File size: 22,230 Bytes
518343a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
// doctrine-scanner-exempt: legacy live-product surface; rename tracked as separate engineering debt — see scripts/check-doctrine-v6.mjs header.
// MYTHOS LAYER — A11oy × Sentra defensive orchestration architecture.
//
// Distillation of the Mythos-class executive summary into a static doctrine
// surface: the A11oy command-agent orchestration loop, the integrated scanner
// toolchain catalogue, the RL pipeline formulation, the guardrails stack, and
// the phased roadmap.
//
// All inputs are public research, official tool documentation, or first-principle
// reconstructions. No leaks, no scraped weights, no telemetry. Adoption requires
// a Sentra approval workflow per the Glasswing doctrine.

export const MYTHOS_LAYER_VERSION = '0.1.0-seed';

export const MYTHOS_LAYER_TAGLINE =
  'A11oy plans, Sentra executes, the human approves. Defensive only, evidence-bound, public-input only.';

// ---------------------------------------------------------------------------
// 1. ORCHESTRATION LOOP
// The 10-step A11oy → Sentra interaction described in the architecture diagram.
// ---------------------------------------------------------------------------

export type LoopActor = 'A11oy' | 'Sentra' | 'Human' | 'External';

export interface OrchestrationStep {
  step: number;
  from: LoopActor;
  to: LoopActor;
  action: string;
  detail: string;
  guardrail: string;
}

export const ORCHESTRATION_STEPS: readonly OrchestrationStep[] = [
  {
    step: 1,
    from: 'A11oy',
    to: 'Sentra',
    action: 'plan tasks',
    detail:
      'A11oy decomposes a natural-language or policy directive (e.g. "scan production Linux fleet for critical CVEs") into a typed task graph and submits it to Sentra\u2019s Policy Engine for admission control.',
    guardrail:
      'Plan is rejected if it requests any action outside the tenant\u2019s registered scope or violates a Constitution rule.',
  },
  {
    step: 2,
    from: 'A11oy',
    to: 'Sentra',
    action: 'execute scans',
    detail:
      'Approved tasks dispatch the scanner toolchain (CodeQL, Semgrep, OSV-Scanner, Trivy, Gitleaks, Checkov) against the target surface defined in the plan.',
    guardrail:
      'Scanners run in least-privilege containers with read-only mounts; no scanner can write to a target system.',
  },
  {
    step: 3,
    from: 'Sentra',
    to: 'Sentra',
    action: 'findings → analyzer',
    detail:
      'Raw scanner output is normalised into the Mythos finding schema and fed to the Risk Assessment Engine for de-duplication, severity scoring, and exploitability triage.',
    guardrail:
      'Every finding is hash-linked to the scanner version, ruleset version, and target snapshot for chain-of-custody.',
  },
  {
    step: 4,
    from: 'Sentra',
    to: 'A11oy',
    action: 'context info → model',
    detail:
      'Risk-scored findings, the relevant code/IaC context, and the dependency graph are returned to the A11oy LLM agent for patch reasoning.',
    guardrail:
      'Context is bounded by the original plan\u2019s scope; no out-of-scope source is delivered to the model.',
  },
  {
    step: 5,
    from: 'A11oy',
    to: 'Sentra',
    action: 'propose patch',
    detail:
      'A11oy generates a candidate patch (diff) and a rationale trace, then submits both to Sentra\u2019s Patch Engine for static validation.',
    guardrail:
      'Patch payload is schema-validated; any non-diff side-channel (shell calls, network IO) is stripped before submission.',
  },
  {
    step: 6,
    from: 'Sentra',
    to: 'Human',
    action: 'patch suggestion → approval',
    detail:
      'The Patch Engine attaches a static-analysis verdict and posts the suggestion to the Approval Queue with severity, blast radius, and rollback plan.',
    guardrail:
      'Production patches require dual-key approval; staging patches require single-key approval; no auto-merge above low risk.',
  },
  {
    step: 7,
    from: 'Human',
    to: 'Sentra',
    action: 'commit approved patch',
    detail:
      'On approve, Sentra commits the patch to the staging branch with the approver\u2019s OIDC-attested signature recorded in the audit ledger.',
    guardrail:
      'Approval evidence is appended to the Cerberus vault before the commit lands; rollback is a single-click revert.',
  },
  {
    step: 8,
    from: 'Sentra',
    to: 'External',
    action: 'apply patch',
    detail:
      'Sentra deploys the approved patch through the customer\u2019s existing CI/CD path and verifies the targeted finding is resolved on the post-deploy scan.',
    guardrail:
      'Apply step is idempotent and auto-reverts if the post-deploy scan introduces new criticals.',
  },
  {
    step: 9,
    from: 'A11oy',
    to: 'A11oy',
    action: 'retrain',
    detail:
      'Outcome (verified fix, false positive, regression) is logged as an RL episode for the Silver planner; reward signal is composed from finding-closed, no-new-bugs, and time-to-resolve.',
    guardrail:
      'Training data is tenant-isolated; no cross-tenant gradient sharing without explicit opt-in.',
  },
  {
    step: 10,
    from: 'Human',
    to: 'Sentra',
    action: 'review',
    detail:
      'Analyst reviews the closed loop, marks the disposition (correct / over-cautious / wrong), and the feedback updates the reward model used by step 9.',
    guardrail:
      'Reviewer disposition is itself logged as evidence and feeds the Mirror Eval slice for the next quarterly model audit.',
  },
];

// ---------------------------------------------------------------------------
// 2. SCANNER TOOLCHAIN
// Public, defensively-licensed tools the orchestration layer composes.
// ---------------------------------------------------------------------------

export type ScannerSurface =
  | 'source-code'
  | 'dependencies'
  | 'containers'
  | 'iac'
  | 'secrets';

export type ScannerLicense = 'MIT' | 'Apache-2.0' | 'MPL-2.0' | 'LGPL-3.0' | 'GPL-3.0';

export type ScannerId =
  | 'codeql'
  | 'semgrep'
  | 'osv-scanner'
  | 'trivy'
  | 'gitleaks'
  | 'checkov';

export interface Scanner {
  id: ScannerId;
  name: string;
  surface: ScannerSurface;
  oneLine: string;
  primitive: string;
  license: ScannerLicense;
  source: { repo: string; url: string; org: string; lang: string };
  sentraBinding: string;
  guardrail: string;
}

export const SCANNERS: readonly Scanner[] = [
  {
    id: 'codeql',
    name: 'CodeQL',
    surface: 'source-code',
    oneLine:
      'GitHub\u2019s semantic code-analysis engine. Treats the codebase as a database and runs typed queries to find vulnerability classes (SQLi, RCE, taint flows).',
    primitive: 'Datalog-style queries over a code property graph, with curated query packs per language.',
    license: 'MIT',
    source: { repo: 'github/codeql', url: 'https://github.com/github/codeql', org: 'GitHub', lang: 'C++ / QL' },
    sentraBinding:
      'Invoked by the Static Scanner module on every plan that touches source. Findings normalise into the Mythos finding schema with rule-id provenance.',
    guardrail:
      'Runs in a sandboxed container against a read-only checkout. The query pack version is pinned and recorded in the evidence ledger.',
  },
  {
    id: 'semgrep',
    name: 'Semgrep',
    surface: 'source-code',
    oneLine:
      'Fast, syntax-aware pattern matcher for source code. Lower-cost first pass that catches the long-tail of policy-driven anti-patterns.',
    primitive:
      'AST pattern templates with metavariables and taint analysis, executed against many languages from a single ruleset.',
    license: 'LGPL-3.0',
    source: { repo: 'semgrep/semgrep', url: 'https://github.com/semgrep/semgrep', org: 'Semgrep', lang: 'Python / OCaml' },
    sentraBinding:
      'Runs in parallel with CodeQL as the cheap first sieve. Its findings carry lower default severity unless escalated by the Risk Engine.',
    guardrail:
      'Custom rules require Constitution review before they are added to the active ruleset; no inline rule loading.',
  },
  {
    id: 'osv-scanner',
    name: 'OSV-Scanner',
    surface: 'dependencies',
    oneLine:
      'Google\u2019s frontend to the Open Source Vulnerabilities database. Resolves project lockfiles and reports known CVEs against pinned versions.',
    primitive:
      'Deterministic lockfile parser + queries against the public OSV.dev advisory feed.',
    license: 'Apache-2.0',
    source: { repo: 'google/osv-scanner', url: 'https://github.com/google/osv-scanner', org: 'Google', lang: 'Go' },
    sentraBinding:
      'Runs on every dependency manifest in scope. Findings link to the upstream OSV entry plus the targeted lockfile path.',
    guardrail:
      'Uses the public OSV API only. No customer SBOM is uploaded; queries are purl-based and opaque.',
  },
  {
    id: 'trivy',
    name: 'Trivy',
    surface: 'containers',
    oneLine:
      'Container, OS package, and IaC scanner. The default Sentra layer for any image or registry surface.',
    primitive:
      'Layered scanner that combines distro vulnerability databases, SBOM extraction, and a Rego-driven misconfig pass.',
    license: 'Apache-2.0',
    source: { repo: 'aquasecurity/trivy', url: 'https://github.com/aquasecurity/trivy', org: 'Aqua Security', lang: 'Go' },
    sentraBinding:
      'Wired to both the Static Scanner (image scan) and the Patch Engine (post-deploy verification scan).',
    guardrail:
      'Operates against image digests, not running workloads. No exec-into-container path is exposed.',
  },
  {
    id: 'gitleaks',
    name: 'Gitleaks',
    surface: 'secrets',
    oneLine:
      'High-entropy and pattern-based secrets scanner for git history. Catches leaked credentials before they reach production.',
    primitive:
      'Configurable regex + entropy ruleset over git diffs and historical blobs.',
    license: 'MIT',
    source: { repo: 'gitleaks/gitleaks', url: 'https://github.com/gitleaks/gitleaks', org: 'Zachary Rice', lang: 'Go' },
    sentraBinding:
      'Triggered on every plan that includes a repository scan. Findings are routed straight to the Approval Queue with auto-redaction in the audit log.',
    guardrail:
      'Findings are stored as hashed fingerprints, not the raw secret. The raw match is held in the Cerberus vault under analyst-only access.',
  },
  {
    id: 'checkov',
    name: 'Checkov',
    surface: 'iac',
    oneLine:
      'Infrastructure-as-code scanner for Terraform, CloudFormation, Kubernetes, Helm, and ARM. Catches misconfigurations before they ship.',
    primitive:
      'Built-in policy library + custom policies expressed in Python or YAML, executed against parsed IaC graphs.',
    license: 'Apache-2.0',
    source: { repo: 'bridgecrewio/checkov', url: 'https://github.com/bridgecrewio/checkov', org: 'Bridgecrew / Prisma Cloud', lang: 'Python' },
    sentraBinding:
      'Runs on every IaC change in a plan. Verdicts feed the Compliance Engine for SOC2 / CIS-Benchmark mapping.',
    guardrail:
      'No --download-external-modules in CI; only vendored modules are evaluated to keep the supply chain attested.',
  },
];

// ---------------------------------------------------------------------------
// 3. RL PIPELINE
// Public-research formulation: state, action, reward, training regime.
// ---------------------------------------------------------------------------

export interface RLAxis {
  id: 'state' | 'action' | 'reward' | 'regime' | 'humanFeedback' | 'safety';
  axis: string;
  detail: string;
  citation: string;
}

export const RL_PIPELINE: readonly RLAxis[] = [
  {
    id: 'state',
    axis: 'State',
    detail:
      'Current security posture: vulnerable code context, dependency graph, recent scanner output, telemetry stream, and the agent\u2019s long-term memory of prior episodes.',
    citation:
      'Anthropic public research on agentic context windows; Team Atlanta patch RL state design.',
  },
  {
    id: 'action',
    axis: 'Action',
    detail:
      'Discrete, schema-validated operations: scan repo X with tool Y, fetch code piece Z, generate patch for CVE C, apply staging deploy, request human approval. No free-form shell.',
    citation:
      'GRPO action-space framing; Anthropic public RL guardrail constraints.',
  },
  {
    id: 'reward',
    axis: 'Reward',
    detail:
      'Composite: +1 for verified vulnerability closed, +0.5 for high-severity CVE detected, \u22120.5 for false positive, \u22121 for regression introduced, \u22122 for any policy violation attempt.',
    citation:
      'Team Atlanta patch RL: reward=1 when patch compiles and fixes the bug. Sentra extends with regression and policy penalties.',
  },
  {
    id: 'regime',
    axis: 'Training regime',
    detail:
      'Offline replay of historical incidents and CTF corpora bootstraps the policy. Online self-play in isolated lab networks refines it. PPO and GRPO are the primary on-policy algorithms.',
    citation:
      'PPO (Schulman 2017); GRPO (DeepSeek-Math 2024); Team Atlanta agentic patch loop.',
  },
  {
    id: 'humanFeedback',
    axis: 'Human feedback (RLHF)',
    detail:
      'Every high-impact decision (production patch, isolation order, exploit confirmation) routes through an analyst. The accept/reject/revise signal updates the reward model used in the next training run.',
    citation:
      'Anthropic RLHF technique reports; NIST AI RMF guidance on human oversight.',
  },
  {
    id: 'safety',
    axis: 'Safety constraints',
    detail:
      'Hard rules sit outside the RL loop: schema-validated outputs, sandboxed action execution, no patch to prod without dual-key approval, no scanner against out-of-scope assets.',
    citation:
      'Public Anthropic agentic guardrail patterns; OWASP LLM Top-10 input/output validation.',
  },
];

// ---------------------------------------------------------------------------
// 4. GUARDRAILS STACK
// Multi-layer defensive constraints applied to every model action.
// ---------------------------------------------------------------------------

export interface Guardrail {
  layer: string;
  control: string;
  enforcedBy: string;
}

export const GUARDRAILS: readonly Guardrail[] = [
  {
    layer: 'Input',
    control: 'Schema validation on every prompt and tool argument. Reject on parse failure, no silent coercion.',
    enforcedBy: 'A11oy intent router + Mythos doctrine schemas (JSON Schema 2020-12).',
  },
  {
    layer: 'Action',
    control: 'Action space restricted to a typed registry. No free-form shell, no dynamic code eval, no network egress outside the tenant scope.',
    enforcedBy: 'Sentra capability compartments + connector firewall.',
  },
  {
    layer: 'Output',
    control: 'Patch diffs static-validated; analysis text PII-redacted; secrets hashed before storage.',
    enforcedBy: 'Patch Engine validators + Cerberus vault redaction policy.',
  },
  {
    layer: 'Approval',
    control: 'Risk-tiered gating: low \u2192 single-key, medium \u2192 dual-key, high \u2192 dual-key plus 24h soak in staging.',
    enforcedBy: 'Sentra Approval Queue + Constitution rule engine.',
  },
  {
    layer: 'Audit',
    control: 'Every action, approval, and rejection appended to a hash-linked evidence ledger with OIDC-attested signatures.',
    enforcedBy: 'Cerberus evidence vault (append-only, signed).',
  },
  {
    layer: 'Provenance',
    control: 'Model version, prompt hash, tool versions, and ruleset versions captured per episode for reproducibility and SLSA-style supply-chain attestation.',
    enforcedBy: 'Hephaestus provenance gate.',
  },
];

// ---------------------------------------------------------------------------
// 5. PHASED ROADMAP
// Translation of the research brief\u2019s 0\u201324 month plan onto Sentra modules.
// ---------------------------------------------------------------------------

export type RoadmapPhase = '0\u20136' | '7\u201312' | '13\u201324';

export interface RoadmapMilestone {
  id: string;
  phase: RoadmapPhase;
  title: string;
  detail: string;
  module: string;
}

export const ROADMAP: readonly RoadmapMilestone[] = [
  {
    id: 'phase1-toolchain',
    phase: '0\u20136',
    title: 'Scanner toolchain wired end-to-end',
    detail:
      'CodeQL + Semgrep + OSV-Scanner + Trivy + Gitleaks + Checkov each producing normalised findings into the Mythos schema, with a deterministic merge pass and a single Risk Engine entry point.',
    module: 'Sentra Static Scanner + Risk Engine',
  },
  {
    id: 'phase1-orchestrator',
    phase: '0\u20136',
    title: 'A11oy plan \u2192 Sentra dispatch path',
    detail:
      'A11oy decomposes natural-language directives into typed plans; Sentra admits or rejects against the Constitution; approved plans dispatch to the scanner toolchain.',
    module: 'A11oy Planner + Sentra Policy Engine',
  },
  {
    id: 'phase1-approval-loop',
    phase: '0\u20136',
    title: 'Approval Queue with dual-key gating',
    detail:
      'Patch suggestions appear in the Approval Queue with severity, blast radius, and rollback plan. Production changes require two operators; staging requires one.',
    module: 'Sentra Approval Queue',
  },
  {
    id: 'phase2-rl-offline',
    phase: '7\u201312',
    title: 'Offline RL bootstrap on historical incidents',
    detail:
      'Replay anonymised incident logs as RL episodes. Train Silver to prefer plans that historically closed the finding without regression. Reward model anchored on analyst dispositions.',
    module: 'Silver RL Planner',
  },
  {
    id: 'phase2-context-retrieval',
    phase: '7\u201312',
    title: 'RL contextual code retrieval',
    detail:
      'Per Team Atlanta, train a retrieval policy that fetches the exact code pieces required to verify or fix a finding. Cuts token spend and false-positive rate.',
    module: 'A11oy Retriever + Sentra Patch Engine',
  },
  {
    id: 'phase2-redteam',
    phase: '7\u201312',
    title: 'Continuous defensive red-team in sandbox',
    detail:
      'Strict-policy hacker agent runs in an isolated tenant clone, attempting attacks against staging snapshots. Outcomes feed the Risk Engine\u2019s exploitability score.',
    module: 'Sentra Sandbox + Adversarial Resilience',
  },
  {
    id: 'phase3-self-play',
    phase: '13\u201324',
    title: 'Online self-play in isolated lab networks',
    detail:
      'Synthetic corpora and CTF-style environments host longer-horizon self-play episodes. PPO / GRPO policy updates ship behind a Mirror Eval gate.',
    module: 'Silver RL Planner + Mirror Eval',
  },
  {
    id: 'phase3-compliance',
    phase: '13\u201324',
    title: 'SOC2 Type II + ISO 42001 attestation',
    detail:
      'The full A11oy \u00d7 Sentra control set passes a Type II audit with the evidence ledger as the primary control surface. ISO 42001 added for AI-management-system coverage.',
    module: 'Sentra Compliance Engine + Cerberus Evidence Vault',
  },
  {
    id: 'phase3-defender-network',
    phase: '13\u201324',
    title: 'Defender Credits federated learning',
    detail:
      'Tenants opt in to a federated update channel where reward-model deltas (never raw evidence) cross-pollinate. Each contribution earns Defender Credits redeemable against compute.',
    module: 'A11oy Federation + Defender Credits',
  },
];

// ---------------------------------------------------------------------------
// 6. PUBLIC MODEL REFERENCES
// Candidates from Table 1 of the brief. Public access only \u2014 no claim of
// integration with proprietary endpoints we do not have.
// ---------------------------------------------------------------------------

export interface ModelReference {
  id: string;
  family: string;
  contextWindow: string;
  strengths: string;
  access: string;
}

export const MODEL_REFERENCES: readonly ModelReference[] = [
  {
    id: 'claude-sonnet-4-6',
    family: 'Anthropic Claude Sonnet 4.6',
    contextWindow: '~1M tokens',
    strengths: 'Frontier coding and agentic tool use; long-running multi-file workflows.',
    access: 'Public Anthropic API. Used as the default A11oy planner where customer policy permits.',
  },
  {
    id: 'claude-opus-4-7',
    family: 'Anthropic Claude Opus 4.7',
    contextWindow: '~512K\u20131M tokens',
    strengths: 'Deepest reasoning chain; first Mythos-class safety filters in the public Claude line.',
    access: 'Public Anthropic API. Reserved for high-risk patch reasoning routed via A11oy escalation.',
  },
  {
    id: 'openmythos-rdt',
    family: 'OpenMythos (community RDT reconstruction)',
    contextWindow: '~1M tokens (target)',
    strengths: 'Independent looped-transformer reconstruction. Useful as a transparent reasoning baseline.',
    access: 'Open-source weights, public training recipe. No leaked Anthropic artifacts ever ingested.',
  },
  {
    id: 'gpt-frontier',
    family: 'OpenAI frontier (GPT-4o / GPT-5 line)',
    contextWindow: 'Vendor-disclosed',
    strengths: 'Multimodal agent capabilities; strong tool-call reliability.',
    access: 'Public OpenAI API. Selectable via the Model Router when customer policy prefers it.',
  },
];

// ---------------------------------------------------------------------------
// 7. CITATIONS — anchor refs back to the source brief.
// ---------------------------------------------------------------------------

export const MYTHOS_LAYER_CITATIONS: ReadonlyArray<{ tag: string; source: string }> = [
  { tag: '66', source: 'Anthropic public posting on Mythos Preview vulnerability discovery and exploitation.' },
  { tag: '70', source: 'Project Glasswing public announcement and partner stack.' },
  { tag: '72', source: 'Claude Sonnet 4.6 public release notes (1M-token context, agentic coding).' },
  { tag: '73', source: 'Claude Opus 4.7 public release notes (Mythos-safe filtering).' },
  { tag: '74', source: 'GRPO algorithm description, DeepSeek-Math 2024.' },
  { tag: '75', source: 'Team Atlanta agentic patch RL writeup; reward shaping for compile/fix.' },
  { tag: '86', source: 'UK AI Security Institute (AISI) Mythos threat-level assessment.' },
  { tag: '91', source: 'OpenMythos community reconstruction notes on recurrent-depth transformers.' },
  { tag: '95', source: 'Public agentic guardrail patterns (schema validation, sandboxed actions, output constraints).' },
];