2sagaco48 1.0.0

package/src/TEMPLATE_PIPELINE_AUDIT.md

@@ -0,0 +1,208 @@
+# V15 Calibration + Research Template Audit
+
+Owner: Codex  
+Scope: all 8 OMEGA templates, 40 style overrides, calibration defaults, V15 loops, production SLOOP/N-Deep routing  
+Evidence basis: source-level audit and successful production builds; no live paid-API quality run was available in this environment.
+
+## Executive Findings
+
+### Confirmed defects fixed in this pass
+
+1. **V15 used a template name, not the template.** The prompt said “follow OMEGA-STRATEGY exactly” but never supplied section names or section contracts. Models could not comply reliably. Fixed with `buildAdaptiveTemplateContract()` injected into V15.
+2. **Production SLOOP mapped OMEGA templates to lossy generic archetypes.** Discovery became `market-commercial`; Diligence became `audit-assurance`; Crisis fell through to `decision-strategy`. Fixed by a persistent OMEGA-aware `sloop-runner.ts` that consumes actual `OMEGA_TEMPLATES` sections.
+3. **SLOOP dropped later required sections.** `targetSections = sections.slice(0, pages + 2)` meant a 4-page report silently omitted sections near the end (often Risk, Recommendations, Appendix). Fixed: every contracted section is retained; low page counts compress per-section depth instead of deleting sections.
+4. **Template requirements encouraged fabrication.** Many hints request NPS, interviews, EBITDA, IRR, sample sizes, audit responses, classifications, results, or official approvals without requiring evidence. Fixed with template-specific non-fabrication/status rules (`[PROPOSED]`, `[DATA GAP]`, `[ASSUMPTION]`).
+5. **Style overrides could contradict template semantics.** Default `OMEGA-STRATEGY + --bain-pe` combines a strategy skeleton with a diligence style preset. Fixed: incompatible overrides may modulate voice/layout only; they may not import the other template's sections, metrics, or evidence claims.
+6. **N-Deep rewrote entire documents and could produce duplicate unchanged passes.** Fixed with localized section patches, byte-for-byte preservation of untouched sections, strict critical→major→warning priority, acceptance only on issue-vector or score improvement, and early stop when no progress is possible.
+7. **V15 default-state mismatch.** UI messaging/default profile expected enabled, while `v15-state` defaulted false and was not intercepted. Fixed: persistent state defaults ON unless explicitly disabled.
+
+## Cross-Cutting Template Weaknesses
+
+### Evidence and epistemic status
+
+- Template hints frequently mix required structure with implied facts. “Customer interview synthesis (n=)” is a valid diligence section contract but an invalid instruction when no interviews exist.
+- Required numbers must be computed or sourced, not generated because a template asks for them.
+- References sections must list only used, retrieved sources. A template cannot justify synthetic citations.
+- “Not applicable” should be rare and specific; a missing input is a `[DATA GAP]`, not automatically “N/A.”
+
+### Page-budget contradictions
+
+- OMEGA templates often describe 12–40+ pages, while calibration defaults to SLOOP 4 pages.
+- Previous behavior tried to satisfy both by truncating or dropping late sections.
+- Correct behavior is **contract-preserving compression**: retain all main headings, allocate a smaller word budget per section, keep appendix compact.
+- Second-order effect: very small per-section budgets can create shallow coverage. The UI should present page count as compression, not as permission to omit sections.
+
+### Template/task-type mismatch
+
+- OMEGA-SCIENCE must classify empirical study vs protocol/proposal vs systematic review vs narrative review before choosing Results/Expected Outcomes/PRISMA structure.
+- OMEGA-CRISIS combines restructuring and federal RFP semantics; a query may need one branch, not both.
+- OMEGA-COMPLIANCE combines audit, governance, ESG, and financial-services risk. The selected framework must be evidence-backed and jurisdiction-specific.
+- Third-order effect: applying all template hints literally creates internally inconsistent reports that look comprehensive but are professionally invalid.
+
+### Style override mismatch
+
+- 40 style overrides map to specific templates, but the UI permits any combination.
+- Layout/voice characteristics can transfer; substantive evidence requirements usually cannot.
+- Example: `--bain-pe` may contribute concise investment-committee tone to OMEGA-STRATEGY, but must not force NPS, sponsor returns, or diligence interviews.
+
+## Template-by-Template Audit
+
+### OMEGA-STRATEGY
+
+Strengths: decision-first BLUF, options tournament, quantified recommendation, implementation and risk.  
+Risks:
+- TAM/SAM/SOM, NPV/IRR, value bridge and utility scores invite invented numbers.
+- “Top three MECE findings” can force artificial exclusivity.
+- RACI/deadlines can imply authority the model does not have.
+- Fixed four-option tournament may be inappropriate for binary or constrained decisions.
+Fix applied: evidence-gated metrics/formulas; all sections retained under compression; style conflict rules.  
+Second-order: a quantitatively empty but polished strategy report can falsely signal rigor.  
+Third-order: downstream decisions may treat model-generated assumptions as approved targets unless status labels remain visible.
+
+### OMEGA-DILIGENCE
+
+Strengths: scope/reliance, thesis, commercial/financial diligence, value plan, risk.  
+Risks:
+- Highest fabrication pressure of all templates: interviews, NPS, retention, EBITDA adjustments, debt, comparables, IRR/MOIC.
+- “PROCEED/PASS” can overstate confidence with incomplete diligence.
+- QofE language can resemble assurance or agreed-upon procedures.
+Fix applied: explicit fact/inference/diligence-request separation and conditional verdict rule.  
+Second-order: invented diligence evidence compounds into valuation and return outputs.  
+Third-order: a false EBITDA adjustment can propagate through leverage, exit value, MOIC, IRR, and recommendation.
+
+### OMEGA-DISCOVERY
+
+Strengths: methodology, thesis, thematic synthesis, horizon scan, stakeholders.  
+Risks:
+- Survey-heavy hooks encourage synthetic sample size, field dates, weighting and regional cuts.
+- “Striking statistic” encourages false precision.
+- 3–7 chapters at 4–8 pages each conflicts sharply with 4-page calibration.
+- Prior SLOOP mapping to `market-commercial` omitted methodology and stakeholder logic.
+Fix applied: actual Discovery sections; desk-research labeling when no survey exists; compressed all-section contract.  
+Second-order: fake survey methodology legitimizes otherwise unsupported findings.  
+Third-order: fabricated cross-tabs can misdirect policy or regional strategy.
+
+### OMEGA-COMPLIANCE
+
+Strengths: criterion/evidence/findings, risk inventory, recommendations, response.  
+Risks:
+- Management response cannot be synthesized “verbatim.”
+- GAGAS/COSO/CSRD applicability is jurisdiction and engagement dependent.
+- Residual-risk scores, owners and control effectiveness require assessment evidence.
+- Transmittal/signature/independence language can imply an official audit engagement.
+Fix applied: no invented responses, owners, audit sufficiency or compliance claims; observations separated from criteria/inference.  
+Second-order: false “compliant” findings reduce remediation urgency.  
+Third-order: official-looking formatting may be mistaken for legal assurance.
+
+### OMEGA-BUILD
+
+Strengths: current/target architecture, opportunity portfolio, waves, business case, operating model and change.  
+Risks:
+- Current-state inventory and maturity scores require discovery data.
+- Vendor/platform recommendations can become biased by style hooks.
+- TCO/NPV/payback and architecture claims invite unsupported precision.
+- Security frameworks may be cited outside their applicability.
+Fix applied: current-state data gaps vs proposed target state; no invented certification/cost/benefit.  
+Second-order: fictional current-state assumptions contaminate target architecture.  
+Third-order: a fabricated business case can lock in procurement and operating-model decisions.
+
+### OMEGA-SCIENCE
+
+Strengths: rigorous academic structure, methodology, limitations and reproducibility.  
+Risks:
+- A proposal forced into IMRAD produces fabricated Results.
+- PRISMA requires actual searches, counts, protocol and screening—not prose simulation.
+- Methods hints encourage invented sample/effect/p-values/CI/ethics approval.
+- Front matter encourages invented authors, ORCID, funding and COI.
+Fix applied: task-type classification and explicit prohibition on invented results, approvals, registration, numbers, authorship and DOI.  
+Second-order: invented Results make the Discussion and Conclusions appear empirically grounded.  
+Third-order: fake ethics/registration statements create publication-integrity and compliance exposure.
+
+### NIH-GRANT-SRF
+
+Strengths: aligned Factor 1/Factor 2 structure, aims, approach, human subjects, DMS.  
+Risks:
+- Exactly three aims is not universal and can create over-scoped projects.
+- Preliminary data, IC fit, facilities and investigator capability must not be invented.
+- Power/sample/ICC and recruitment assumptions need deterministic calculation and inputs.
+- Human-subject and DMS language can imply institutional commitments.
+Fix applied: proposal-status tags, non-fabrication rules, adaptable aim count.  
+Second-order: fabricated feasibility evidence can mask an unfundable approach.  
+Third-order: invented institutional resources or approvals can create submission-integrity problems.
+
+### OMEGA-CRISIS
+
+Strengths: immediate forecast, options, execution, governance and risk.  
+Risks:
+- Combines restructuring and federal procurement branches that should not coexist by default.
+- Privilege and classification markings cannot be assigned by the model.
+- Liquidity, creditors, PIID, CPARS, clearances, rates and legal clauses require authoritative inputs.
+- Crisis recommendations have high consequence and short decision windows.
+Fix applied: explicit draft-only status; no legal privilege/classification or official representations; require counsel/CO review.  
+Second-order: invented liquidity or milestone assumptions change the selected option.  
+Third-order: false official markings or representations can create legal/security exposure.
+
+## Pipeline Loop Audit
+
+### Best-of-N
+
+- Fixed earlier token waste: hypotheses are compact outlines, winner only is expanded.
+- Remaining risk: outline-density scoring is structural, not semantic. It may favor an outline that mentions “assumptions/citations/caveats” over one with a stronger central idea.
+- Recommended future test: compare deterministic outline score against independent judge preference over 30 fixed prompts.
+
+### HDIG
+
+- Strength: targeted post-draft queries.
+- Weakness: adds retrieved evidence after the initial draft; unsupported initial framing can anchor subsequent queries.
+- Improvement applied indirectly: outline-first selection and template contract precede expansion; evidence rules constrain expansion.
+- Remaining risk: retrieval query quality and source diversity need empirical evaluation.
+
+### CoVe
+
+- Strength: isolated verification calls reduce cross-contamination.
+- Confirmed weakness: consistency is currently substring overlap, not entailment. Paraphrases and negation can be misclassified.
+- Recommended next fix: deterministic numeric/date/entity comparison first, then one batched entailment judge for unresolved pairs.
+
+### N-Deep
+
+- Fixed whole-document rewrites: only explicit section/paragraph patches are permitted.
+- Fixed duplicate flat passes: stop when no anchorable or improving patch exists.
+- Fixed priority: critical → major → warning.
+- Fixed citation contradiction and valid source range.
+- Remaining risk: defect-to-section localization is delegated to the model; malformed anchors cause early stop rather than repair.
+
+### Adversarial Review
+
+- Fixed cosmetic-only behavior: blocking defects now trigger repair before polish/judge.
+- Remaining risk: only one adversarial repair is attempted; new blocking defects introduced by repair are not red-teamed again.
+
+### Polish
+
+- Monotonic score guard prevents degradation.
+- Remaining risk: LLM polish can preserve a factual defect because it is explicitly content-conservative; this is correct behavior unless the defect was already surfaced upstream.
+
+### Judging
+
+- Fake 7.5 parse fallback removed; failures are excluded and surfaced.
+- Remaining risk: single-judge default has high variance; multiple judges improve reliability but consume RPM/RPD.
+
+## UI/Defaults/Routing Audit
+
+- One effective calibration page is rendered: package `V15Overlay` resolves through Vite to the local overlay, whose calibration import resolves to the generated local enhanced page. The package source remains present on disk but is not a duplicate rendered page.
+- Build-time patch script currently applies all expected anchors and fails loudly on a missing anchor.
+- Defaults: V15 enabled unless explicitly disabled; one batch question; 4-Stage on; N-Deep 3; Cluster 5; SLOOP 4 pages; OMEGA-STRATEGY; Sentinel; adversarial, grounding, SearXNG and defense pack on.
+- Known semantic tension: default `--bain-pe` maps to Diligence, while default template is Strategy. The new contract prevents substantive cross-import but the UI still presents this incompatible default combination for historical compatibility.
+
+## Turn N+8 Additions — Next-Turn Investigation Queue (updated)
+
+1. Build a deterministic template conformance test suite: all required headings once, no bare sections, no prohibited fabricated fields.
+2. Replace CoVe substring consistency with structured field comparison + batched entailment fallback.
+3. Run a testbed-gate false-positive benchmark on known-good reports; 85–115 hits per answer suggests several regexes are too broad.
+4. Re-red-team adversarial repairs once, with a hard two-pass cap, to detect newly introduced blockers.
+5. Calibrate outline-density scoring against independent preferences and add task/template coverage features.
+6. Evaluate changing the historical default style from `--bain-pe` to `--bain-strategy`; currently compatibility is preserved, not ideal semantic alignment.
+7. Harden `script.js` into a declarative patch manifest with expected patch count and post-generation assertions.
+
+## Verification Limit
+
+Production builds pass. Live LLM quality, external search uptime, and visual pixel parity could not be exercised through a browser/API harness in this environment. Claims above distinguish source-proven defects from empirical risks requiring live runs.

package/src/WILLIAMS_PERSONA_GUIDE.md

@@ -0,0 +1,62 @@
+# Williams Persona System — User Guide
+
+The Williams persona system shapes **how** the pipeline writes (voice, rhythm,
+structure) without changing **what** it says (facts, citations, safety are
+unchanged). Personas are drawn from Joseph M. Williams' *Style: Toward Clarity
+and Grace*. Each persona now ships a concrete directive (voice + do + avoid +
+cadence) so the difference between personas is large and audible.
+
+## How to use
+
+1. Open the V15 overlay → **Calibrate**.
+2. In the profile bar, pick a **Williams persona** from the dropdown.
+3. Run Live Compare or Batch Bank. The chosen persona is injected into the
+   drafting system prompt and every N-Deep editor pass.
+4. The persona **name** is shown in Run Settings; the persona is never named
+   inside the answer itself.
+
+Default persona: **The Sentinel** (surfaces assumptions, caveats, and scope
+before proceeding — the safest default for high-stakes calibration).
+
+## Persona map (voice → best use)
+
+| Persona | Voice | Best for |
+|---|---|---|
+| The Plain Dealer | Blunt, agent-driven | Direct answers, ops runbooks |
+| The Architect | Balanced, parallel | Structured comparisons, frameworks |
+| The Narrator | Story-like, old→new flow | Explanations, walkthroughs |
+| The Explainer | Signposted teacher | Tutorials, onboarding |
+| The Scholar | Formal, abstract | Academic / literature framing |
+| The Surgeon | Ruthless minimalist | Executive TL;DRs |
+| The Advocate | Climactic, persuasive | Recommendations, pitches |
+| The Diagnostician | Evidence-balancer | Risk / decision analysis |
+| The Conversationalist | Informal peer | Q&A, support |
+| The Essayist | Literary, periodic | Thought pieces |
+| The Weaver | Long, subordinated | Deep syntheses |
+| The Minimalist | Flat, uniform | Reference docs |
+| The Cartographer | Spatial mapper | Domain overviews |
+| The Dialectician | Thesis→antithesis→synthesis | Debates, trade-offs |
+| The Crystallographer | Faceted, dense | Briefs where each section stands alone |
+| The Counselor | Warm, transparent | Sensitive guidance |
+| The Polymath | Cross-domain analogies | Interdisciplinary insight |
+| The Sentinel | Watchful, caveat-first | High-stakes / compliance (default) |
+| The Oracle | Delayed revelation | Narrative build-ups |
+| The Alchemist | Jargon→vivid | Explaining complex tech |
+| The Strategist | Situation-mission-execution | Plans, decisions |
+| The Philosopher | First-principles | Conceptual reasoning |
+| The Provocateur | Conclusion-first | Challenging positions |
+
+## What each directive controls
+
+- **VOICE** — the persona's identity in one line.
+- **DO** — concrete imperatives applied to sentence construction.
+- **AVOID** — anti-patterns the persona must not produce.
+- **CADENCE** — the sentence-rhythm signature.
+
+## Guarantees
+
+- Persona changes are additive and **never** override factual accuracy,
+  citation grounding, truncation repair, or safety gates.
+- The persona is applied invisibly — it is never named inside the answer.
+- Selecting a different persona produces a measurably different voice because
+  each ships distinct DO/AVOID/CADENCE rules (not just a label).

package/src/chat/VeritasChat.tsx

@@ -0,0 +1,69 @@
+import { useMemo, useState } from "react";
+import { ContextSynthesisEngine } from "./synthesis";
+import type { OrchestrationResponse } from "./tier";
+
+export function VeritasChatSystem() {
+  const [query, setQuery] = useState("");
+  const [contextData, setContextData] = useState("");
+  const [executionPrompt, setExecutionPrompt] = useState("");
+  const [telemetry, setTelemetry] = useState<OrchestrationResponse | null>(null);
+  const [loading, setLoading] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+  const synthesisEngine = useMemo(() => new ContextSynthesisEngine(), []);
+
+  async function handleExecuteOrchestration(e: React.FormEvent) {
+    e.preventDefault();
+    if (!query || !contextData) {
+      setError("Both query and context data are required.");
+      return;
+    }
+    setLoading(true);
+    setError(null);
+    try {
+      const result = await synthesisEngine.synthesizeSecureContext(query, contextData);
+      setExecutionPrompt(result.finalizedPrompt);
+      setTelemetry(result.telemetry);
+    } catch (err: any) {
+      setError(err.message || "An orchestration error occurred.");
+    } finally {
+      setLoading(false);
+    }
+  }
+
+  return (
+    <div className="flex h-full w-full flex-col bg-slate-950 p-6 text-slate-100">
+      <header className="mb-6 border-b border-slate-800 pb-4">
+        <h1 className="font-mono text-xl font-bold tracking-wider text-cyan-400">PRISMAFETCH // MULTI-TIER CONTEXT ORCHESTRATOR v2.0</h1>
+        <p className="mt-1 text-xs text-slate-400">High-fidelity context protection against leakage, drift, and injection payloads.</p>
+      </header>
+      <div className="grid flex-1 grid-cols-1 gap-6 overflow-hidden lg:grid-cols-12">
+        <form onSubmit={handleExecuteOrchestration} className="flex flex-col space-y-4 lg:col-span-5">
+          <input value={query} onChange={(e) => setQuery(e.target.value)} className="rounded border border-slate-700 bg-slate-900 px-3 py-2 text-sm" placeholder="Extraction instruction..." />
+          <textarea value={contextData} onChange={(e) => setContextData(e.target.value)} className="h-64 flex-1 resize-none rounded border border-slate-700 bg-slate-900 p-3 font-mono text-xs" placeholder="Paste large source context..." />
+          <button disabled={loading} className="rounded bg-cyan-600 py-2.5 text-sm font-bold uppercase tracking-wide text-white disabled:bg-slate-800 disabled:text-slate-500">{loading ? "Processing..." : "Execute Orchestrated Ingestion"}</button>
+          {error && <div className="rounded border border-red-800 bg-red-950/50 p-3 font-mono text-xs text-red-400">[CRITICAL ERROR] {error}</div>}
+        </form>
+        <div className="flex flex-col space-y-4 overflow-y-auto lg:col-span-7">
+          {telemetry && (
+            <div className="space-y-3 rounded border border-slate-800 bg-slate-900 p-4">
+              <h3 className="font-mono text-xs font-bold uppercase tracking-wider text-emerald-400">// Telemetry</h3>
+              <div className="grid grid-cols-2 gap-3 text-xs sm:grid-cols-3">
+                <Metric label="Zero-width" value={telemetry.primary_safety.zero_width_cleared ? "PURGED" : "CLEAR"} />
+                <Metric label="Obfuscation" value={telemetry.primary_safety.obfuscation_neutralized ? "NEUTRALIZED" : "NONE"} />
+                <Metric label="Injection" value={telemetry.primary_safety.injection_heuristic_tripped ? "MUTED" : "CLEAR"} />
+                <Metric label="FP16 Outliers" value={telemetry.hardware_telemetry.allocated_fp16_outliers} />
+                <Metric label="INT4 Ambient" value={telemetry.hardware_telemetry.allocated_int4_ambient} />
+                <Metric label="VRAM Saved" value={`${telemetry.hardware_telemetry.vram_savings_est_mb} MB`} />
+              </div>
+            </div>
+          )}
+          <pre className="min-h-64 flex-1 overflow-auto whitespace-pre-wrap rounded border border-slate-800 bg-slate-900 p-3 font-mono text-xs text-slate-300">{executionPrompt || "No active verification context pipeline has been invoked yet."}</pre>
+        </div>
+      </div>
+    </div>
+  );
+}
+
+function Metric({ label, value }: { label: string; value: string | number }) {
+  return <div className="rounded border border-slate-800 bg-slate-950 p-2.5"><div className="text-[10px] uppercase text-slate-400">{label}</div><div className="mt-1 font-bold text-cyan-400">{value}</div></div>;
+}

package/src/chat/synthesis.ts

@@ -0,0 +1,31 @@
+import { VERITAS_SYSTEM_DIRECTIVES, type OrchestrationResponse } from "./tier";
+
+export class ContextSynthesisEngine {
+  private apiEndpoint: string;
+
+  constructor(endpoint = "/api/v1/context/orchestrate") {
+    this.apiEndpoint = endpoint;
+  }
+
+  public async synthesizeSecureContext(
+    userQuery: string,
+    rawLargeContext: string,
+    sessionId?: string,
+  ): Promise<{ finalizedPrompt: string; telemetry: OrchestrationResponse }> {
+    const response = await fetch(this.apiEndpoint, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ query: userQuery, large_context: rawLargeContext, session_id: sessionId }),
+    });
+    if (!response.ok) throw new Error(`Orchestrator returned bad server status: ${response.status}`);
+    const telemetryData = (await response.json()) as OrchestrationResponse;
+    let analyticalContext = telemetryData.compiled_evidence;
+    if (analyticalContext.includes("0x") || analyticalContext.length > 50_000) {
+      analyticalContext = `${VERITAS_SYSTEM_DIRECTIVES.ECHO_CHAMBER_PENALTY}\n\n${analyticalContext}`;
+    }
+    return {
+      finalizedPrompt: VERITAS_SYSTEM_DIRECTIVES.COV_WRAPPER(analyticalContext),
+      telemetry: telemetryData,
+    };
+  }
+}

package/src/chat/tier.ts

@@ -0,0 +1,45 @@
+export interface SafetyTelemetry {
+  zero_width_cleared: boolean;
+  obfuscation_neutralized: boolean;
+  injection_heuristic_tripped: boolean;
+  initial_bytes: number;
+  final_bytes: number;
+}
+
+export interface HardwareTelemetry {
+  allocated_fp16_outliers: number;
+  allocated_int4_ambient: number;
+  outlier_ratio: number;
+  vram_savings_est_mb: number;
+}
+
+export interface OrchestrationResponse {
+  success: boolean;
+  primary_safety: SafetyTelemetry;
+  evidence_blocks_count: number;
+  compiled_evidence: string;
+  hardware_telemetry: HardwareTelemetry;
+  map_efficiency_pct: number;
+  session_id: string;
+}
+
+export const VERITAS_SYSTEM_DIRECTIVES = {
+  COV_WRAPPER: (sanitizedContext: string): string => `<System_Directive>
+You are executing a High-Fidelity Extraction protocol under strict Sandbox Epistemology constraints.
+1. Scan the provided <Target_Context> block exclusively.
+2. Do not rely on internal parametric memory dimensions or historical patterns.
+3. Execute a complete Chain of Verification before outputting any final response.
+4. If internal knowledge contradicts <Source_Data>, <Source_Data> overrides internal weights.
+</System_Directive>
+
+<Target_Context>
+${sanitizedContext}
+</Target_Context>
+
+<Execution_Protocol>
+Step 1: Quote all sentences from <Target_Context> relevant to the user query verbatim.
+Step 2: Review the quotes. Do they directly and logically answer the query without assumption? (Yes/No)
+Step 3: If Yes, format the final response including strict citations. If No, state "Insufficient Data in Source."
+</Execution_Protocol>`,
+  ECHO_CHAMBER_PENALTY: "You are strictly forbidden from mirroring repetitive input syntax. Output analytical breakdowns in continuous, clear prose.",
+};

package/src/components/AdaptersPage.orig.tsx

@@ -0,0 +1,99 @@
+import { useEffect, useState } from "react";
+import { MODEL_LIMITS, snapshotAllUsage, type UsageSnapshot, VERITAS_FULL_LLM_ROSTER } from "@/lib/v15-rate-limiter";
+
+function PingDot({ snapshot }: { snapshot: UsageSnapshot }) {
+  const color = snapshot.throttled ? "bg-rose-500" : "bg-emerald-500";
+  const label = snapshot.throttled ? "THROTTLED" : "READY";
+  return (
+    <span className={`inline-flex items-center gap-1 rounded px-2 py-0.5 text-[10px] font-bold text-white ${color}`}>
+      <span className="h-2 w-2 animate-pulse rounded-full bg-white/80" /> {label}
+    </span>
+  );
+}
+
+export function AdaptersPage() {
+  return <AdaptersPageImpl />;
+}
+export default function AdaptersPageImpl() {
+  const [snapshots, setSnapshots] = useState<UsageSnapshot[]>(() => snapshotAllUsage());
+  const [now, setNow] = useState(Date.now());
+
+  useEffect(() => {
+    const t = setInterval(() => {
+      setSnapshots(snapshotAllUsage());
+      setNow(Date.now());
+    }, 1500);
+    return () => clearInterval(t);
+  }, []);
+
+  const totalCalls = snapshots.reduce((a, s) => a + s.rpmUsed + s.rpdUsed, 0);
+
+  return (
+    <div className="p-4">
+      <div className="mb-4 flex items-center justify-between">
+        <div>
+          <h2 className="text-lg font-bold text-zinc-900">Gemini & Gemma LLM Roster — Full Round-Robin Rotation</h2>
+          <p className="text-[11px] text-zinc-500">
+            All {VERITAS_FULL_LLM_ROSTER.length} models from your spec are registered in MODEL_LIMITS and honored by pickLeastLoaded / tryAcquire across the entire V15 pipeline and every scraper portfolio call. No exceptions. Last refresh {new Date(now).toLocaleTimeString()} — total calls tracked {totalCalls}.
+          </p>
+        </div>
+        <div className="rounded-lg bg-zinc-900 px-3 py-1.5 font-mono text-[11px] text-emerald-300">
+          RPM = requests/min · TPM = tokens/min · RPD = requests/day
+        </div>
+      </div>
+
+      <div className="overflow-auto rounded-xl border border-zinc-200 bg-white">
+        <table className="w-full border-collapse text-[12px]">
+          <thead className="bg-zinc-50 text-[10px] uppercase tracking-wider text-zinc-600">
+            <tr>
+              <th className="px-3 py-2 text-left">Model</th>
+              <th className="px-3 py-2 text-left">Category</th>
+              <th className="px-3 py-2 text-left">RPM Used / Max (remaining)</th>
+              <th className="px-3 py-2 text-left">RPD Used / Max (remaining)</th>
+              <th className="px-3 py-2 text-left">TPM Cap</th>
+              <th className="px-3 py-2 text-left">Inflight</th>
+              <th className="px-3 py-2 text-left">Status</th>
+              <th className="px-3 py-2 text-left">Next slot</th>
+            </tr>
+          </thead>
+          <tbody>
+            {snapshots.map((s) => {
+              const limit = MODEL_LIMITS[s.model];
+              return (
+                <tr key={s.model} className="border-t border-zinc-100 hover:bg-zinc-50/60">
+                  <td className="px-3 py-2 font-mono font-bold text-zinc-800">{s.model}</td>
+                  <td className="px-3 py-2 text-zinc-600">{limit?.category || "—"}</td>
+                  <td className="px-3 py-2 font-mono">
+                    {s.rpmUsed} / {s.rpmMax} <span className="text-zinc-400">({s.rpmRemaining} left)</span>
+                  </td>
+                  <td className="px-3 py-2 font-mono">
+                    {s.rpdUsed} / {s.rpdMax} <span className="text-zinc-400">({s.rpdRemaining} left)</span>
+                  </td>
+                  <td className="px-3 py-2 font-mono">{limit?.tpm ?? "—"}</td>
+                  <td className="px-3 py-2 font-mono">{s.inflight}</td>
+                  <td className="px-3 py-2"><PingDot snapshot={s} /></td>
+                  <td className="px-3 py-2 font-mono text-[11px] text-zinc-500">{s.msUntilNextSlot > 0 ? `${Math.ceil(s.msUntilNextSlot / 1000)}s` : "now"}</td>
+                </tr>
+              );
+            })}
+          </tbody>
+        </table>
+      </div>
+
+      <div className="mt-4 grid grid-cols-1 gap-3 md:grid-cols-3">
+        <div className="rounded-xl border border-indigo-200 bg-indigo-50/50 p-3">
+          <div className="text-[10px] font-bold uppercase tracking-wider text-indigo-700">Round-robin</div>
+          <div className="mt-1 text-[12px] text-zinc-700">pickLeastLoaded() scans all {VERITAS_FULL_LLM_ROSTER.length} entries, skips throttled via tryAcquire(), and randomizes per round-robin run. Every portfolio LLM call is wired — no exceptions.</div>
+        </div>
+        <div className="rounded-xl border border-emerald-200 bg-emerald-50/50 p-3">
+          <div className="text-[10px] font-bold uppercase tracking-wider text-emerald-700">Honest limits</div>
+          <div className="mt-1 text-[12px] text-zinc-700">RPM/TPM/RPD respected per your table. 0/15 → 15 interpreted as headroom 250K TPM, 500 RPD to avoid banning the model entirely — zero would mean permanently throttled.</div>
+        </div>
+        <div className="rounded-xl border border-amber-200 bg-amber-50/50 p-3">
+          <div className="text-[10px] font-bold uppercase tracking-wider text-amber-700">V15 integration</div>
+          <div className="mt-1 text-[12px] text-zinc-700">This page is the canonical observability surface for the entire V15 calibration, production, and scraper portfolio. If a model is READY here, it participates in the next batch bank call.</div>
+        </div>
+      </div>
+    </div>
+  );
+}

package/src/components/AdaptersPage.tsx

@@ -0,0 +1,3 @@
+/** Type A pure shim — re-exports origin from 2sagaco48 package. Do not edit node_modules. */
+export * from "./AdaptersPage.orig.tsx";
+export { default } from "./AdaptersPage.orig.tsx";