolInput);
280
+ const runSummary = validateRuns(protocol, runInputs);
281
+ const mapping = parseMapping(mappingInput);
282
+ validateMappingOpening(protocol, mapping);
283
+ if (!Array.isArray(contentsInput))
284
+ fail('blind_contents_invalid');
285
+ const contents = contentsInput.map((item) => { const entry = record(item); exactKeys(entry, ['caseId', 'A', 'B']); text(entry.caseId); text(entry.A); text(entry.B); return entry; });
286
+ if (/baseline|stage\s*[-_ ]?1|550ea24|4e626912/i.test(canonicalJson(contents)))
287
+ fail('blind_label_leak');
288
+ if (!Array.isArray(nonReviewableInput))
289
+ fail('non_reviewable_invalid');
290
+ const allowedOutcomes = ['completed', 'hard-failure', 'hard-gate-failed', 'timeout', 'abandoned'];
291
+ const nonReviewable = nonReviewableInput.map((item) => { const entry = record(item); exactKeys(entry, ['caseId', 'baselineOutcome', 'stage1Outcome']); text(entry.caseId); if (!allowedOutcomes.includes(String(entry.baselineOutcome)) || !allowedOutcomes.includes(String(entry.stage1Outcome)))
292
+ fail('non_reviewable_invalid'); return entry; });
293
+ const completedCases = protocol.cases.filter((item) => runInputs.filter((run) => run.caseId === item.caseId && effectiveRunOutcome(run) === 'completed').length === 2).map((item) => item.caseId);
294
+ const incompleteCases = protocol.cases.map((item) => item.caseId).filter((id) => !completedCases.includes(id));
295
+ if (contents.length !== completedCases.length || new Set(contents.map((item) => item.caseId)).size !== completedCases.length || contents.some((item) => !completedCases.includes(item.caseId)))
296
+ fail('blind_case_set_mismatch');
297
+ if (nonReviewable.length !== incompleteCases.length || new Set(nonReviewable.map((item) => item.caseId)).size !== incompleteCases.length || nonReviewable.some((item) => !incompleteCases.includes(item.caseId)))
298
+ fail('non_reviewable_mismatch');
299
+ for (const entry of nonReviewable)
300
+ for (const arm of ['baseline', 'stage1']) {
301
+ const run = runInputs.find((candidate) => candidate.caseId === entry.caseId && candidate.arm === arm);
302
+ if (!run || effectiveRunOutcome(run) !== entry[`${arm}Outcome`])
303
+ fail('non_reviewable_mismatch');
304
+ }
305
+ for (const entry of contents) {
306
+ for (const label of ['A', 'B']) {
307
+ const arm = mapping.labels[label];
308
+ const run = runInputs.find((item) => item.caseId === entry.caseId && item.arm === arm && effectiveRunOutcome(item) === 'completed');
309
+ if (!run || run.outputDigest !== sha256Text(entry[label]))
310
+ fail('blind_content_digest_mismatch');
311
+ }
312
+ }
313
+ const base = { kind: 'hyv-stage1-blind-packet', version: '1', protocolDigest: protocol.protocolDigest, candidateDigest: candidateDigest(protocol), runsDigest: runSummary.runsDigest, mappingDigest: sha256Canonical(mappingCore(mapping)), contentDigest: sha256Canonical(contents), nonReviewableDigest: sha256Canonical(nonReviewable), reviewableCount: contents.length, nonReviewableCount: nonReviewable.length, encryptedAtRest: true, approvedStorage: true };
314
+ return { packet: { ...base, packetDigest: sha256Canonical(base) } };
315
+ }
316
+ function validateAttestation(value, artifactDigest, protocolDigest, packetDigest) {
317
+ const item = record(value, 'human_attestation_required');
318
+ exactKeys(item, ['kind', 'version', 'verified', 'artifactDigest', 'trustStoreDigest', 'keyId', 'purpose', 'protocolDigest', 'packetDigest', 'nonce', 'issuedAt', 'expiresAt', 'verifiedAt']);
319
+ if (item.kind !== 'external-verification-receipt' || item.version !== '1' || item.verified !== true)
320
+ fail('human_attestation_required');
321
+ if (artifactDigest && item.artifactDigest !== artifactDigest)
322
+ fail('attestation_binding_mismatch');
323
+ else
324
+ digest(item.artifactDigest, 'human_attestation_required');
325
+ digest(item.trustStoreDigest, 'human_attestation_required');
326
+ text(item.keyId, 'human_attestation_required');
327
+ text(item.purpose, 'human_attestation_required');
328
+ text(item.nonce, 'human_attestation_required');
329
+ if (protocolDigest && item.protocolDigest !== protocolDigest)
330
+ fail('attestation_binding_mismatch');
331
+ else
332
+ digest(item.protocolDigest, 'human_attestation_required');
333
+ if (packetDigest && item.packetDigest !== packetDigest)
334
+ fail('attestation_binding_mismatch');
335
+ else
336
+ digest(item.packetDigest, 'human_attestation_required');
337
+ for (const key of ['issuedAt', 'expiresAt', 'verifiedAt'])
338
+ timestamp(item[key], 'human_attestation_required');
339
+ if (Date.parse(item.expiresAt) <= Date.parse(item.issuedAt) || Date.parse(item.verifiedAt) < Date.parse(item.issuedAt) || Date.parse(item.verifiedAt) > Date.parse(item.expiresAt))
340
+ fail('human_attestation_required');
341
+ }
342
+ function parseRating(packet, input) {
343
+ const value = record(input);
344
+ const baseKeys = ['kind', 'version', 'evidenceClass', 'protocolDigest', 'candidateDigest', 'packetDigest', 'reviewerId', 'identityKey', 'recordId', 'sequence', 'previousRecordDigest', 'recordDigest', 'caseId', 'workflow', 'attestation'];
345
+ exactKeys(value, value.workflow === 'completed' ? [...baseKeys, 'preferredLabel', 'correctionVersusConfirm'] : baseKeys);
346
+ if (value.kind !== 'hyv-stage1-reviewer-record' || value.version !== '1')
347
+ fail('reviewer_record_version_invalid');
348
+ if (value.protocolDigest !== packet.protocolDigest || value.candidateDigest !== packet.candidateDigest || value.packetDigest !== packet.packetDigest)
349
+ fail('reviewer_digest_chain_mismatch');
350
+ const reviewerId = text(value.reviewerId);
351
+ text(value.caseId);
352
+ digest(value.identityKey, 'reviewer_identity_invalid');
353
+ digest(value.recordId, 'reviewer_record_id_invalid');
354
+ digest(value.recordDigest, 'reviewer_record_digest_invalid');
355
+ if (value.identityKey !== sha256Canonical({ reviewerId }))
356
+ fail('reviewer_identity_mismatch');
357
+ if (value.recordId !== sha256Canonical({ protocolDigest: packet.protocolDigest, packetDigest: packet.packetDigest, identityKey: value.identityKey, caseId: value.caseId }))
358
+ fail('reviewer_record_id_mismatch');
359
+ if (!Number.isInteger(value.sequence) || value.sequence < 1)
360
+ fail('reviewer_sequence_invalid');
361
+ if (value.previousRecordDigest !== null)
362
+ digest(value.previousRecordDigest, 'previous_record_digest_invalid');
363
+ if (!['completed', 'abandoned'].includes(String(value.workflow)))
364
+ fail('workflow_invalid');
365
+ if (value.workflow === 'completed') {
366
+ if (!['A', 'B', 'tie'].includes(String(value.preferredLabel)))
367
+ fail('preference_invalid');
368
+ const correction = record(value.correctionVersusConfirm);
369
+ exactKeys(correction, ['A', 'B']);
370
+ if (!['correction', 'confirm'].includes(String(correction.A)) || !['correction', 'confirm'].includes(String(correction.B)))
371
+ fail('correction_measure_invalid');
372
+ }
373
+ const { attestation, recordDigest, ...core } = value;
374
+ if (recordDigest !== sha256Canonical(core))
375
+ fail('reviewer_record_digest_mismatch');
376
+ if (value.evidenceClass === 'human') {
377
+ validateAttestation(attestation, recordDigest, packet.protocolDigest, packet.packetDigest);
378
+ if (attestation.purpose !== 'stage1-blind-review')
379
+ fail('attestation_purpose_mismatch');
380
+ }
381
+ else if (value.evidenceClass !== 'synthetic-dry-run' || attestation !== null)
382
+ fail('evidence_class_mismatch');
383
+ return value;
384
+ }
385
+ export function recordRating(packetInput, existingInput, ratingInput) {
386
+ const packet = parsePacket(packetInput);
387
+ if (!Array.isArray(existingInput))
388
+ fail('reviewer_log_invalid');
389
+ const existing = existingInput.map((item) => parseRating(packet, item));
390
+ const rating = parseRating(packet, ratingInput);
391
+ const same = existing.find((item) => item.identityKey === rating.identityKey && item.caseId === rating.caseId);
392
+ if (same)
393
+ fail(same.preferredLabel === rating.preferredLabel && (same.correctionVersusConfirm === undefined ? rating.correctionVersusConfirm === undefined : sha256Canonical(same.correctionVersusConfirm) === sha256Canonical(rating.correctionVersusConfirm)) && same.workflow === rating.workflow ? 'duplicate_reviewer_record' : 'conflicting_reviewer_record');
394
+ for (let index = 0; index < existing.length; index += 1) {
395
+ if (existing[index].sequence !== index + 1 || existing[index].previousRecordDigest !== (index === 0 ? null : existing[index - 1].recordDigest))
396
+ fail('reviewer_chain_invalid');
397
+ }
398
+ if (rating.sequence !== existing.length + 1 || rating.previousRecordDigest !== (existing.length === 0 ? null : existing[existing.length - 1].recordDigest))
399
+ fail('reviewer_chain_invalid');
400
+ return [...existing, rating];
401
+ }
402
+ function validateReviewerLog(protocol, packet, ratingsInput, requireComplete) {
403
+ if (!Array.isArray(ratingsInput))
404
+ fail('reviewer_log_invalid');
405
+ const ratings = ratingsInput.map((item) => parseRating(packet, item));
406
+ for (let index = 0; index < ratings.length; index += 1) {
407
+ if (ratings[index].sequence !== index + 1 || ratings[index].previousRecordDigest !== (index === 0 ? null : ratings[index - 1].recordDigest))
408
+ fail('reviewer_chain_invalid');
409
+ if (protocol && !protocol.cases.some((item) => item.caseId === ratings[index].caseId))
410
+ fail('reviewer_case_unknown');
411
+ }
412
+ const pairs = ratings.map((item) => `${item.identityKey}\0${item.caseId}`);
413
+ if (new Set(pairs).size !== pairs.length)
414
+ fail('duplicate_reviewer_record');
415
+ if (protocol && requireComplete) {
416
+ const identities = new Set(ratings.map((item) => item.identityKey));
417
+ if (identities.size !== protocol.intentToTreat.expectedReviewers || ratings.length !== protocol.intentToTreat.expectedReviewers * protocol.cases.length)
418
+ fail('reviewer_matrix_mismatch');
419
+ for (const identity of identities)
420
+ for (const item of protocol.cases)
421
+ if (!pairs.includes(`${identity}\0${item.caseId}`))
422
+ fail('reviewer_matrix_mismatch');
423
+ }
424
+ return ratings;
425
+ }
426
+ export function sealRatings(packetInput, mappingInput, ratingsInput) {
427
+ const packet = parsePacket(packetInput);
428
+ const mapping = parseMapping(mappingInput);
429
+ if (sha256Canonical(mappingCore(mapping)) !== packet.mappingDigest)
430
+ fail('mapping_digest_mismatch');
431
+ const ratings = validateReviewerLog(undefined, packet, ratingsInput, false);
432
+ const base = { kind: 'hyv-stage1-ratings-seal', version: '1', protocolDigest: packet.protocolDigest, candidateDigest: packet.candidateDigest, packetDigest: packet.packetDigest, mappingDigest: packet.mappingDigest, ratingsDigest: sha256Canonical(ratings), recordCount: ratings.length };
433
+ return { ...base, sealDigest: sha256Canonical(base) };
434
+ }
435
+ export function reduceEvaluation(protocolInput, runsInput, packetInput, mappingInput, ratingsInput, sealInput, releaseAuditInput) {
436
+ const protocol = assertCommittedProtocol(protocolInput);
437
+ const summary = validateRuns(protocol, runsInput);
438
+ const packet = parsePacket(packetInput);
439
+ if (packet.protocolDigest !== protocol.protocolDigest || packet.candidateDigest !== candidateDigest(protocol) || packet.runsDigest !== summary.runsDigest)
440
+ fail('packet_digest_chain_mismatch');
441
+ const runRows = runsInput;
442
+ const expectedReviewableCount = protocol.cases.filter((entry) => runRows.filter((run) => run.caseId === entry.caseId && effectiveRunOutcome(run) === 'completed').length === 2).length;
443
+ if (packet.reviewableCount !== expectedReviewableCount || packet.nonReviewableCount !== protocol.cases.length - expectedReviewableCount)
444
+ fail('blind_packet_count_mismatch');
445
+ const mapping = parseMapping(mappingInput);
446
+ validateMappingOpening(protocol, mapping);
447
+ const ratings = validateReviewerLog(protocol, packet, ratingsInput, true);
448
+ const seal = sealRatings(packet, mapping, ratings);
449
+ if (sha256Canonical(seal) !== sha256Canonical(sealInput))
450
+ fail('ratings_seal_mismatch');
451
+ const audit = record(releaseAuditInput);
452
+ exactKeys(audit, ['kind', 'version', 'candidateCommit', 'protocolDigest', 'contractDigest', 'passed', 'digest']);
453
+ const { digest: auditDigest, ...auditCore } = audit;
454
+ if (auditDigest !== sha256Canonical(auditCore))
455
+ fail('release_audit_digest_mismatch');
456
+ if (audit.kind !== 'hyv-release-audit' || audit.version !== '1' || audit.candidateCommit !== protocol.stage1.sourceCommit || audit.protocolDigest !== protocol.protocolDigest || audit.contractDigest !== protocol.releaseAuditContractDigest || audit.passed !== true)
457
+ fail('release_audit_binding_mismatch');
458
+ const blockers = ['human_writer_evidence_deferred'];
459
+ if (protocol.benchmark.synthetic || ratings.some((r) => r.evidenceClass === 'synthetic-dry-run'))
460
+ blockers.push('synthetic_fixture_evidence');
461
+ if (protocol.mode !== 'locked-human')
462
+ blockers.push('locked_human_evidence_required');
463
+ if (protocol.mode === 'locked-human') {
464
+ blockers.push('reviewer_roster_verification_deferred');
465
+ const { attestation, ...openingCore } = mapping;
466
+ try {
467
+ validateAttestation(attestation, sha256Canonical(openingCore), protocol.protocolDigest, packet.packetDigest);
468
+ if (attestation.purpose !== 'stage1-mapping-custody')
469
+ throw new Error('purpose');
470
+ }
471
+ catch {
472
+ blockers.push('mapping_custody_attestation_required');
473
+ }
474
+ }
475
+ if (ratings.length !== protocol.intentToTreat.expectedReviewers * protocol.cases.length)
476
+ blockers.push('reviewer_denominator_mismatch');
477
+ if (protocol.cases.length < protocol.analysis.minimumCases || ratings.length < protocol.analysis.minimumRatings)
478
+ blockers.push('preregistered_minimum_not_met');
479
+ if (runsInput.some((run) => run.hardGates && Object.values(run.hardGates).some((passed) => !passed)))
480
+ blockers.push('hard_gate_regression');
481
+ const completedRatings = ratings.filter((rating) => rating.workflow === 'completed');
482
+ const preferred = completedRatings.reduce((total, rating) => total + (rating.preferredLabel === 'tie' ? 0.5 : mapping.labels[rating.preferredLabel] === 'stage1' ? 1 : 0), 0);
483
+ const preferenceDenominator = protocol.intentToTreat.expectedReviewers * protocol.cases.length;
484
+ const preferenceRate = preferenceDenominator ? preferred / preferenceDenominator : 0;
485
+ const stage1Confirm = completedRatings.filter((rating) => rating.correctionVersusConfirm[mapping.labels.A === 'stage1' ? 'A' : 'B'] === 'confirm').length;
486
+ const baselineConfirm = completedRatings.filter((rating) => rating.correctionVersusConfirm[mapping.labels.A === 'baseline' ? 'A' : 'B'] === 'confirm').length;
487
+ const correctionMarginMet = (stage1Confirm - baselineConfirm) / preferenceDenominator >= protocol.analysis.margin;
488
+ if (preferenceRate < 0.5 + protocol.analysis.margin && !correctionMarginMet)
489
+ blockers.push('checkpoint_threshold_not_met');
490
+ const workflowByArm = (arm) => ({ completed: runRows.filter((run) => run.arm === arm && effectiveRunOutcome(run) === 'completed').length, denominator: runRows.filter((run) => run.arm === arm).length });
491
+ const baselineWorkflow = workflowByArm('baseline');
492
+ const stage1Workflow = workflowByArm('stage1');
493
+ if (stage1Workflow.denominator === 0 || baselineWorkflow.denominator === 0 || stage1Workflow.completed / stage1Workflow.denominator < baselineWorkflow.completed / baselineWorkflow.denominator)
494
+ blockers.push('workflow_regression');
495
+ const stage1Correction = completedRatings.filter((rating) => rating.correctionVersusConfirm[mapping.labels.A === 'stage1' ? 'A' : 'B'] === 'correction').length;
496
+ const baselineCorrection = completedRatings.filter((rating) => rating.correctionVersusConfirm[mapping.labels.A === 'baseline' ? 'A' : 'B'] === 'correction').length;
497
+ const metrics = {
498
+ preference: { numerator: preferred, denominator: preferenceDenominator, rate: preferenceRate, uncertainty95: wilson95(preferred, preferenceDenominator) },
499
+ correctionVersusConfirm: { stage1: { corrections: stage1Correction, confirms: stage1Confirm, denominator: preferenceDenominator, correctionRate: stage1Correction / preferenceDenominator, uncertainty95: wilson95(stage1Correction, preferenceDenominator) }, baseline: { corrections: baselineCorrection, confirms: baselineConfirm, denominator: preferenceDenominator, correctionRate: baselineCorrection / preferenceDenominator, uncertainty95: wilson95(baselineCorrection, preferenceDenominator) } },
500
+ completion: { numerator: completedRatings.length, denominator: preferenceDenominator, rate: completedRatings.length / preferenceDenominator, uncertainty95: wilson95(completedRatings.length, preferenceDenominator) },
501
+ abandonment: { numerator: ratings.length - completedRatings.length, denominator: preferenceDenominator, rate: (ratings.length - completedRatings.length) / preferenceDenominator, uncertainty95: wilson95(ratings.length - completedRatings.length, preferenceDenominator) },
502
+ providerRuns: { completed: summary.completed, abandoned: summary.abandonments, hardFailures: summary.hardFailures, timeouts: summary.timeouts, denominator: summary.denominator },
503
+ };
504
+ const intentToTreat = { ...summary, expectedRatings: preferenceDenominator, observedRatings: ratings.length, missingRatings: Math.max(0, preferenceDenominator - ratings.length), reconciled: summary.denominator === protocol.intentToTreat.expectedAssignments && ratings.length === preferenceDenominator };
505
+ const base = { kind: 'hyv-stage1-evaluation-report', version: '1', protocolDigest: protocol.protocolDigest, candidateDigest: candidateDigest(protocol), runsDigest: summary.runsDigest, packetDigest: packet.packetDigest, mappingDigest: packet.mappingDigest, sealDigest: seal.sealDigest, releaseAuditDigest: audit.digest, intentToTreat, metrics, decision: blockers.length ? 'BLOCKED' : 'PASS', promotable: blockers.length === 0, blockers };
506
+ return { ...base, reportDigest: sha256Canonical(base) };
507
+ }
508
+ export function preflight(protocolInput) {
509
+ const protocol = assertCommittedProtocol(protocolInput);
510
+ return { ready: true, protocolDigest: protocol.protocolDigest, humanEvidenceRequired: protocol.mode === 'locked-human' };
511
+ }
512
+ export function recordCheckpointDisposition(reportInput, disposition, attestation) {
513
+ const report = record(reportInput);
514
+ exactKeys(report, ['kind', 'version', 'protocolDigest', 'candidateDigest', 'runsDigest', 'packetDigest', 'mappingDigest', 'sealDigest', 'releaseAuditDigest', 'intentToTreat', 'metrics', 'decision', 'promotable', 'blockers', 'reportDigest']);
515
+ if (report.kind !== 'hyv-stage1-evaluation-report' || report.version !== '1')
516
+ fail('report_contract_invalid');
517
+ for (const field of ['protocolDigest', 'candidateDigest', 'runsDigest', 'packetDigest', 'mappingDigest', 'sealDigest', 'releaseAuditDigest'])
518
+ digest(report[field], 'report_digest_chain_invalid');
519
+ if (report.decision !== 'BLOCKED' || report.promotable !== false)
520
+ fail('report_must_be_blocked');
521
+ if (!Array.isArray(report.blockers))
522
+ fail('report_blockers_invalid');
523
+ const blockers = report.blockers;
524
+ if (!blockers.every((blocker) => typeof blocker === 'string') || !blockers.includes('human_writer_evidence_deferred'))
525
+ fail('report_blockers_invalid');
526
+ const intentToTreat = record(report.intentToTreat, 'report_intent_to_treat_invalid');
527
+ exactKeys(intentToTreat, ['denominator', 'completed', 'hardFailures', 'timeouts', 'abandonments', 'runsDigest', 'expectedRatings', 'observedRatings', 'missingRatings', 'reconciled']);
528
+ for (const field of ['denominator', 'completed', 'hardFailures', 'timeouts', 'abandonments', 'expectedRatings', 'observedRatings', 'missingRatings'])
529
+ count(intentToTreat[field], 'report_intent_to_treat_invalid');
530
+ digest(intentToTreat.runsDigest, 'report_intent_to_treat_invalid');
531
+ if (typeof intentToTreat.reconciled !== 'boolean')
532
+ fail('report_intent_to_treat_invalid');
533
+ const metrics = record(report.metrics, 'report_metrics_invalid');
534
+ exactKeys(metrics, ['preference', 'correctionVersusConfirm', 'completion', 'abandonment', 'providerRuns']);
535
+ const rateMetric = (value, fields) => {
536
+ const metric = record(value, 'report_metrics_invalid');
537
+ exactKeys(metric, fields);
538
+ for (const field of fields) {
539
+ if (field === 'uncertainty95') {
540
+ if (!Array.isArray(metric[field]) || metric[field].length !== 2)
541
+ fail('report_metrics_invalid');
542
+ for (const bound of metric[field])
543
+ finiteNumber(bound, 0, 1, 'report_metrics_invalid');
544
+ }
545
+ else if (field === 'rate' || field === 'correctionRate')
546
+ finiteNumber(metric[field], 0, 1, 'report_metrics_invalid');
547
+ else
548
+ count(metric[field], 'report_metrics_invalid');
549
+ }
550
+ };
551
+ rateMetric(metrics.preference, ['numerator', 'denominator', 'rate', 'uncertainty95']);
552
+ rateMetric(metrics.completion, ['numerator', 'denominator', 'rate', 'uncertainty95']);
553
+ rateMetric(metrics.abandonment, ['numerator', 'denominator', 'rate', 'uncertainty95']);
554
+ const correction = record(metrics.correctionVersusConfirm, 'report_metrics_invalid');
555
+ exactKeys(correction, ['stage1', 'baseline']);
556
+ rateMetric(correction.stage1, ['corrections', 'confirms', 'denominator', 'correctionRate', 'uncertainty95']);
557
+ rateMetric(correction.baseline, ['corrections', 'confirms', 'denominator', 'correctionRate', 'uncertainty95']);
558
+ const providerRuns = record(metrics.providerRuns, 'report_metrics_invalid');
559
+ exactKeys(providerRuns, ['completed', 'abandoned', 'hardFailures', 'timeouts', 'denominator']);
560
+ for (const field of ['completed', 'abandoned', 'hardFailures', 'timeouts', 'denominator'])
561
+ count(providerRuns[field], 'report_metrics_invalid');
562
+ const reportDigest = digest(report.reportDigest, 'report_digest_invalid');
563
+ const { reportDigest: _reportDigest, ...reportCore } = report;
564
+ if (reportDigest !== sha256Canonical(reportCore))
565
+ fail('report_digest_mismatch');
566
+ const protocolDigest = digest(report.protocolDigest, 'report_digest_chain_invalid');
567
+ const packetDigest = digest(report.packetDigest, 'report_digest_chain_invalid');
568
+ const releaseAuditDigest = digest(report.releaseAuditDigest, 'report_digest_chain_invalid');
569
+ if (!['PROCEED_TO_MAR_363', 'STOP', 'REPEAT_PROTOCOL'].includes(String(disposition)))
570
+ fail('checkpoint_disposition_invalid');
571
+ if (disposition === 'PROCEED_TO_MAR_363')
572
+ fail('human_evidence_deferred');
573
+ const base = { kind: 'hyv-stage1-checkpoint-disposition', version: '1', protocolDigest, packetDigest, releaseAuditDigest, reportDigest, disposition };
574
+ validateAttestation(attestation, sha256Canonical(base), protocolDigest, packetDigest);
575
+ if (attestation.purpose !== 'stage1-checkpoint-disposition')
576
+ fail('attestation_purpose_mismatch');
577
+ const artifact = { ...base, attestation };
578
+ return { ...artifact, dispositionDigest: sha256Canonical(artifact) };
579
+ }