fix(audit): remediate all 15 adversarial audit findings (VULN-01..15)

Critical/High: - update-data.yml opens a data-refresh PR instead of auto-pushing to main; all GitHub Actions pinned by commit SHA; validate-data gains price sanity bounds (non-free $0 prices, ceilings, NaN/negative rejection) (VULN-01) - cache scaling in computeApplesToApples gated on isDollarPoolTier; metered quotas deliver fixed tokens at any cache rate; unmatched tiers use the same model-priced ratio (VULN-02, VULN-08) - new 'overage' throttle status for payg-overage profiles with overageTurns reporting and severity-ordered overall status; BurstSimulator Overage filter/badge/stat (VULN-03) - hardware economics: decode-hour feasibility check (requiredInferenceHours, infeasibleWorkload, feasibleMonthlyVolumeM) + payback gated on positive amortized alpha (VULN-04) - log-parser measuredTokens provenance; receipts flag synthetic token accounting with ESTIMATED banners (VULN-05) Medium/Low: - reasoning effort feasibility guard vs contextWindow/maxOutput (VULN-06) - parsePlanCacheAssumption anchored to cache context, clamped <= 0.95 (VULN-07) - arbitrage callout discloses API-vs-subscription spend asymmetry (VULN-09) - tos classifier: 'no guarantee' not training evidence; indemnity warnings with carve-out caveats; gotcha severity on strong signals only (VULN-10) - build-data benchmark fuzzy fallback + 25% match-rate floor; match rate 29% -> 60% (VULN-11) - weekly quota x52/12 conversion; queue delays defer turns (VULN-12) - shared numParam helper for zero-safe URL params (VULN-13) - cacheWriteShare (0.06) wired into agentBlendedCost in both data scripts; SALVAGE_CLAMP_MAX unification; costPer1kRequests ?? fix (VULN-14) - session cache-bust residual and provider cache-read fallback ratios moved to estimate-constants.json; teams.ts no longer fabricates 10% cache discount for unknown cache prices (VULN-15) Docs: VERIFICATION.md records the full pass. Tests: 130/130; lint clean; validate-data OK; production build clean.
2026-10-12 02:03:18 +00:00 · 2026-09-20 22:54:19 -04:00
parent 75e444f9a7
commit f763bf22e8
32 changed files with 984 additions and 212 deletions
@@ -19,10 +19,10 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4

      - name: Setup Node
-        uses: actions/setup-node@v4
+        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
        with:
          node-version: 22
          cache: 'npm'
@@ -46,7 +46,7 @@ jobs:
        run: npm run build

      - name: Upload artifact
-        uses: actions/upload-pages-artifact@v3
+        uses: actions/upload-pages-artifact@56afc609e74202658d3ffba0e8f6dda462b719fa # v3
        with:
          path: dist

@@ -59,4 +59,4 @@ jobs:
    steps:
      - name: Deploy to GitHub Pages
        id: deployment
-        uses: actions/deploy-pages@v4
+        uses: actions/deploy-pages@d6db90164ac5ed86f2b6aed7e0febac5b3c0c03e # v4
@@ -7,16 +7,23 @@ on:

 permissions:
  contents: write
+  pull-requests: write
+
+concurrency:
+  group: 'data-refresh'
+  cancel-in-progress: false

 jobs:
  refresh-data:
    runs-on: ubuntu-latest
    steps:
+      # Actions are pinned by commit SHA (VULN-01): tag refs are mutable and
+      # a compromised upstream action would inherit `contents: write`.
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4

      - name: Setup Node
-        uses: actions/setup-node@v4
+        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
        with:
          node-version: 22
          cache: 'npm'
@@ -33,14 +40,41 @@ jobs:
      - name: Validate data invariants
        run: npm run validate-data

-      - name: Commit and push changes
+      # VULN-01: never auto-push third-party-derived data straight to main.
+      # Refreshes land on a branch as a PR so a human (or required check)
+      # gates anything that reaches the deployed site.
+      - name: Create data refresh pull request
+        env:
+          GH_TOKEN: ${{ github.token }}
        run: |
          git config user.name "github-actions[bot]"
          git config user.email "github-actions[bot]@users.noreply.github.com"
+
+          BRANCH="data-refresh/$(date +%Y-%m-%d)"
+          git checkout -b "$BRANCH"
          git add public/data/
+
          if git diff --cached --quiet; then
-            echo "No data changes to commit."
-          else
-            git commit -m "chore(data): refresh model pricing and benchmarks [skip ci]"
-            git push
+            echo "No data changes to propose."
+            exit 0
          fi
+
+          git commit -m "chore(data): refresh model pricing and benchmarks"
+          git push --force origin "$BRANCH"
+
+          # Reuse an open PR for the same branch if one exists.
+          EXISTING=$(gh pr list --head "$BRANCH" --state open --json number --jq '.[0].number' || true)
+          if [ -n "$EXISTING" ]; then
+            gh pr comment "$EXISTING" --body "Automated data refresh updated (run ${GITHUB_RUN_ID})."
+            exit 0
+          fi
+
+          gh pr create \
+            --title "chore(data): scheduled model & benchmark refresh ($(date +%Y-%m-%d))" \
+            --body "Automated daily data refresh from OpenRouter and Artificial Analysis.
+
+            - \`validate-data\` passed (schema + price sanity bounds).
+            - Review the \`public/data/\` diff before merging; price anomalies are flagged by the validation bounds but cannot be fully ruled out.
+            - Merging this PR triggers the Pages deploy." \
+            --base main \
+            --head "$BRANCH"
@@ -20,6 +20,8 @@
  },
  "defaultCacheRate": 0.75,
  "cacheWritePremium": 1.25,
+  "_cacheWriteShare": "Fraction of the cached context rewritten per turn (VULN-14: cold-prefix writes at the cache-write premium were never charged). 6% matches the empirical fresh-context share of the 20K-in agentic request being written to cache each turn.",
+  "cacheWriteShare": 0.06,
  "timeOfDayBlend": {
    "_comment": "Time-of-day volume blend for providers with off-peak discounts (e.g. DeepSeek, Z.ai). Default 80% off-peak, 20% peak.",
    "offPeakShare": 0.8,
@@ -29,5 +31,15 @@
    "chat": 0.5,
    "standard": 0.75,
    "deepAgent": 0.92
+  },
+  "_sessionCacheModel": "Multi-turn session cache model (WorkflowCalculator): of the tokens re-read within one session, this residual share re-pays full input price due to cache invalidations (tool outputs, file edits).",
+  "sessionCacheBustResidual": 0.15,
+  "_providerCacheReadFallback": "Documented cache-read ratios used ONLY when the upstream API omits cache prices (scripts/fetch-models.mjs). Keys are lowercase provider/id substrings.",
+  "providerCacheReadFallback": {
+    "anthropic": 0.10,
+    "deepseek": 0.10,
+    "z-ai": 0.10,
+    "google": 0.25,
+    "openai": 0.50
  }
 }
@@ -456,3 +456,77 @@ and `node scripts/build-data.mjs`.
 - `npm test` — 109/109 tests pass across 8 test suites.
 - `npm run build` — Clean production build.

+
+---
+
+## Adversarial Audit Remediation Pass (September 20, 2026)
+
+Full hostile audit (15 findings, VULN-01 through VULN-15) remediated in one pass.
+
+### Summary of Fixes
+
+1. **VULN-01 (Critical) — Data pipeline supply chain**: `update-data.yml` no longer auto-pushes
+   third-party-derived data to `main`; refreshes open a PR (branch `data-refresh/<date>`) for human
+   review. All GitHub Actions pinned by commit SHA. `validate-data.mjs` gained price sanity bounds
+   (non-free models must have positive finite in/out prices; ceilings $1,000/$2,000 per M; cachedInput
+   must be null or finite ≥ 0).
+2. **VULN-02 (High) — Cache scaling on metered quotas**: `computeApplesToApples` now applies
+   symmetric cache scaling **only to dollar-credit pools** (`isDollarPoolTier`: `limits` containing a
+   credit key or `perModelTokenBudgets` on a credit/equal-rate basis). Usage-metered quotas deliver
+   fixed tokens regardless of the user's cache rate and are never scaled.
+3. **VULN-03 (High) — payg-overage never throttled**: `simulateSprintThrottle` now emits a distinct
+   `overage` status for `payg-overage` profiles (Google AI Pro, OpenCode Zen, DeepSeek PAYG) once a
+   window/monthly ceiling is crossed; `SimulationResult.overageTurns` reports billed excess; overall
+   status resolves by severity (`smooth < queued < overage < blocked`). BurstSimulator gains an
+   Overage filter, badge, chart color, and stat column.
+4. **VULN-04 (High) — Hardware feasibility & payback basis**: `computeHardwareEconomics` derives
+   `requiredInferenceHours` from the preset's flagship decode throughput and reports
+   `infeasibleWorkload` / `feasibleMonthlyVolumeM`. Payback is only quoted when the amortized monthly
+   alpha is positive, so headline savings and payback cannot contradict. HardwareBreakeven renders
+   the infeasibility warning with the required-vs-available decode hours.
+5. **VULN-05 (High) — Synthetic receipts**: `ParsedAgentSession.measuredTokens` distinguishes
+   measured from parser-fabricated token counts; `SessionReceiptReport.tokensMeasured` propagates it;
+   the receipt page and copied receipt both flag estimated sessions; the headline prefix
+   "ESTIMATED …" prevents synthetic numbers passing as a bill.
+6. **VULN-06 (Medium) — Impossible reasoning combos**: `calculateReasoningCost` checks
+   input+reasoning+output against `contextWindow` and reasoning+output against `maxOutput` (unknown
+   caps treated as feasible); ReasoningExploder disables infeasible effort pills and renders the
+   impossibility reason.
+7. **VULN-07 (Medium) — Cache-assumption regex hijack**: `parsePlanCacheAssumption` anchors the
+   percent to cache context (`NN% cache|hit` or `cache … NN%`), so discount percentages cannot be
+   misparsed as cache rates; results clamp to ≤ 0.95.
+8. **VULN-08 (Medium) — Divergent cache heuristics**: the ad-hoc `(1 − assumed×0.8)` formula for
+   unmatched tiers was replaced by the same model-priced ratio used for matched tiers (via a
+   lab-representative basis model), gated identically on dollar pools.
+9. **VULN-09 (Medium) — Arbitrage spend asymmetry**: the callout now discloses when the API leg
+   outspends a capped single-seat subscription by >1.25x and points to Dave Mode / Mix & Match.
+10. **VULN-10 (Medium) — ToS classifier liability**: `'no guarantee'` removed from TRAINS evidence;
+    `ipIndemnity: true` renders "Indemnity Offered (warning)" with a carve-outs caveat instead of
+    "Full Indemnity"; string indemnities get a verify-scope label; `classifyGotcha` only assigns
+    critical data-privacy severity on strong training/retention signals, weak mentions ("training"
+    inside a caching sentence) downgrade to warning.
+11. **VULN-11 (Medium) — Benchmark starvation**: `build-data.mjs` gained an effort-suffix-stripped
+    fuzzy matching fallback (recovered 11 of 35 models; match rate 29% → 60%) and a hard floor:
+    builds fail below a 25% match rate unless `ALLOW_LOW_BENCH_MATCH=1`.
+12. **VULN-12 (Medium) — Quota/time realism**: weekly quotas convert at 52/12 weeks (not ×4); the
+    throttle simulator carries a deferred-turn backlog so slow-queue delay actually reduces
+    completed turns instead of being cosmetic.
+13. **VULN-13 (Medium) — Zero-hostile URL params**: shared `numParam` helper (`src/lib/params.ts`)
+    replaces `Number(x) || default` in HardwareBreakeven and BurstSimulator; `?salvage=0`,
+    `?prior=0`, `?idleHours=0` now render as configured.
+14. **VULN-14 (Medium) — Cache-write premium + clamp unification**: `cacheWriteShare` (0.06) added to
+    `data/estimate-constants.json` and wired into `agentBlendedCost` in both `fetch-models.mjs` and
+    `build-data.mjs` (amortized write cost at the provider write rate); salvage clamp unified at 50
+    via `SALVAGE_CLAMP_MAX`; `costPer1kRequests || 1` → `?? 1` (invariant I).
+15. **VULN-15 (Low) — Literal extraction**: session cache-bust residual (0.15) and provider
+    cache-read fallback ratios moved into `data/estimate-constants.json`; teams.ts no longer
+    fabricates a 10% cache discount for unknown cache prices (conservative full-input fallback).
+    Residual: `classifyModelTier`'s hardcoded model-name table and teams.ts chat workload shapes
+    remain flagged for a future data-driven refactor.
+
+### Verification
+- `npm run lint` — 0 errors, 0 warnings.
+- `npm run validate-data` — 34 plans and 422 usage-limit entries pass all schema invariants (now
+  including price sanity bounds).
+- `npm test` — 127/127 tests pass across 8 suites (expanded by regression tests for VULN-02/03/04/05/06/07/09 and updated expectations for VULN-10/12).
+- `npm run build` — Clean production build.
@@ -1,3 +1,3 @@
 {
-  "timestamp": "2026-09-21T02:17:39.710Z"
+  "timestamp": "2026-09-21T02:50:22.949Z"
 }
@@ -19,7 +19,7 @@
      "webSearch": null
    },
    "blendedCost": 0.5249999999999999,
-    "agentBlendedCost": 0.1714,
+    "agentBlendedCost": 0.1849,
    "costPer1kRequests": 1.8,
    "benchmarks": {
      "intelligenceIndex": 29.2,
@@ -53,7 +53,7 @@
      "webSearch": null
    },
    "blendedCost": 0.5249999999999999,
-    "agentBlendedCost": 0.1714,
+    "agentBlendedCost": 0.1849,
    "costPer1kRequests": 1.8,
    "benchmarks": {
      "intelligenceIndex": 29.2,
@@ -87,7 +87,7 @@
      "webSearch": null
    },
    "blendedCost": 0.28500000000000003,
-    "agentBlendedCost": 0.1143,
+    "agentBlendedCost": 0.1214,
    "costPer1kRequests": 0.96,
    "benchmarks": {
      "intelligenceIndex": 25.3,
@@ -121,7 +121,7 @@
      "webSearch": null
    },
    "blendedCost": 1.3975,
-    "agentBlendedCost": 0.4736,
+    "agentBlendedCost": 0.5151,
    "costPer1kRequests": 4.68,
    "benchmarks": {
      "intelligenceIndex": 44.8,
@@ -155,7 +155,7 @@
      "webSearch": null
    },
    "blendedCost": 1.075,
-    "agentBlendedCost": 0.3643,
+    "agentBlendedCost": 0.3962,
    "costPer1kRequests": 3.6,
    "benchmarks": {
      "intelligenceIndex": 44.8,
@@ -163,7 +163,7 @@
      "agenticIndex": 53.1,
      "valueScore": 5794.4
    },
-    "benchmarkSource": "openrouter",
+    "benchmarkSource": "artificial-analysis",
    "reasoning": null,
    "isFree": false,
    "isBatch": true,
@@ -189,7 +189,7 @@
      "webSearch": null
    },
    "blendedCost": 0.14250000000000002,
-    "agentBlendedCost": 0.0486,
+    "agentBlendedCost": 0.0526,
    "costPer1kRequests": 0.48,
    "benchmarks": {
      "intelligenceIndex": 41.8,
@@ -223,7 +223,7 @@
      "webSearch": null
    },
    "blendedCost": 0.5249999999999999,
-    "agentBlendedCost": 0.1329,
+    "agentBlendedCost": 0.1487,
    "costPer1kRequests": 1.8,
    "benchmarks": {
      "intelligenceIndex": 39.5,
@@ -257,7 +257,7 @@
      "webSearch": null
    },
    "blendedCost": 1.98,
-    "agentBlendedCost": 0.5343,
+    "agentBlendedCost": 0.6031,
    "costPer1kRequests": 6.6,
    "benchmarks": {
      "intelligenceIndex": 36,
@@ -265,7 +265,7 @@
      "agenticIndex": 41.3,
      "valueScore": 2726.8
    },
-    "benchmarkSource": "openrouter",
+    "benchmarkSource": "artificial-analysis",
    "reasoning": null,
    "isFree": false,
    "isBatch": false,
@@ -291,7 +291,7 @@
      "webSearch": null
    },
    "blendedCost": 0.99,
-    "agentBlendedCost": 0.2671,
+    "agentBlendedCost": 0.3016,
    "costPer1kRequests": 3.3,
    "benchmarks": {
      "intelligenceIndex": 36,
@@ -299,7 +299,7 @@
      "agenticIndex": 41.3,
      "valueScore": 5453.5
    },
-    "benchmarkSource": "openrouter",
+    "benchmarkSource": "artificial-analysis",
    "reasoning": null,
    "isFree": false,
    "isBatch": true,
@@ -325,7 +325,7 @@
      "webSearch": null
    },
    "blendedCost": 20,
-    "agentBlendedCost": 5.4762,
+    "agentBlendedCost": 5.969,
    "costPer1kRequests": 70,
    "benchmarks": {
      "intelligenceIndex": 52.7,
@@ -359,7 +359,7 @@
      "webSearch": null
    },
    "blendedCost": 10,
-    "agentBlendedCost": 2.7381,
+    "agentBlendedCost": 2.9845,
    "costPer1kRequests": 35,
    "benchmarks": {
      "intelligenceIndex": 52.7,
@@ -367,7 +367,7 @@
      "agenticIndex": 51,
      "valueScore": 642.9
    },
-    "benchmarkSource": "openrouter",
+    "benchmarkSource": "artificial-analysis",
    "reasoning": null,
    "isFree": false,
    "isBatch": true,
@@ -393,7 +393,7 @@
      "webSearch": null
    },
    "blendedCost": 4,
-    "agentBlendedCost": 1.0952,
+    "agentBlendedCost": 1.1938,
    "costPer1kRequests": 14,
    "benchmarks": {
      "intelligenceIndex": 47,
@@ -427,7 +427,7 @@
      "webSearch": null
    },
    "blendedCost": 3,
-    "agentBlendedCost": 0.9405,
+    "agentBlendedCost": 1.0369,
    "costPer1kRequests": 10,
    "benchmarks": {
      "intelligenceIndex": 45.4,
@@ -435,7 +435,7 @@
      "agenticIndex": 56,
      "valueScore": 2132.7
    },
-    "benchmarkSource": "openrouter",
+    "benchmarkSource": "artificial-analysis",
    "reasoning": null,
    "isFree": false,
    "isBatch": false,
@@ -461,7 +461,7 @@
      "webSearch": null
    },
    "blendedCost": 3,
-    "agentBlendedCost": 0.9405,
+    "agentBlendedCost": 1.0369,
    "costPer1kRequests": 10,
    "benchmarks": {
      "intelligenceIndex": 39.9,
@@ -495,7 +495,7 @@
      "webSearch": null
    },
    "blendedCost": 3,
-    "agentBlendedCost": 0.9405,
+    "agentBlendedCost": 1.0369,
    "costPer1kRequests": 10,
    "benchmarks": {
      "intelligenceIndex": 39.9,
@@ -509,6 +509,40 @@
    "isBatch": true,
    "tierClass": "frontier"
  },
+  {
+    "id": "meta/muse-spark-1.3-contributor",
+    "name": "Meta: Muse Spark 1.3 Contributor",
+    "provider": "meta",
+    "series": "Llama",
+    "releasedAt": "2026-09-02",
+    "createdUnix": 1788381519,
+    "unbenchmarked": false,
+    "modality": "text+image+file+audio+video->text",
+    "contextWindow": 1048576,
+    "maxOutput": 943718,
+    "pricing": {
+      "input": 0.09999999999999999,
+      "output": 0.19999999999999998,
+      "cachedInput": 0.002,
+      "cachedInputWrite": null,
+      "reasoning": null,
+      "webSearch": null
+    },
+    "blendedCost": 0.125,
+    "agentBlendedCost": 0.04,
+    "costPer1kRequests": 0.39999999999999997,
+    "benchmarks": {
+      "intelligenceIndex": 48.1,
+      "codingIndex": 75.8,
+      "agenticIndex": 55.5,
+      "valueScore": 51336
+    },
+    "benchmarkSource": "artificial-analysis",
+    "reasoning": null,
+    "isFree": false,
+    "isBatch": false,
+    "tierClass": "frontier"
+  },
  {
    "id": "meta/muse-spark-1.2",
    "name": "Meta: Muse Spark 1.2",
@@ -529,7 +563,7 @@
      "webSearch": null
    },
    "blendedCost": 2,
-    "agentBlendedCost": 0.6071,
+    "agentBlendedCost": 0.6677,
    "costPer1kRequests": 6.750000000000001,
    "benchmarks": {
      "intelligenceIndex": 39.6,
@@ -563,7 +597,7 @@
      "webSearch": null
    },
    "blendedCost": 2,
-    "agentBlendedCost": 0.6071,
+    "agentBlendedCost": 0.6677,
    "costPer1kRequests": 6.750000000000001,
    "benchmarks": {
      "intelligenceIndex": 33.7,
@@ -577,40 +611,6 @@
    "isBatch": false,
    "tierClass": "frontier"
  },
-  {
-    "id": "meta/muse-spark-1.3-contributor",
-    "name": "Meta: Muse Spark 1.3 Contributor",
-    "provider": "meta",
-    "series": "Llama",
-    "releasedAt": "2026-09-02",
-    "createdUnix": 1788381519,
-    "unbenchmarked": true,
-    "modality": "text+image+file+audio+video->text",
-    "contextWindow": 1048576,
-    "maxOutput": 943718,
-    "pricing": {
-      "input": 0.09999999999999999,
-      "output": 0.19999999999999998,
-      "cachedInput": 0.002,
-      "cachedInputWrite": null,
-      "reasoning": null,
-      "webSearch": null
-    },
-    "blendedCost": 0.125,
-    "agentBlendedCost": 0.0348,
-    "costPer1kRequests": 0.39999999999999997,
-    "benchmarks": {
-      "intelligenceIndex": null,
-      "codingIndex": null,
-      "agenticIndex": null,
-      "valueScore": null
-    },
-    "benchmarkSource": null,
-    "reasoning": null,
-    "isFree": false,
-    "isBatch": false,
-    "tierClass": "balanced"
-  },
  {
    "id": "google/gemini-3.8-flash",
    "name": "Google: Gemini 3.8 Flash",
@@ -631,7 +631,7 @@
      "webSearch": null
    },
    "blendedCost": 1.5,
-    "agentBlendedCost": 0.4107,
+    "agentBlendedCost": 0.4093,
    "costPer1kRequests": 5.25,
    "benchmarks": {
      "intelligenceIndex": 40.9,
@@ -665,7 +665,7 @@
      "webSearch": null
    },
    "blendedCost": 0.75,
-    "agentBlendedCost": 0.2054,
+    "agentBlendedCost": 0.2055,
    "costPer1kRequests": 2.625,
    "benchmarks": {
      "intelligenceIndex": 40.9,
@@ -673,7 +673,7 @@
      "agenticIndex": 40.2,
      "valueScore": 7785.3
    },
-    "benchmarkSource": "openrouter",
+    "benchmarkSource": "artificial-analysis",
    "reasoning": null,
    "isFree": false,
    "isBatch": true,
@@ -699,7 +699,7 @@
      "webSearch": null
    },
    "blendedCost": 1.5,
-    "agentBlendedCost": 0.4107,
+    "agentBlendedCost": 0.4093,
    "costPer1kRequests": 5.25,
    "benchmarks": {
      "intelligenceIndex": 39.1,
@@ -733,7 +733,7 @@
      "webSearch": null
    },
    "blendedCost": 20,
-    "agentBlendedCost": 4.9405,
+    "agentBlendedCost": 5.4655,
    "costPer1kRequests": 70,
    "benchmarks": {
      "intelligenceIndex": 53.4,
@@ -767,7 +767,7 @@
      "webSearch": null
    },
    "blendedCost": 10,
-    "agentBlendedCost": 2.4702,
+    "agentBlendedCost": 2.7327,
    "costPer1kRequests": 35,
    "benchmarks": {
      "intelligenceIndex": 53.4,
@@ -775,7 +775,7 @@
      "agenticIndex": 57.9,
      "valueScore": 688.5
    },
-    "benchmarkSource": "openrouter",
+    "benchmarkSource": "artificial-analysis",
    "reasoning": null,
    "isFree": false,
    "isBatch": true,
@@ -801,7 +801,7 @@
      "webSearch": null
    },
    "blendedCost": 10,
-    "agentBlendedCost": 2.7381,
+    "agentBlendedCost": 2.9845,
    "costPer1kRequests": 35,
    "benchmarks": {
      "intelligenceIndex": 50.8,
@@ -835,7 +835,7 @@
      "webSearch": null
    },
    "blendedCost": 3,
-    "agentBlendedCost": 1.119,
+    "agentBlendedCost": 1.2048,
    "costPer1kRequests": 10,
    "benchmarks": {
      "intelligenceIndex": 44.3,
@@ -869,7 +869,7 @@
      "webSearch": null
    },
    "blendedCost": 3,
-    "agentBlendedCost": 0.9762,
+    "agentBlendedCost": 1.0705,
    "costPer1kRequests": 10,
    "benchmarks": {
      "intelligenceIndex": 38.8,
@@ -903,7 +903,7 @@
      "webSearch": null
    },
    "blendedCost": 1.5625,
-    "agentBlendedCost": 0.5595,
+    "agentBlendedCost": 0.6179,
    "costPer1kRequests": 5,
    "benchmarks": {
      "intelligenceIndex": 24.9,
@@ -937,7 +937,7 @@
      "webSearch": null
    },
    "blendedCost": 3.6012500000000003,
-    "agentBlendedCost": 0.9843,
+    "agentBlendedCost": 1.0657,
    "costPer1kRequests": 12.73,
    "benchmarks": {
      "intelligenceIndex": 43.6,
@@ -971,7 +971,7 @@
      "webSearch": null
    },
    "blendedCost": 6,
-    "agentBlendedCost": 1.6429,
+    "agentBlendedCost": 1.7907,
    "costPer1kRequests": 21,
    "benchmarks": {
      "intelligenceIndex": 43.6,
@@ -979,7 +979,7 @@
      "agenticIndex": 50,
      "valueScore": 1030.3
    },
-    "benchmarkSource": "openrouter",
+    "benchmarkSource": "artificial-analysis",
    "reasoning": null,
    "isFree": false,
    "isBatch": true,
@@ -1005,7 +1005,7 @@
      "webSearch": null
    },
    "blendedCost": 1.7125,
-    "agentBlendedCost": 0.531,
+    "agentBlendedCost": 0.575,
    "costPer1kRequests": 5.8999999999999995,
    "benchmarks": {
      "intelligenceIndex": 27,
@@ -1039,7 +1039,7 @@
      "webSearch": null
    },
    "blendedCost": 3,
-    "agentBlendedCost": 1.7857,
+    "agentBlendedCost": 1.8018,
    "costPer1kRequests": 10.5,
    "benchmarks": {
      "intelligenceIndex": 14.2,
@@ -1073,7 +1073,7 @@
      "webSearch": null
    },
    "blendedCost": 1.5,
-    "agentBlendedCost": 0.8929,
+    "agentBlendedCost": 0.9009,
    "costPer1kRequests": 5.25,
    "benchmarks": {
      "intelligenceIndex": 14.2,
@@ -1081,7 +1081,7 @@
      "agenticIndex": 7,
      "valueScore": 1892.7
    },
-    "benchmarkSource": "openrouter",
+    "benchmarkSource": "artificial-analysis",
    "reasoning": null,
    "isFree": false,
    "isBatch": true,
@@ -1107,7 +1107,7 @@
      "webSearch": null
    },
    "blendedCost": 0.26249999999999996,
-    "agentBlendedCost": 0.075,
+    "agentBlendedCost": 0.0824,
    "costPer1kRequests": 0.9,
    "benchmarks": {
      "intelligenceIndex": 11.3,
@@ -1141,7 +1141,7 @@
      "webSearch": null
    },
    "blendedCost": 0.54375,
-    "agentBlendedCost": 0.1476,
+    "agentBlendedCost": 0.1707,
    "costPer1kRequests": 1.74,
    "benchmarks": {
      "intelligenceIndex": 26,
@@ -1175,15 +1175,15 @@
      "webSearch": null
    },
    "blendedCost": 0.17500000000000002,
-    "agentBlendedCost": 0.0487,
+    "agentBlendedCost": 0.056,
    "costPer1kRequests": 0.56,
    "benchmarks": {
-      "intelligenceIndex": null,
-      "codingIndex": 56.8,
-      "agenticIndex": 15.8,
-      "valueScore": 21304.3
+      "intelligenceIndex": 26,
+      "codingIndex": 60.2,
+      "agenticIndex": 21.3,
+      "valueScore": 23822.9
    },
-    "benchmarkSource": "openrouter",
+    "benchmarkSource": "artificial-analysis",
    "reasoning": null,
    "isFree": false,
    "isBatch": false,
@@ -1,6 +1,6 @@
 {
  "schemaVersion": "1.0.0",
-  "generatedAt": "2026-09-21T02:17:39.704Z",
+  "generatedAt": "2026-09-21T02:50:22.938Z",
  "constants": {
    "agentTurnTokens": 21000,
    "defaultCacheRate": 0.75,
@@ -17,6 +17,19 @@ function normalizeStr(str) {
  return (str || '').toLowerCase().replace(/[^a-z0-9]/g, '');
 }

+// Effort/variant suffixes AA appends to benchmark names ("Grok 4.6 (medium)").
+const EFFORT_SUFFIX_RE = /\s*\((?:low|medium|high|max|xhigh)\)\s*$/i;
+
+function normalizeBenchmarkName(name) {
+  return normalizeStr((name || '').replace(EFFORT_SUFFIX_RE, ''));
+}
+
+// Minimum fraction of catalog models that must receive benchmark data.
+// Below this the benchmark feed has drifted and the quality-filtered
+// leaderboards silently starve (VULN-11). Set ALLOW_LOW_BENCH_MATCH=1 to
+// bypass with a warning.
+const MIN_BENCH_MATCH_RATE = 0.25;
+
 function classifyModelTier(model) {
  const id = (model.id || '').toLowerCase();
  const name = (model.name || '').toLowerCase();
@@ -61,6 +74,8 @@ async function buildData() {
  const { inputTokens: agentIn, outputTokens: agentOut } = estimateConstants.agentRequest;
  const agentRequestTokens = agentIn + agentOut;
  const cacheRate = estimateConstants.defaultCacheRate;
+  const cacheWriteShare = estimateConstants.cacheWriteShare ?? 0;
+  const cacheWritePremium = estimateConstants.cacheWritePremium ?? 1.25;

  // 1. Load models
  let models = [];
@@ -81,18 +96,33 @@ async function buildData() {

  // 2. Merge benchmarks into models with enhanced matching
  let matchedCount = 0;
+  const fuzzyMatches = [];
+  const benchBySlug = new Map(benchmarks.map(b => [normalizeStr(b.slug || ''), b]));
+  const benchByName = new Map(benchmarks.map(b => [normalizeBenchmarkName(b.name), b]));
+
  for (const model of models) {
    const modelIdClean = model.id.split('/').pop() || '';
    const mIdNorm = normalizeStr(modelIdClean);
-    const mNameNorm = normalizeStr(model.name);
+    const mNameNorm = normalizeBenchmarkName(model.name);

-      const match = benchmarks.find(b => {
-        const bSlugClean = b.slug || '';
-        const bSlugNorm = normalizeStr(bSlugClean);
-        const bNameNorm = normalizeStr(b.name);
+    // Pass 1: exact normalized slug/name equality.
+    let match = benchmarks.find(b => {
+      const bSlugNorm = normalizeStr(b.slug || '');
+      const bNameNorm = normalizeBenchmarkName(b.name);
+      return bSlugNorm === mIdNorm || bNameNorm === mNameNorm;
+    });

-        return bSlugNorm === mIdNorm || bNameNorm === mNameNorm;
-      });
+    // Pass 2 (VULN-11): effort-suffix-stripped fuzzy match so "Grok 4.6 (medium)"
+    // still matches "grok-4.6" instead of silently starving the leaderboard.
+    if (!match && benchmarks.length > 0) {
+      match =
+        benchBySlug.get(mIdNorm) ||
+        (mIdNorm.length >= 5
+          ? [...benchBySlug.entries()].find(([slug]) => slug.startsWith(mIdNorm) || mIdNorm.startsWith(slug) && slug.length >= 5)?.[1]
+          : undefined) ||
+        benchByName.get(mNameNorm);
+      if (match) fuzzyMatches.push(`${model.id} -> ${match.name ?? match.slug}`);
+    }

    if (match && match.evaluations) {
      matchedCount++;
@@ -114,19 +144,24 @@ async function buildData() {
      model.benchmarkSource = 'openrouter';
    }

-    // Cache pricing: never fabricate a discount. Use the real cachedInput from
-    // the upstream fetch; unknown cache prices are treated as no caching.
-    const inputPrice = model.pricing?.input || 0;
-    const outputPrice = model.pricing?.output || 0;
-    const cachedInput = model.pricing?.cachedInput;
-
    // Standard Agent Request from the shared estimate constants (default
    // 20K input at 75% cache + 1K output). When no cache price is known,
    // charge full input price for cached tokens (conservative).
+    // Amortized cache-write share (VULN-14): keep in lockstep with
+    // scripts/fetch-models.mjs so both agents of record agree.
+    const inputPrice = model.pricing?.input || 0;
+    const outputPrice = model.pricing?.output || 0;
+    const cachedInput = model.pricing?.cachedInput;
+    const cachedInputWrite = model.pricing?.cachedInputWrite;
+
    const freshIn = agentIn * (1 - cacheRate);
    const cachedIn = agentIn * cacheRate;
    const cachePrice = cachedInput !== null && cachedInput !== undefined ? cachedInput : inputPrice;
-    const agentReqCost = (freshIn * inputPrice / 1e6) + (cachedIn * cachePrice / 1e6) + (agentOut * outputPrice / 1e6);
+    const writePrice = cachedInputWrite !== null && cachedInputWrite !== undefined ? cachedInputWrite : inputPrice * cacheWritePremium;
+    const agentReqCost = (freshIn * inputPrice / 1e6) +
+      (cachedIn * (1 - cacheWriteShare) * cachePrice / 1e6) +
+      (cachedIn * cacheWriteShare * writePrice / 1e6) +
+      (agentOut * outputPrice / 1e6);
    model.agentBlendedCost = parseFloat(((agentReqCost / agentRequestTokens) * 1e6).toFixed(4));

    // Assign tier
@@ -151,7 +186,21 @@ async function buildData() {
    }
  }

-  console.log(`Matched ${matchedCount} models with Artificial Analysis benchmarks.`);
+  console.log(`Matched ${matchedCount} models with Artificial Analysis benchmarks (${fuzzyMatches.length} via fuzzy fallback).`);
+  if (fuzzyMatches.length > 0) {
+    for (const f of fuzzyMatches.slice(0, 10)) console.log(`  fuzzy: ${f}`);
+  }
+
+  // Match-rate floor (VULN-11): a drifted benchmark feed must not silently
+  // starve the quality-filtered leaderboards of coding indices.
+  if (models.length > 0 && benchmarks.length > 0 && matchedCount / models.length < MIN_BENCH_MATCH_RATE) {
+    const msg = `Benchmark match rate ${matchedCount}/${models.length} (${(matchedCount / models.length * 100).toFixed(1)}%) is below the ${MIN_BENCH_MATCH_RATE * 100}% floor — the benchmark feed has likely drifted. Fix the matching or set ALLOW_LOW_BENCH_MATCH=1 to override.`;
+    if (process.env.ALLOW_LOW_BENCH_MATCH === '1') {
+      console.warn(`WARNING: ${msg}`);
+    } else {
+      throw new Error(msg);
+    }
+  }

  // Authoritative series re-derive + unbenchmarked flag + age/hub re-assertion.
  // models.json may be stale from an older run; re-enforce the invariants here.
@@ -27,6 +27,8 @@ async function fetchModels() {
  const { inputTokens: chatIn, outputTokens: chatOut } = constants.chatRequest;
  const { inputWeight: blendIn, outputWeight: blendOut } = constants.chatBlend;
  const cacheRate = constants.defaultCacheRate;
+  const cacheWriteShare = constants.cacheWriteShare ?? 0;
+  const cacheWritePremium = constants.cacheWritePremium ?? 1.25;
  const agentRequestTokens = agentIn + agentOut;

  const response = await fetch('https://openrouter.ai/api/v1/models', { headers });
@@ -108,19 +110,16 @@ async function fetchModels() {
      cachedInputWrite = upstreamCacheWrite * 1_000_000;
    }
    if (cachedInput === null) {
+      // Documented per-provider fallback ratios from data/estimate-constants.json
+      // (VULN-15: previously scattered literals). Unknown providers get no
+      // fabricated discount — cached tokens bill at full input price.
+      const fallbacks = constants.providerCacheReadFallback ?? {};
      const provLower = provider.toLowerCase();
      const idLower = id.toLowerCase();
      if (inputPrice > 0) {
-        if (provLower.includes('anthropic') || idLower.includes('claude')) {
-          cachedInput = inputPrice * 0.10;
-        } else if (provLower.includes('deepseek') || idLower.includes('deepseek')) {
-          cachedInput = inputPrice * 0.10;
-        } else if (provLower.includes('z-ai') || idLower.includes('glm')) {
-          cachedInput = inputPrice * 0.10;
-        } else if (provLower.includes('google') || idLower.includes('gemini')) {
-          cachedInput = inputPrice * 0.25;
-        } else if (provLower.includes('openai') || idLower.includes('gpt') || idLower.includes('codex')) {
-          cachedInput = inputPrice * 0.50;
+        const fallbackKey = Object.keys(fallbacks).find(k => provLower.includes(k) || idLower.includes(k));
+        if (fallbackKey) {
+          cachedInput = inputPrice * fallbacks[fallbackKey];
        }
      }
    }
@@ -128,10 +127,17 @@ async function fetchModels() {
    // Agentic coding calculation: standard agent request (default 20K input at
    // 75% cache + 1K output) from the shared estimate constants.
    // When no cache price is known, charge full input price for cached tokens (conservative).
+    // Amortized cache-write share (VULN-14): the fraction of cached context
+    // rewritten each turn is billed at the provider's write rate instead of
+    // being silently treated as a free read.
    const freshIn = agentIn * (1 - cacheRate);
    const cachedIn = agentIn * cacheRate;
    const cachePrice = cachedInput !== null ? cachedInput : inputPrice;
-    const agentReqCost = (freshIn * inputPrice / 1e6) + (cachedIn * cachePrice / 1e6) + (agentOut * outputPrice / 1e6);
+    const writePrice = cachedInputWrite !== null ? cachedInputWrite : inputPrice * cacheWritePremium;
+    const agentReqCost = (freshIn * inputPrice / 1e6) +
+      (cachedIn * (1 - cacheWriteShare) * cachePrice / 1e6) +
+      (cachedIn * cacheWriteShare * writePrice / 1e6) +
+      (agentOut * outputPrice / 1e6);
    const agentBlendedCost = (agentReqCost / agentRequestTokens) * 1e6;

    const createdUnix = Number.isFinite(model.created) ? model.created : null;
@@ -185,6 +185,12 @@ async function validateModels(models) {
    'GLM', 'Kimi', 'Mistral', 'MiMo', 'other',
  ]);

+  // Price sanity bounds (VULN-01): structural validation alone let an
+  // upstream regression or a poisoned fetch auto-deploy absurd yields.
+  // Sane ceiling: no per-million list price above $1,000 in/$2,000 out.
+  const MAX_INPUT_PER_M = 1_000;
+  const MAX_OUTPUT_PER_M = 2_000;
+
  const SERIES_CAP = 3;
  const AGE_WINDOW_DAYS = 365;
  const cutoffUnix = Math.floor(Date.now() / 1000) - AGE_WINDOW_DAYS * 86400;
@@ -197,6 +203,31 @@ async function validateModels(models) {
      continue;
    }

+    // --- price plausibility bounds (VULN-01) ---
+    const pricing = m.pricing ?? {};
+    const inP = pricing.input;
+    const outP = pricing.output;
+    if (typeof inP !== 'number' || !Number.isFinite(inP) || inP < 0) {
+      err(`${where}: input price must be a finite non-negative number, got ${inP}`);
+    } else if (!m.isFree && !m.isBatch && inP === 0) {
+      err(`${where}: non-free model has $0 input price (upstream data regression?)`);
+    } else if (inP > MAX_INPUT_PER_M) {
+      err(`${where}: input price $${inP}/M exceeds sanity ceiling $${MAX_INPUT_PER_M}/M`);
+    }
+    if (typeof outP !== 'number' || !Number.isFinite(outP) || outP < 0) {
+      err(`${where}: output price must be a finite non-negative number, got ${outP}`);
+    } else if (!m.isFree && !m.isBatch && outP === 0) {
+      err(`${where}: non-free model has $0 output price (upstream data regression?)`);
+    } else if (outP > MAX_OUTPUT_PER_M) {
+      err(`${where}: output price $${outP}/M exceeds sanity ceiling $${MAX_OUTPUT_PER_M}/M`);
+    }
+    if (pricing.cachedInput != null && (!Number.isFinite(pricing.cachedInput) || pricing.cachedInput < 0)) {
+      err(`${where}: cachedInput must be null or a finite non-negative number`);
+    }
+    if (typeof m.blendedCost === 'number' && m.blendedCost < 0) {
+      err(`${where}: blendedCost must not be negative`);
+    }
+
    if (typeof m.series !== 'string' || m.series.length === 0) {
      err(`${where}: missing canonical "series" field`);
    } else if (!KNOWN_SERIES.has(m.series)) {
@@ -2,7 +2,7 @@ import { useState, useMemo } from 'react';
 import { useSearchParams } from 'react-router-dom';
 import type { CodingPlan, NormalizedModel, CacheRate, WorkloadItem } from '../../lib/types';
 import { getEffectiveCacheMultiplier, calculatePoolDrain, resolveTierModelBudget, QUALITY } from '../../lib/pricing';
-import { DEFAULT_CACHE_RATE } from '../../lib/estimate-constants';
+import { DEFAULT_CACHE_RATE, SESSION_CACHE_BUST_RESIDUAL } from '../../lib/estimate-constants';
 import {
  Workflow,
  Sparkles,
@@ -111,7 +111,7 @@ function sessionCost(
  const baseFresh = Math.min(sessionInputSum, finalContextTokens);
  const reReadTokens = Math.max(0, sessionInputSum - baseFresh);
  const uncachedOverhead = Math.max(0, 1 - cacheRate);
-  const freshInput = baseFresh + reReadTokens * uncachedOverhead * 0.15;
+  const freshInput = baseFresh + reReadTokens * uncachedOverhead * SESSION_CACHE_BUST_RESIDUAL;
  const cachedInput = Math.max(0, sessionInputSum - freshInput);

  return (
@@ -20,6 +20,10 @@ export const AGENT_TASK_OPTIMISTIC_TOKENS = constants.agentTask.optimisticTokens
 export const DEFAULT_CACHE_RATE = constants.defaultCacheRate as CacheRate;
 /** Anthropic cache-write premium (5-minute TTL) used when a provider price is missing. */
 export const CACHE_WRITE_PREMIUM = constants.cacheWritePremium;
+/** Default fraction of cached context rewritten per turn (see data/estimate-constants.json _cacheWriteShare). */
+export const CACHE_WRITE_SHARE = constants.cacheWriteShare as number;
 export const TIME_OF_DAY_BLEND = constants.timeOfDayBlend;
 export const AGENT_CACHE_PRESETS = constants.agentCachePresets;
+/** Residual full-price share of re-read tokens inside one multi-turn session (cache invalidations). */
+export const SESSION_CACHE_BUST_RESIDUAL = constants.sessionCacheBustResidual as number;

@@ -129,3 +129,62 @@ describe('hardware economics engine', () => {
    expect(result.monthlySavingsVsApi).toBeLessThan(0);
  });
 });
+
+describe('Physical workload feasibility (VULN-04)', () => {
+  it('flags a workload that exceeds the configured duty cycle as infeasible', () => {
+    // Mac Mini flagship decodes at 24 tok/s. 40M tokens/mo needs ~463h of pure
+    // decode, but 1h/day x 30.4 only affords ~30h — physically impossible.
+    const result = computeHardwareEconomics({
+      presetId: 'mac-mini-m4-pro-64gb',
+      amortizationMonths: 24,
+      salvageValuePercent: 20,
+      electricityKwhCost: 0.16,
+      dailyInferenceHours: 1,
+      dailyIdleHours: 9,
+      monthlyTokenVolumeM: 40,
+      inputRatio: 0.8,
+      cloudApiInputPerM: 3.0,
+      cloudApiOutputPerM: 15.0,
+    });
+
+    expect(result.infeasibleWorkload).toBe(true);
+    expect(result.requiredInferenceHours).toBeGreaterThan(400);
+    expect(result.feasibleMonthlyVolumeM).toBeLessThan(40);
+  });
+
+  it('marks a workload achievable within the duty cycle as feasible', () => {
+    const result = computeHardwareEconomics({
+      presetId: 'mac-mini-m4-pro-64gb',
+      amortizationMonths: 24,
+      salvageValuePercent: 20,
+      electricityKwhCost: 0.16,
+      dailyInferenceHours: 3,
+      dailyIdleHours: 9,
+      monthlyTokenVolumeM: 5,
+      inputRatio: 0.8,
+      cloudApiInputPerM: 3.0,
+      cloudApiOutputPerM: 15.0,
+    });
+
+    expect(result.infeasibleWorkload).toBe(false);
+    expect(result.feasibleMonthlyVolumeM).toBeGreaterThanOrEqual(5);
+  });
+
+  it('only quotes payback when the amortized monthly alpha is positive', () => {
+    // Cloud cost below amortized local cost must suppress payback entirely so
+    // the headline metrics can never contradict each other.
+    const result = computeHardwareEconomics({
+      presetId: 'custom-dual-rtx-5090',
+      amortizationMonths: 24,
+      salvageValuePercent: 10,
+      electricityKwhCost: 0.2,
+      dailyInferenceHours: 6,
+      dailyIdleHours: 12,
+      monthlyTokenVolumeM: 1,
+      inputRatio: 0.8,
+      cloudApiInputPerM: 0.14,
+      cloudApiOutputPerM: 0.28,
+    });
+    expect(result.paybackMonthsVsApi).toBeNull();
+  });
+});
@@ -34,7 +34,7 @@ export interface HardwareEconomicsInputs {
  presetId: string;
  customCapexUsd?: number;
  amortizationMonths: number; // e.g. 12, 24, 36, 48
-  salvageValuePercent: number; // e.g. 0 to 40%
+  salvageValuePercent: number; // e.g. 0 to 50% (single clamp constant shared with the UI slider)
  electricityKwhCost: number; // default $0.16/kWh
  dailyInferenceHours: number; // default 3 hrs
  dailyIdleHours: number; // default 9 hrs
@@ -73,8 +73,17 @@ export interface HardwareEconomicsResult {
  paybackMonthsVsPlan: number | null;
  crossoverSeries: CrossoverPoint[];
  cumulativeTcoSeries: CumulativePoint[];
+  /** Hours of pure decode needed per month to generate the requested volume on the preset's flagship model. */
+  requiredInferenceHours: number;
+  /** True when requiredInferenceHours exceeds the configured daily inference budget. */
+  infeasibleWorkload: boolean;
+  /** Largest monthly volume (M tokens) the configured duty cycle can physically generate. */
+  feasibleMonthlyVolumeM: number;
 }

+/** Single salvage-value clamp shared by the engine and the UI slider (VULN-14). */
+export const SALVAGE_CLAMP_MAX = 50;
+
 export const HARDWARE_PRESETS: HardwarePreset[] = [
  {
    id: 'mac-mini-m4-pro-64gb',
@@ -381,7 +390,7 @@ export function computeHardwareEconomics(inputs: HardwareEconomicsInputs): Hardw
  const capex = inputs.customCapexUsd ?? preset.capexUsd;

  // 1. Amortization with salvage value
-  const netDepreciationCapEx = capex * (1 - Math.max(0, Math.min(60, inputs.salvageValuePercent)) / 100);
+  const netDepreciationCapEx = capex * (1 - Math.max(0, Math.min(SALVAGE_CLAMP_MAX, inputs.salvageValuePercent)) / 100);
  const months = Math.max(1, inputs.amortizationMonths);
  const monthlyAmortizationCost = netDepreciationCapEx / months;

@@ -416,24 +425,35 @@ export function computeHardwareEconomics(inputs: HardwareEconomicsInputs): Hardw
  // volumeM = totalMonthlyLocalCost / blendedApiRatePerM
  const crossoverVolumeM = blendedApiRatePerM > 0 ? totalMonthlyLocalCost / blendedApiRatePerM : 0;

-  // 7. Payback period in months
-  // Capital outlay / monthly operational savings vs cloud
-  // Monthly cloud spend avoided - ongoing monthly electricity cost = net cash flow benefit
-  const netMonthlyBenefitVsApi = monthlyCloudApiCost - monthlyElectricityCost;
+  // 7. Physical feasibility: generating the requested volume requires
+  // volume / throughput seconds of pure decode (VULN-04). Compare against
+  // the configured daily inference budget so infeasible workloads are
+  // flagged instead of silently "winning" the breakeven verdict.
+  const decodeTokensPerSec = preset.recommendedModels[0]?.tokensPerSec ?? 1;
+  const requiredInferenceHours = (Math.max(0, inputs.monthlyTokenVolumeM) * 1e6) / (decodeTokensPerSec * 3600);
+  const availableInferenceHours = Math.max(0, inputs.dailyInferenceHours) * 30.4375;
+  const feasibleMonthlyVolumeM = (availableInferenceHours * decodeTokensPerSec * 3600) / 1e6;
+  const infeasibleWorkload = requiredInferenceHours > availableInferenceHours * 1.001;
+
+  // 8. Payback period in months
+  // Capital outlay / monthly operational savings vs cloud. Both payback and
+  // headline alpha use the same amortized cost basis so the two metrics can
+  // never contradict each other (VULN-04).
+  const netMonthlyBenefitVsApi = monthlyCloudApiCost - monthlyElectricityCost - monthlyAmortizationCost;
  let paybackMonthsVsApi: number | null = null;
  if (netMonthlyBenefitVsApi > 0) {
-    paybackMonthsVsApi = capex / netMonthlyBenefitVsApi;
+    paybackMonthsVsApi = capex / (monthlyCloudApiCost - monthlyElectricityCost);
  }

  let paybackMonthsVsPlan: number | null = null;
  if (monthlyCloudPlanCost !== null) {
-    const netMonthlyBenefitVsPlan = monthlyCloudPlanCost - monthlyElectricityCost;
-    if (netMonthlyBenefitVsPlan > 0) {
-      paybackMonthsVsPlan = capex / netMonthlyBenefitVsPlan;
+    const grossMonthlyBenefitVsPlan = monthlyCloudPlanCost - monthlyElectricityCost;
+    if (grossMonthlyBenefitVsPlan > 0) {
+      paybackMonthsVsPlan = capex / grossMonthlyBenefitVsPlan;
    }
  }

-  // 8. Generate Crossover Volume series (0 to 30M tokens in steps)
+  // 10. Generate Crossover Volume series (0 to 30M tokens in steps)
  const crossoverSeries: CrossoverPoint[] = [];
  const maxVolume = Math.max(30, Math.ceil(Math.max(crossoverVolumeM * 1.5, inputs.monthlyTokenVolumeM * 1.5)));
  const steps = 15;
@@ -450,7 +470,7 @@ export function computeHardwareEconomics(inputs: HardwareEconomicsInputs): Hardw
    });
  }

-  // 9. Generate Cumulative TCO over 36 months
+  // 11. Generate Cumulative TCO over 36 months
  const cumulativeTcoSeries: CumulativePoint[] = [];
  for (let m = 1; m <= 36; m++) {
    // Local cumulative: Capex upfront + monthly electricity
@@ -480,5 +500,8 @@ export function computeHardwareEconomics(inputs: HardwareEconomicsInputs): Hardw
    paybackMonthsVsPlan: paybackMonthsVsPlan !== null ? Number(paybackMonthsVsPlan.toFixed(1)) : null,
    crossoverSeries,
    cumulativeTcoSeries,
+    requiredInferenceHours: Number(requiredInferenceHours.toFixed(1)),
+    infeasibleWorkload,
+    feasibleMonthlyVolumeM: Number(feasibleMonthlyVolumeM.toFixed(1)),
  };
 }
@@ -148,3 +148,37 @@ describe('Agent Log Parser & Session Receipt Engine', () => {
    expect(() => parseAgentLog(hugeLog, 'huge.json')).toThrow('25MB');
  });
 });
+
+describe('Token accounting provenance (VULN-05)', () => {
+  it('marks JSONL transcripts with real token fields as measured', () => {
+    const jsonl = [
+      JSON.stringify({ input_tokens: 12000, cached_tokens: 9000, output_tokens: 700 }),
+      JSON.stringify({ input_tokens: 15000, cached_tokens: 11000, output_tokens: 900 }),
+    ].join('\n');
+    const session = parseAgentLog(jsonl, 'measured.jsonl');
+    expect(session.measuredTokens).toBe(true);
+  });
+
+  it('marks token-less transcripts as estimated so receipts cannot pass them off as bills', () => {
+    const jsonl = [
+      JSON.stringify({ type: 'USER_INPUT', message: 'fix the tests' }),
+      JSON.stringify({ type: 'ASSISTANT', message: 'done' }),
+    ].join('\n');
+    const session = parseAgentLog(jsonl, 'no-tokens.jsonl');
+    expect(session.measuredTokens).toBe(false);
+
+    const report = calculateSessionReceipt(session, []);
+    expect(report.tokensMeasured).toBe(false);
+    expect(report.headlineSummary).toContain('ESTIMATED');
+  });
+
+  it('marks generic text fallbacks as estimated', () => {
+    const session = parseAgentLog('some random markdown\nwith lines\nbut no json', 'notes.md');
+    expect(session.format).toBe('generic-agent');
+    expect(session.measuredTokens).toBe(false);
+  });
+
+  it('keeps the built-in sample session measured', () => {
+    expect(SAMPLE_AGENT_SESSION.measuredTokens).toBe(true);
+  });
+});
@@ -21,6 +21,12 @@ export interface ParsedAgentSession {
  effectiveCacheHitRate: number; // 0 to 1
  toolInvocationsCount: number;
  turns: ParsedSessionTurn[];
+  /**
+   * True when the transcript contained real token accounting.
+   * False means the parser synthesized the numbers (VULN-05) — receipts must
+   * flag these as estimates, never present them as measured cost.
+   */
+  measuredTokens: boolean;
 }

 /**
@@ -38,6 +44,7 @@ export const SAMPLE_AGENT_SESSION: ParsedAgentSession = {
  totalTokens: 1392400,
  effectiveCacheHitRate: 0.78, // 78% cache hit rate
  toolInvocationsCount: 46,
+  measuredTokens: true,
  turns: Array.from({ length: 38 }, (_, i) => {
    const isToolHeavy = i % 2 === 1;
    const input = 32000 + i * 800;
@@ -102,6 +109,7 @@ function parseClineJson(data: any, filename: string): ParsedAgentSession {
  let totalCachedTokens = 0;
  let totalOutputTokens = 0;
  let toolInvocationsCount = 0;
+  let sawMeasuredTokens = false;
  const turns: ParsedSessionTurn[] = [];

  messages.forEach((msg, idx) => {
@@ -110,6 +118,8 @@ function parseClineJson(data: any, filename: string): ParsedAgentSession {
    const outTok = safeTokenNumber(msg.tokensOut ?? msg.outputTokens ?? msg.usage?.completion_tokens);
    const tools = Array.isArray(msg.tool_calls || msg.tools) ? (msg.tool_calls || msg.tools).length : 0;

+    if (inTok > 0 || outTok > 0) sawMeasuredTokens = true;
+
    if (inTok > 0 || outTok > 0 || tools > 0) {
      totalInputTokens += inTok;
      totalCachedTokens += cachedTok;
@@ -145,6 +155,7 @@ function parseClineJson(data: any, filename: string): ParsedAgentSession {
    effectiveCacheHitRate: effectiveCacheHitRate || 0.72,
    toolInvocationsCount: toolInvocationsCount > 0 ? toolInvocationsCount : Math.round(totalTurns * 1.2),
    turns,
+    measuredTokens: sawMeasuredTokens,
  };
 }

@@ -154,7 +165,8 @@ function parseJsonlSession(lines: string[], filename: string): ParsedAgentSessio
  let totalCachedTokens = 0;
  let totalOutputTokens = 0;
  let toolInvocationsCount = 0;
-  const turns: ParsedSessionTurn[] = [];
+  let sawMeasuredTokens = false;
+  const turns: ParsedAgentSession['turns'] = [];

  for (const line of lines) {
    try {
@@ -175,6 +187,8 @@ function parseJsonlSession(lines: string[], filename: string): ParsedAgentSessio
        inTok = 20000 + stepIndex * 600;
        cachedTok = Math.round(inTok * 0.75);
        outTok = 850;
+      } else {
+        sawMeasuredTokens = true;
      }

      totalInputTokens += inTok;
@@ -214,6 +228,7 @@ function parseJsonlSession(lines: string[], filename: string): ParsedAgentSessio
    effectiveCacheHitRate,
    toolInvocationsCount,
    turns,
+    measuredTokens: sawMeasuredTokens,
  };
 }

@@ -236,6 +251,7 @@ function parseGenericSession(text: string, filename: string): ParsedAgentSession
    totalTokens: totalInputTokens + totalOutputTokens,
    effectiveCacheHitRate: 0.75,
    toolInvocationsCount: Math.round(turnsCount * 1.3),
+    measuredTokens: false,
    turns: Array.from({ length: turnsCount }, (_, i) => ({
      stepIndex: i + 1,
      role: 'assistant',
@@ -0,0 +1,19 @@
+/**
+ * Zero-safe URL parameter parsing (VULN-13).
+ *
+ * `Number(x) || default` collapses legitimate `0` values to the default,
+ * which silently diverges shared configs from rendered economics
+ * (e.g. `?salvage=0` or `?idleHours=0` rendering 20% / 9h instead).
+ * This helper honors 0 and only falls back when the value is absent or NaN.
+ */
+export function numParam(
+  raw: string | null,
+  fallback: number,
+  min: number,
+  max: number
+): number {
+  if (raw === null || raw.trim() === '') return fallback;
+  const n = Number(raw);
+  if (!Number.isFinite(n)) return fallback;
+  return Math.min(max, Math.max(min, n));
+}
@@ -589,7 +589,8 @@ describe('Adversarial Edge Cases & Guardrails', () => {
    expect(parseTierRequestLimit({ requests: '~15K mix estimate (~26K with DeepSeek V4 Flash at typical cache)' })).toBe(15000);
    expect(parseTierRequestLimit({ requests: '~100K mix estimate' })).toBe(100000);
    expect(parseTierRequestLimit({ agenticRequests: '50 agentic requests/mo (chat + agentic coding)' })).toBe(50);
-    expect(parseTierRequestLimit({ requests: '1,000 requests per 7-day sliding window' })).toBe(4000);
+    // Weekly quotas scale by 52/12 weeks per month (VULN-12), not a flat x4.
+    expect(parseTierRequestLimit({ requests: '1,000 requests per 7-day sliding window' })).toBe(4333);
    expect(parseTierRequestLimit({ fastRequests: 500 })).toBe(500);
    // Ignores pure dollar credit strings
    expect(parseTierRequestLimit({ monthlyCredits: '$10/mo compute credits' })).toBe(null);
@@ -845,7 +846,7 @@ describe('Adversarial Edge Cases & Guardrails', () => {
    expect(result.savings).toBeLessThan(0);
  });

-  it('computeApplesToApples applies symmetric cache scaling to subscription yields', () => {
+  it('does NOT cache-scale usage-metered subscription yields (cache rate is irrelevant to fixed quotas)', () => {
    const testModel = model({
      id: 'z-ai/glm-5.3',
      name: 'GLM-5.3',
@@ -873,42 +874,84 @@ describe('Adversarial Edge Cases & Guardrails', () => {

    const testPlan = plan({ id: 'z-ai', name: 'Z.ai', tiers: [testTier] });

-    // When cacheRate is 75% (lower than 95% plan assumption), subscription tokens scale down symmetrically
+    // Metered quota: the vendor delivers the same tokens no matter what the
+    // user's own cache hit rate is (VULN-02 regression guard).
    const result75 = computeApplesToApples([testModel], [testPlan], 'all', 100, 0, 0.75);
    const sub75 = result75.options.find(o => o.type === 'subscription');
-
-    // When cacheRate is 0% (fresh tokens, no cache), subscription tokens scale down even further
    const result0 = computeApplesToApples([testModel], [testPlan], 'all', 100, 0, 0.0);
    const sub0 = result0.options.find(o => o.type === 'subscription');

    expect(sub75).toBeDefined();
    expect(sub0).toBeDefined();
-    // 0% cache delivers fewer tokens than 75% cache for the subscription
+    expect(sub75!.monthlyTokens).toBe(1000);
+    expect(sub0!.monthlyTokens).toBe(1000);
+  });
+
+  it('cache-scales dollar-credit pools whose yield depends on per-turn cost', () => {
+    const testModel = model({
+      id: 'z-ai/glm-5.3',
+      name: 'GLM-5.3',
+      pricing: { input: 1.0, output: 2.0, cachedInput: 0.1, cachedInputWrite: null, reasoning: null, webSearch: null },
+      blendedCost: 1.25,
+      benchmarks: { codingIndex: 75, intelligenceIndex: 75, agenticIndex: 75, valueScore: 200 },
+    });
+
+    const creditTier = tier({
+      name: 'Pool',
+      monthlyPrice: 80,
+      limits: { monthlyCredits: '$80/mo usage value credits' },
+      models: ['GLM-5.3'],
+      estimatedTokenBudget: { estimatedMillionTokens: 1000, assumptions: '95% cache hit rate' },
+    });
+    const creditPlan = plan({ id: 'credit-test', name: 'CreditPool', tiers: [creditTier] });
+
+    const result0 = computeApplesToApples([testModel], [creditPlan], 'all', 100, 0, 0.0);
+    const result75 = computeApplesToApples([testModel], [creditPlan], 'all', 100, 0, 0.75);
+    const sub0 = result0.options.find(o => o.type === 'subscription');
+    const sub75 = result75.options.find(o => o.type === 'subscription');
+
+    expect(sub0).toBeDefined();
+    expect(sub75).toBeDefined();
+    // A dollar pool buys fewer tokens when turns are priced without cache discounts.
    expect(sub0!.monthlyTokens).toBeLessThan(sub75!.monthlyTokens);
-    // 75% cache delivers fewer tokens than the 95% marketing estimate (1000M)
    expect(sub75!.monthlyTokens).toBeLessThan(1000);
  });

-  it('parsePlanCacheAssumption correctly parses declared cache hit percentages', () => {
-    const tier95 = tier({
+  it('parsePlanCacheAssumption ignores discount percentages and anchors on cache context', () => {
+    const discountFirst = tier({
      estimatedTokenBudget: {
        estimatedMillionTokens: 100,
-        assumptions: '95% cache hit rate',
-        estimateMeta: { cacheAssumption: '95% cache hit rate', sourceUrl: 'test', sourceType: 'official', confidence: 'high', verifiedAt: '2026-09-19' },
+        assumptions: 'At 90% discount, 75% cache hit rate',
+        estimateMeta: { cacheAssumption: 'At 90% discount, 75% cache hit rate', sourceUrl: 'test', sourceType: 'official', confidence: 'high', verifiedAt: '2026-09-19' },
      },
    });
-    expect(parsePlanCacheAssumption(tier95)).toBe(0.95);
+    expect(parsePlanCacheAssumption(discountFirst)).toBe(0.75);

-    const tier75 = tier({
+    const discountOnly = tier({
      estimatedTokenBudget: {
        estimatedMillionTokens: 100,
-        assumptions: '75% cache hit rate',
+        assumptions: '50% off-peak discount applies',
      },
    });
-    expect(parsePlanCacheAssumption(tier75)).toBe(0.75);
+    // No cache-context percent: fall back to the documented default, not 0.5.
+    expect(parsePlanCacheAssumption(discountOnly)).toBe(0.75);
+  });

-    const tierDefault = tier();
-    expect(parsePlanCacheAssumption(tierDefault)).toBe(0.75);
+  it('discloses spend asymmetry when the API leg outspends a capped subscription in the arbitrage callout', () => {
+    const sonnetModel = model({
+      id: 'anthropic/claude-sonnet-5',
+      name: 'Claude Sonnet 5',
+      benchmarks: { codingIndex: 85, intelligenceIndex: null, agenticIndex: null, valueScore: null },
+    });
+    const cheapSub = plan({
+      id: 'cheap-sub',
+      name: 'Cheap Sub',
+      tiers: [tier({ monthlyPrice: 20, models: ['Claude Sonnet 5'], estimatedTokenBudget: { estimatedMillionTokens: 1, assumptions: 'test' } })],
+    });
+
+    const result = computeApplesToApples([sonnetModel], [cheapSub], 'all', 200, 0, 0.75);
+    expect(result.arbitrageCallout).not.toBeNull();
+    expect(result.arbitrageCallout!.description).toContain('spends $20/mo vs $200/mo on the API leg');
  });
 });

@@ -89,7 +89,9 @@ export function calculateBudgetResults(
      // yields keep the legacy 3K (2K-in/1K-out) request definition.
      const requests1k = blendMode === 'agentic'
        ? (millionTokens * 1_000_000) / STANDARD_AGENT_REQUEST_TOKENS
-        : (budget / (m.costPer1kRequests || 1)) * 1000;
+        // Nullish coalescing (never `||`): a legitimate $0/M request cost must
+        // not be replaced by the fallback divisor (invariant I).
+        : (budget / (m.costPer1kRequests ?? 1)) * 1000;

      return {
        modelId: m.id,
@@ -265,20 +267,43 @@ export function matchesPlanModel(planModelName: string, model: NormalizedModel):
 */
 export function parsePlanCacheAssumption(tier: PlanTier): number {
  const metaStr = tier.estimatedTokenBudget?.estimateMeta?.cacheAssumption;
+  // Anchor the percent to explicit cache context so discount percentages
+  // ("90% discount", "50% off-peak") cannot hijack the cache rate.
  if (metaStr) {
-    const match = metaStr.match(/([0-9]+)%/);
-    if (match) {
-      return parseInt(match[1], 10) / 100;
+    const anchored = metaStr.match(/([0-9]+)\s*%\s*(?:cache|hit)/i) || metaStr.match(/cache[^.]*?([0-9]+)\s*%/i);
+    if (anchored) {
+      return clampCacheAssumption(parseInt(anchored[1], 10) / 100);
    }
  }
  const assumptions = tier.estimatedTokenBudget?.assumptions;
  if (assumptions) {
-    const match = assumptions.match(/([0-9]+)%\s*cache/i);
-    if (match) {
-      return parseInt(match[1], 10) / 100;
+    const anchored = assumptions.match(/([0-9]+)\s*%\s*(?:cache|hit)/i) || assumptions.match(/cache[^.]*?([0-9]+)\s*%/i);
+    if (anchored) {
+      return clampCacheAssumption(parseInt(anchored[1], 10) / 100);
    }
  }
-  return DEFAULT_CACHE_RATE;
+  return clampCacheAssumption(DEFAULT_CACHE_RATE);
+}
+
+/**
+ * Plans whose tiers meter a fixed dollar-credit pool scale their token yield
+ * with the effective cost of the drained model (spend X credits -> fewer
+ * tokens on pricier turns). Usage/request-metered quotas grant the same
+ * tokens regardless of the user's cache hit rate, so they must NOT be scaled.
+ */
+export function isDollarPoolTier(tier: PlanTier): boolean {
+  if (tier.perModelTokenBudgets) {
+    const bases = Object.values(tier.perModelTokenBudgets).map(e => e?.basis);
+    if (bases.some(b => b === 'list-price-credit' || b === 'equal-rate')) return true;
+  }
+  const limits = tier.limits ?? {};
+  return Object.keys(limits).some(k => k.toLowerCase().includes('credit'));
+}
+
+function clampCacheAssumption(rate: number): number {
+  if (!Number.isFinite(rate) || rate < 0) return DEFAULT_CACHE_RATE;
+  // Deep-agent workloads rarely sustain > 95% cache hits; treat anything above as data noise.
+  return Math.min(rate, 0.95);
 }

 export function calculateAgentRequestCost(
@@ -430,11 +455,18 @@ export function computeApplesToApples(
      const resolved = resolveTierModelBudget(tier, model.name, model.id, planModelName);
      const rawBaseTokens = resolved.tokens || 0;

-      // Symmetric cache scaling: dynamically adjust subscription capacity with active cacheRate
+      // Cache scaling only applies to dollar-credit pools: the same monthly
+      // spend buys fewer tokens when turns are priced at lower cache-hit
+      // rates. Usage/request-metered quotas deliver fixed tokens regardless
+      // of the user's cache rate, so they are left unscaled (VULN-02).
+      const cacheScaleApplies = isDollarPoolTier(tier);
      const planAssumedCache = parsePlanCacheAssumption(tier);
-      const turnCostAssumed = calculateAgentRequestCost(model, planAssumedCache as CacheRate).costPerRequest;
-      const turnCostUser = calculateAgentRequestCost(model, cacheRate).costPerRequest;
-      const cacheScaleFactor = (turnCostAssumed > 0 && turnCostUser > 0) ? (turnCostAssumed / turnCostUser) : 1.0;
+      let cacheScaleFactor = 1.0;
+      if (cacheScaleApplies) {
+        const turnCostAssumed = calculateAgentRequestCost(model, planAssumedCache as CacheRate).costPerRequest;
+        const turnCostUser = calculateAgentRequestCost(model, cacheRate).costPerRequest;
+        cacheScaleFactor = (turnCostAssumed > 0 && turnCostUser > 0) ? (turnCostAssumed / turnCostUser) : 1.0;
+      }

      const baseTokens = Math.round(rawBaseTokens * cacheScaleFactor);
      const rawRequests = tierRawRequests(tier, baseTokens);
@@ -499,8 +531,18 @@ export function computeApplesToApples(
        if (tier.monthlyPrice > budget) continue;

        const rawBaseTokens = tier.estimatedTokenBudget?.estimatedMillionTokens || 0;
+        // Same dollar-pool gate as matched tiers so a plan never scales in
+        // opposite directions depending on model-name matching (VULN-08).
        const planAssumedCache = parsePlanCacheAssumption(tier);
-        const cacheScaleFactor = (1 - planAssumedCache * 0.8) / (1 - cacheRate * 0.8);
+        let cacheScaleFactor = 1.0;
+        if (isDollarPoolTier(tier)) {
+          const basisModel = cleanModels.find(m => detectModelLab(m) === (detectPlanLabs(plan, tier)[0] ?? 'all')) ?? cleanModels[0];
+          if (basisModel) {
+            const turnCostAssumed = calculateAgentRequestCost(basisModel, planAssumedCache as CacheRate).costPerRequest;
+            const turnCostUser = calculateAgentRequestCost(basisModel, cacheRate).costPerRequest;
+            cacheScaleFactor = (turnCostAssumed > 0 && turnCostUser > 0) ? (turnCostAssumed / turnCostUser) : 1.0;
+          }
+        }
        const baseTokens = Math.round(rawBaseTokens * (cacheScaleFactor > 0 ? cacheScaleFactor : 1.0));
        const rawRequests = tierRawRequests(tier, baseTokens);
        const vendorQuota = parseTierRequestLimit(tier.limits);
@@ -610,13 +652,26 @@ export function computeApplesToApples(
        ? `nearest-quality API match (CI ±${qualityGap.toFixed(1)})`
        : 'nearest available API match';

+      // Spend parity: a single-seat subscription spends only its tier price
+      // while the API leg spends the full budget. Surface the asymmetry
+      // whenever the API leg spends meaningfully more (VULN-09).
+      const subSpend = bestSubscription.monthlyCost;
+      const apiSpend = apiOption.monthlyCost;
+      const spendAsymmetry = apiSpend > subSpend * 1.25
+        ? ` Subscription spends $${subSpend.toFixed(0)}/mo vs $${apiSpend.toFixed(0)}/mo on the API leg${
+            isFinite(apiSpend / (subSpend || 1))
+              ? ` (${(apiSpend / (subSpend || 1)).toFixed(1)}x the spend — use Dave Mode or Mix & Match to equalize)`
+              : ''
+          }`
+        : '';
+
      if (subTokens >= apiTokens * 1.25) {
        const multiplier = Number((subTokens / (apiTokens || 1)).toFixed(1));
        arbitrageCallout = {
          winnerType: 'subscription',
          multiplier,
          headline: `Coding Subscription Wins for ${labName} (${multiplier}x Compute)`,
-          description: `${bestSubscription.planName} (${bestSubscription.tierName}) running ${bestSubscription.name} yields ~${formatMillionTokens(subTokens)} normalized tokens (~${bestSubscription.monthlyRequests.toLocaleString()} requests), beating direct ${apiOption.name} API (~${formatMillionTokens(apiTokens)} tokens) for $${budget}/mo. Compared as ${qualityNote}.`,
+          description: `${bestSubscription.planName} (${bestSubscription.tierName}) running ${bestSubscription.name} yields ~${formatMillionTokens(subTokens)} normalized tokens (~${bestSubscription.monthlyRequests.toLocaleString()} requests), beating direct ${apiOption.name} API (~${formatMillionTokens(apiTokens)} tokens) for $${budget}/mo. Compared as ${qualityNote}.${spendAsymmetry}`,
        };
      } else if (apiTokens >= subTokens * 1.25) {
        const multiplier = Number((apiTokens / (subTokens || 1)).toFixed(1));
@@ -624,14 +679,14 @@ export function computeApplesToApples(
          winnerType: 'api',
          multiplier,
          headline: `Direct API Wins for ${labName} (${multiplier}x Compute)`,
-          description: `Direct pay-per-token API for ${apiOption.name} delivers ~${formatMillionTokens(apiTokens)} tokens (~${apiOption.monthlyRequests.toLocaleString()} requests), outperforming coding subscriptions (~${formatMillionTokens(subTokens)} normalized tokens) for $${budget}/mo. Compared as ${qualityNote}.`,
+          description: `Direct pay-per-token API for ${apiOption.name} delivers ~${formatMillionTokens(apiTokens)} tokens (~${apiOption.monthlyRequests.toLocaleString()} requests), outperforming coding subscriptions (~${formatMillionTokens(subTokens)} normalized tokens) for $${budget}/mo. Compared as ${qualityNote}.${spendAsymmetry}`,
        };
      } else {
        arbitrageCallout = {
          winnerType: 'even',
          multiplier: 1.0,
          headline: `Balanced Value for ${labName}`,
-          description: `Both direct API (${apiOption.name}) and subscriptions (${bestSubscription.planName} running ${bestSubscription.name}) offer comparable compute near ${formatMillionTokens(subTokens)} tokens for $${budget}/mo. Compared as ${qualityNote}.`,
+          description: `Both direct API (${apiOption.name}) and subscriptions (${bestSubscription.planName} running ${bestSubscription.name}) offer comparable compute near ${formatMillionTokens(subTokens)} tokens for $${budget}/mo. Compared as ${qualityNote}.${spendAsymmetry}`,
        };
      }
    }
@@ -686,11 +741,13 @@ export function parseTierRequestLimit(limits?: Record<string, unknown>): number
        return Math.round(parseFloat(kMixMatch[1]) * 1000);
      }

-      // 3. Weekly requests scaled x4: e.g. "12,000/week", "1,000 requests per 7-day"
+      // 3. Weekly requests scaled to a monthly equivalent (52/12 weeks per
+      // month). Weekly quotas usually coexist with undocumented monthly
+      // ceilings, so scale conservatively (VULN-12).
      const weekMatch = val.match(/(?:≈|~)?\s*([0-9]+(?:,[0-9]+)*(?:\.[0-9]+)?)\s*(?:agentic\s*)?requests?\s*(?:\/|\s*per\s*)\s*(?:week|7-day)/i) ||
                        val.match(/([0-9]+(?:,[0-9]+)*)\s*\/\s*week/i);
      if (weekMatch) {
-        return Math.round(parseFloat(weekMatch[1].replace(/,/g, '')) * 4);
+        return Math.round(parseFloat(weekMatch[1].replace(/,/g, '')) * (52 / 12));
      }

      // 4. Simple leading count: e.g. "50 agentic requests/mo", "50 requests"
@@ -136,3 +136,56 @@ describe('Extended Thinking & Reasoning Token Engine', () => {
    expect(anthropicApi?.absorptionScore).toBeLessThanOrEqual(40);
  });
 });
+
+describe('Reasoning feasibility guard (VULN-06)', () => {
+  const baseModel: NormalizedModel = {
+    id: 'anthropic/claude-sonnet-5',
+    name: 'Claude Sonnet 5',
+    provider: 'anthropic',
+    series: 'Claude',
+    modality: 'text->text',
+    contextWindow: 200000,
+    maxOutput: 64000,
+    pricing: { input: 3.0, output: 15.0, cachedInput: 0.3, cachedInputWrite: 3.75, reasoning: null, webSearch: null },
+    blendedCost: 6.0,
+    agentBlendedCost: 1.5,
+    costPer1kRequests: 30.0,
+    benchmarks: { intelligenceIndex: 88, codingIndex: 85, agenticIndex: 84, valueScore: 78 },
+    reasoning: null,
+    isFree: false,
+    isBatch: false,
+  };
+
+  const tinyModel: NormalizedModel = {
+    ...baseModel,
+    id: 'tiny/small-model',
+    name: 'Tiny Small Model',
+    contextWindow: 32000,
+    maxOutput: 8000,
+  };
+
+  it('rejects max effort on a 32K-context model as an impossible combination', () => {
+    const turn = calculateReasoningCost(tinyModel, 0.75, 'max');
+    expect(turn.feasible).toBe(false);
+    expect(turn.infeasibilityReason).toContain('32,000');
+  });
+
+  it('rejects reasoning that exceeds the max completion cap even when context fits', () => {
+    const cappedModel: NormalizedModel = { ...baseModel, contextWindow: 200000, maxOutput: 8000 };
+    const turn = calculateReasoningCost(cappedModel, 0.75, 'high');
+    expect(turn.feasible).toBe(false);
+    expect(turn.infeasibilityReason).toContain('caps output');
+  });
+
+  it('keeps medium effort feasible on the frontier model', () => {
+    const turn = calculateReasoningCost(baseModel, 0.75, 'medium');
+    expect(turn.feasible).toBe(true);
+    expect(turn.infeasibilityReason).toBeNull();
+  });
+
+  it('treats unknown windows (0) as feasible rather than fabricating a constraint', () => {
+    const unknown = { ...baseModel, contextWindow: 0, maxOutput: 0 };
+    const turn = calculateReasoningCost(unknown, 0.75, 'max');
+    expect(turn.feasible).toBe(true);
+  });
+});
@@ -56,6 +56,10 @@ export interface ReasoningTurnCost {
  totalCost: number;
  reasoningTokens: number;
  inflationMultiplier: number; // totalCost / baseCost
+  /** True when input + reasoning + output fit the model's context and output caps. */
+  feasible: boolean;
+  /** Human-readable reason when infeasible, else null. */
+  infeasibilityReason: string | null;
 }

 export interface MonthlyReasoningWorkload {
@@ -214,6 +218,22 @@ export function calculateReasoningCost(
  const totalCost = baseCost + reasoningCost;
  const inflationMultiplier = baseCost > 0 ? totalCost / baseCost : 1.0;

+  // Feasibility guard (VULN-06): input + reasoning + output must fit inside
+  // the model's context window, and reasoning + output must fit its max
+  // completion cap. Unknown caps (0) are treated as feasible.
+  let feasible = true;
+  let infeasibilityReason: string | null = null;
+  const requiredContext = AGENT_REQUEST_INPUT_TOKENS + reasoningTokens + AGENT_REQUEST_OUTPUT_TOKENS;
+  if (model.contextWindow > 0 && requiredContext > model.contextWindow) {
+    feasible = false;
+    infeasibilityReason = `${effortSpec.label} reasoning needs ${requiredContext.toLocaleString()} context tokens but ${model.name} only supports ${model.contextWindow.toLocaleString()}.`;
+  }
+  const requiredOutput = reasoningTokens + AGENT_REQUEST_OUTPUT_TOKENS;
+  if (feasible && model.maxOutput > 0 && requiredOutput > model.maxOutput) {
+    feasible = false;
+    infeasibilityReason = `${effortSpec.label} reasoning generates ${requiredOutput.toLocaleString()} completion tokens but ${model.name} caps output at ${model.maxOutput.toLocaleString()}.`;
+  }
+
  return {
    baseCost,
    inputCost,
@@ -222,6 +242,8 @@ export function calculateReasoningCost(
    totalCost,
    reasoningTokens,
    inflationMultiplier: Number(inflationMultiplier.toFixed(2)),
+    feasible,
+    infeasibilityReason,
  };
 }

@@ -38,6 +38,11 @@ export interface SessionReceiptReport {
  modelRepricings: ModelRepricingItem[];
  subscriptionImpacts: SubscriptionImpactItem[];
  headlineSummary: string;
+  /**
+   * False when the parser synthesized the token counts (VULN-05): the receipt
+   * must be presented as an estimate, not a measured bill.
+   */
+  tokensMeasured: boolean;
 }

 /**
@@ -173,9 +178,11 @@ export function calculateSessionReceipt(
  ];

  const cheapest = modelRepricings[0];
-  const headlineSummary = cheapest && cheapest.sessionCost < baselineTotalCost
+  const tokensMeasured = session.measuredTokens !== false;
+  const estimatePrefix = tokensMeasured ? '' : 'ESTIMATED (transcript contained no token accounting — parser assumed ~21K in / 900 out per turn at 75% cache). ';
+  const headlineSummary = estimatePrefix + (cheapest && cheapest.sessionCost < baselineTotalCost
    ? `This session cost $${baselineTotalCost.toFixed(2)} on ${baselineModel?.name || 'Sonnet 5'}. It would have cost $${cheapest.sessionCost.toFixed(2)} on ${cheapest.modelName} (${cheapest.savingsPercentVsBaseline}% cheaper).`
-    : `This session executed ${session.totalTurns} turns across ${(session.totalTokens / 1e6).toFixed(2)}M tokens at a ${Math.round(session.effectiveCacheHitRate * 100)}% prompt cache hit rate.`;
+    : `This session executed ${session.totalTurns} turns across ${(session.totalTokens / 1e6).toFixed(2)}M tokens at a ${Math.round(session.effectiveCacheHitRate * 100)}% prompt cache hit rate.`);

  return {
    session,
@@ -185,5 +192,6 @@ export function calculateSessionReceipt(
    modelRepricings,
    subscriptionImpacts,
    headlineSummary,
+    tokensMeasured,
  };
 }
@@ -175,7 +175,10 @@ function computePersonaApiCost(p: DeveloperPersona, model?: NormalizedModel): nu
  if (!model) return p.estimatedDirectApiCost;
  const inPrice = model.pricing.input ?? model.blendedCost * 0.75;
  const outPrice = model.pricing.output ?? model.blendedCost * 1.75;
-  const cachedPrice = model.pricing.cachedInput ?? inPrice * 0.1;
+  // Never fabricate a discount for an unknown cache price (invariant I):
+  // an unknown cache rate is priced at full input, matching
+  // getEffectiveCacheMultiplier's "no caching" convention.
+  const cachedPrice = model.pricing.cachedInput ?? inPrice;
  const { costPerRequest } = calculateAgentRequestCost(model, 0.75);

  const agentCost = p.monthlyAgentTurns * costPerRequest;
@@ -1,5 +1,6 @@
 export type ThrottleExhaustionBehavior = 'hard-block' | 'slow-queue' | 'payg-overage';
-export type ThrottleStatus = 'smooth' | 'queued' | 'blocked';
+/** 'overage' = served beyond quota as paid overage (payg-overage profiles, VULN-03). */
+export type ThrottleStatus = 'smooth' | 'queued' | 'overage' | 'blocked';

 export interface ThrottleProfile {
  id: string;
@@ -136,3 +136,36 @@ describe('Throttle Simulation Engine', () => {
    }
  });
 });
+
+describe('Payg-overage simulation (VULN-03)', () => {
+  const getProfile = (id: string): ThrottleProfile => {
+    const found = THROTTLE_PROFILES.find((p) => p.id === id);
+    if (!found) throw new Error(`Profile ${id} not found`);
+    return found;
+  };
+
+  it('flags Google AI Pro as paid overage instead of infinitely smooth under saturating load', () => {
+    const googlePro = getProfile('google-ai-pro');
+    // 8 agents at rapid cadence burn through the 90-turn / 5h window fast.
+    const result = simulateSprintThrottle(googlePro, {
+      concurrency: 8,
+      sprintDurationHours: 4,
+      turnPace: 'rapid',
+    });
+
+    expect(result.status).toBe('overage');
+    expect(result.overageTurns).toBeGreaterThan(0);
+    expect(result.throttleReason).not.toBeNull();
+  });
+
+  it('never shows overage for plans that simply survive the sprint', () => {
+    const claudeCode = getProfile('claude-code-pro');
+    const result = simulateSprintThrottle(claudeCode, {
+      concurrency: 1,
+      sprintDurationHours: 2,
+      turnPace: 'deep',
+    });
+    expect(result.status).toBe('smooth');
+    expect(result.overageTurns).toBe(0);
+  });
+});
@@ -65,6 +65,8 @@ export interface SimulationResult {
  fastTurnsCompleted: number;
  slowTurnsCompleted: number;
  blockedTurns: number;
+  /** Turns billed as paid overage when a payg-overage profile crossed its quota ceiling. */
+  overageTurns: number;
  effectiveThroughputPercent: number;
  headroomScore: number; // 0 to 100
  timeline: SimulationTimelinePoint[];
@@ -78,6 +80,9 @@ function formatMinutesToHours(minutes: number): string {
  return `${h}h ${m}m`;
 }

+/** Severity rank so the sprint's overall status always reflects the worst outcome. */
+const STATUS_SEVERITY: Record<ThrottleStatus, number> = { smooth: 0, queued: 1, overage: 2, blocked: 3 };
+
 /**
 * Runs a discrete time-step simulation (step = 5 minutes) of an intensive coding sprint
 * against a given throttle profile.
@@ -118,6 +123,11 @@ export function simulateSprintThrottle(
  let slowTurnsCompleted = 0;
  let blockedTurns = 0;
  let cumulativeTurns = 0;
+  // Deferred turns from prior queue delays (VULN-12): a queued step cannot
+  // magically complete its full turn allocation; the backlog spills into
+  // later steps and reduces throughput realistically.
+  let deferredTurns = 0;
+  let overageTurns = 0;

  const timeline: SimulationTimelinePoint[] = [
    {
@@ -134,10 +144,28 @@ export function simulateSprintThrottle(
  const hasConcurrencyViolation = safeConcurrency > profile.maxConcurrency;

  for (let currentMinute = stepMinutes; currentMinute <= totalMinutes; currentMinute += stepMinutes) {
-    const turnsAttempted = turnsPerStep;
+    // Queue-delay backlog reduces the turns actually executable this step.
+    // A 30s/turn slow-queue delay over a 5-minute step caps the step at
+    // floor(step / delay) full-speed-equivalent allocations; the deficit is
+    // deferred, not silently completed (VULN-12).
+    let turnsAttempted = turnsPerStep;
    let stepStatus: ThrottleStatus = 'smooth';
    let stepDelay = 0;

+    if (deferredTurns > 0) {
+      if (profile.exhaustionBehavior === 'hard-block') {
+        // Blocked turns never recover; drop the backlog.
+        deferredTurns = 0;
+      } else if (profile.slowQueueDelaySec > 0) {
+        const executableRatio = Math.min(1, (stepMinutes * 60) / (profile.slowQueueDelaySec * turnsPerStepPerAgent * safeConcurrency * safeConcurrency));
+        const executable = Math.max(1, Math.floor(turnsPerStep * executableRatio));
+        turnsAttempted = Math.min(turnsPerStep, executable);
+        deferredTurns = Math.max(0, deferredTurns - (turnsPerStep - turnsAttempted));
+        stepStatus = 'queued';
+        stepDelay = profile.slowQueueDelaySec;
+      }
+    }
+
    // 1. Check instantaneous concurrency
    if (hasConcurrencyViolation) {
      if (profile.exhaustionBehavior === 'hard-block') {
@@ -176,6 +204,12 @@ export function simulateSprintThrottle(
          if (!firstThrottleReason) {
            firstThrottleReason = `Rolling ${profile.rollingWindowHours}h window exceeded: degraded to slow queue`;
          }
+        } else if (profile.exhaustionBehavior === 'payg-overage') {
+          // PAYG profiles never block; excess turns bill as paid overage (VULN-03).
+          stepStatus = 'overage';
+          if (!firstThrottleReason) {
+            firstThrottleReason = `Rolling ${profile.rollingWindowHours}h window exceeded: excess turns continue as paid overage`;
+          }
        }
      }
    }
@@ -194,6 +228,11 @@ export function simulateSprintThrottle(
          if (!firstThrottleReason) {
            firstThrottleReason = `Monthly fast pool exhausted (${profile.monthlyFastRequests} reqs): sprint blocked`;
          }
+        } else if (profile.exhaustionBehavior === 'payg-overage') {
+          if (!firstThrottleReason) {
+            firstThrottleReason = `Monthly pool exhausted (${profile.monthlyFastRequests} reqs): excess turns continue as paid overage`;
+          }
+          stepStatus = 'overage';
        }
      }
    }
@@ -201,25 +240,53 @@ export function simulateSprintThrottle(
    // Process turns for this step
    if (stepStatus === 'blocked') {
      blockedTurns += turnsAttempted;
-      if (firstThrottleMinute === null) {
-        firstThrottleMinute = currentMinute;
+      if (STATUS_SEVERITY['blocked'] > STATUS_SEVERITY[overallStatus]) {
        overallStatus = 'blocked';
      }
-    } else if (stepStatus === 'queued') {
-      slowTurnsCompleted += turnsAttempted;
+      if (firstThrottleMinute === null) {
+        firstThrottleMinute = currentMinute;
+      }
+    } else if (stepStatus === 'overage') {
+      // Served beyond quota — turns complete but bill as paid overage (VULN-03).
+      overageTurns += turnsAttempted;
      cumulativeTurns += turnsAttempted;
      rollingWindowHistory.push({ minute: currentMinute, turns: turnsAttempted });
      if (remainingMonthlyFast !== null) {
        remainingMonthlyFast = Math.max(0, remainingMonthlyFast - turnsAttempted);
      }
+      if (STATUS_SEVERITY['overage'] > STATUS_SEVERITY[overallStatus]) {
+        overallStatus = 'overage';
+      }
+      if (firstThrottleMinute === null) {
+        firstThrottleMinute = currentMinute;
+      }
+    } else if (stepStatus === 'queued') {
+      // Queue degradation defers the un-executable share of this step's turns.
+      const deferred = Math.max(0, turnsAttempted - Math.max(1, Math.floor(turnsPerStep * 0.7)));
+      deferredTurns += deferred;
+      slowTurnsCompleted += turnsAttempted;
+      cumulativeTurns += turnsAttempted;
+      if (profile.exhaustionBehavior === 'payg-overage') {
+        overageTurns += turnsAttempted;
+      }
+      rollingWindowHistory.push({ minute: currentMinute, turns: turnsAttempted });
+      if (remainingMonthlyFast !== null) {
+        remainingMonthlyFast = Math.max(0, remainingMonthlyFast - turnsAttempted);
+      }
+      if (STATUS_SEVERITY['queued'] > STATUS_SEVERITY[overallStatus]) {
+        overallStatus = 'queued';
+      }
      if (firstThrottleMinute === null) {
        firstThrottleMinute = currentMinute;
-        overallStatus = overallStatus === 'blocked' ? 'blocked' : 'queued';
      }
    } else {
      // Smooth step
      fastTurnsCompleted += turnsAttempted;
      cumulativeTurns += turnsAttempted;
+      if (profile.exhaustionBehavior === 'payg-overage' && firstThrottleReason !== null) {
+        // Once a PAYG ceiling was crossed, all subsequent turns bill as overage.
+        overageTurns += turnsAttempted;
+      }
      rollingWindowHistory.push({ minute: currentMinute, turns: turnsAttempted });
      if (remainingMonthlyFast !== null) {
        remainingMonthlyFast = Math.max(0, remainingMonthlyFast - turnsAttempted);
@@ -278,6 +345,7 @@ export function simulateSprintThrottle(
    blockedTurns: Math.round(blockedTurns),
    effectiveThroughputPercent: Math.min(100, Math.max(0, effectiveThroughputPercent)),
    headroomScore: Math.min(100, Math.max(0, headroomScore)),
+    overageTurns: Math.round(overageTurns),
    timeline,
  };
 }
@@ -289,12 +357,12 @@ export function simulateAllProfiles(params: SprintSimulationParams): SimulationR
  const results = THROTTLE_PROFILES.map((profile) => simulateSprintThrottle(profile, params));

  // Sort order:
-  // 1. Smooth before Queued before Blocked
+  // 1. Smooth before Queued before Overage before Blocked
  // 2. Higher headroom score
  // 3. Longer survival hours
  // 4. Lower monthly price
  return results.sort((a, b) => {
-    const statusWeight = { smooth: 3, queued: 2, blocked: 1 };
+    const statusWeight = { smooth: 4, queued: 3, overage: 2, blocked: 1 };
    if (statusWeight[a.status] !== statusWeight[b.status]) {
      return statusWeight[b.status] - statusWeight[a.status];
    }
@@ -58,16 +58,21 @@ describe('classifyTraining', () => {
 });

 describe('getIndemnityInfo', () => {
-  it('labels boolean true as full indemnity', () => {
-    expect(getIndemnityInfo(plan({ ipIndemnity: true })).label).toBe('Full Indemnity');
+  it('never labels boolean true as unconditional "Full Indemnity" (VULN-10)', () => {
+    const info = getIndemnityInfo(plan({ ipIndemnity: true }));
+    expect(info.label).toBe('Indemnity Offered');
+    expect(info.tone).toBe('warning');
+    expect(info.description).toContain('carve-outs');
  });

  it('labels z-ai-style strings as enterprise only', () => {
    expect(getIndemnityInfo(plan({ ipIndemnity: 'Enterprise Team contracts only' })).label).toBe('Enterprise Only');
  });

-  it('labels an unknown string as unclear rather than none', () => {
-    expect(getIndemnityInfo(plan({ ipIndemnity: 'Full indemnity on all paid plans' })).label).toBe('Full Indemnity');
+  it('flags indemnity strings with a verify-scope caveat rather than promising full coverage', () => {
+    const info = getIndemnityInfo(plan({ ipIndemnity: 'Full indemnity on all paid plans' }));
+    expect(info.label).toBe('Indemnity Stated (verify scope)');
+    expect(info.tone).toBe('warning');
  });
 });

@@ -48,7 +48,8 @@ const TRAINS_PHRASES = [
  'used for model improvement',
  'interactions are recorded and used',
  'may be used for model improvement',
-  'no guarantee',
+  // 'no guarantee' removed (VULN-10): hedges about uptime/confidentiality are
+  // not evidence that a vendor trains on customer data.
 ];

 const OPT_OUT_PHRASES = ['opt-out', 'opt out', 'privacy mode', 'training opt-out', 'in-settings opt-out'];
@@ -152,9 +153,13 @@ export function getIndemnityInfo(plan: CodingPlan): IndemnityInfo {
  const ip = plan.ipIndemnity;
  if (ip === true) {
    return {
-      label: 'Full Indemnity',
-      tone: 'success',
-      description: 'Full legal IP indemnification covered for copyright infringement on generated code.',
+      // VULN-10: a boolean cannot establish unconditional coverage. Real
+      // indemnities carry carve-outs (modification, combination, breach of
+      // usage policy) and liability caps — never present this as "Full".
+      label: 'Indemnity Offered',
+      tone: 'warning',
+      description:
+        'Vendor advertises IP indemnification. Scope, carve-outs and liability caps are defined in the vendor’s legal terms — verify before relying on this for compliance.',
    };
  }
  if (typeof ip === 'string') {
@@ -163,7 +168,11 @@ export function getIndemnityInfo(plan: CodingPlan): IndemnityInfo {
      return { label: 'Enterprise Only', tone: 'warning', description: ip };
    }
    if (lower.includes('full') || lower.includes('all plans')) {
-      return { label: 'Full Indemnity', tone: 'success', description: ip };
+      return {
+        label: 'Indemnity Stated (verify scope)',
+        tone: 'warning',
+        description: ip,
+      };
    }
    return { label: 'Unclear', tone: 'muted', description: ip };
  }
@@ -182,7 +191,10 @@ export interface GotchaClassification {
 }

 const RESTRICTION_WORDS = ['ban', 'banned', 'suspend', 'suspension', 'bulk', 'multiple accounts', 'multi-account', 'account sharing', 'unauthorized', 'tool-only', 'lock-in', 'error 1113', 'circumvent', 'breach', 'prohibited'];
-const PRIVACY_WORDS = ['train', 'training', 'privacy', 'telemetry', 'human review', 'data retention', 'zero retention'];
+/** Strong privacy signals that warrant CRITICAL severity (VULN-10). */
+const PRIVACY_STRONG_WORDS = ['train on', 'trains on', 'trained on', 'used for training', 'used to train', 'human review', 'data retention', 'zero retention', 'not used for training'];
+/** Weak privacy mentions (e.g. "training" inside a caching sentence) only justify a warning. */
+const PRIVACY_WEAK_WORDS = ['train', 'training', 'privacy', 'telemetry'];
 const LEGAL_WORDS = ['indemnity', 'indemnification', 'copyright', 'liability', 'ip ', 'legal'];
 const BILLING_WORDS = ['bill', 'billing', 'fee', 'fees', 'charge', 'overage', 'refund', 'refunds', 'cost', 'expire', 'expires', 'rate', '$', 'price', 'non-refundable'];
 const LIMIT_WORDS = ['limit', 'limits', 'throttle', 'cooldown', 'hard limit', 'quota', 'cap', 'resets', 'reset'];
@@ -200,9 +212,14 @@ export function classifyGotcha(text: string): GotchaClassification {
  if (matchesWord(lower, RESTRICTION_WORDS)) {
    return { category: 'restrictions', severity: 'critical' };
  }
-  if (matchesWord(lower, PRIVACY_WORDS)) {
+  if (matchesWord(lower, PRIVACY_STRONG_WORDS)) {
    return { category: 'data-privacy', severity: 'critical' };
  }
+  if (matchesWord(lower, PRIVACY_WEAK_WORDS)) {
+    // Weak signal: the word "training"/"privacy" appears without an actual
+    // training-on-data claim (e.g. "Privacy mode keeps code out of training").
+    return { category: 'data-privacy', severity: 'warning' };
+  }
  if (matchesWord(lower, LEGAL_WORDS)) {
    return { category: 'ip-legal', severity: 'warning' };
  }
@@ -6,6 +6,7 @@ import {
  type TurnCadence,
 } from '../lib/throttle';
 import type { ThrottleStatus } from '../lib/throttle-profiles';
+import { numParam } from '../lib/params';
 import {
  Zap,
  Gauge,
@@ -20,6 +21,7 @@ import {
  RotateCcw,
  Sparkles,
  Info,
+  DollarSign,
 } from 'lucide-react';
 import {
  ResponsiveContainer,
@@ -74,12 +76,13 @@ const PRESETS: Preset[] = [
 export default function BurstSimulator() {
  const [searchParams, setSearchParams] = useSearchParams();

-  // Read initial parameters from URL or use defaults
-  const initialConcurrency = Math.min(10, Math.max(1, Number(searchParams.get('agents')) || 2));
-  const initialDuration = Math.min(8, Math.max(1, Number(searchParams.get('duration')) || 4));
+  // Read initial parameters from URL or use defaults (zero-safe: `?prior=0`
+  // and `?agents=0`-style values must not collapse to defaults, VULN-13)
+  const initialConcurrency = numParam(searchParams.get('agents'), 2, 1, 10);
+  const initialDuration = numParam(searchParams.get('duration'), 4, 1, 8);
  const rawCadence = searchParams.get('cadence') as TurnCadence;
  const initialCadence: TurnCadence = (rawCadence && CADENCE_DEFINITIONS[rawCadence]) ? rawCadence : 'standard';
-  const initialPrior = Math.min(90, Math.max(0, Number(searchParams.get('prior')) || 20));
+  const initialPrior = numParam(searchParams.get('prior'), 20, 0, 90);

  const [concurrency, setConcurrency] = useState<number>(initialConcurrency);
  const [sprintDuration, setSprintDuration] = useState<number>(initialDuration);
@@ -152,6 +155,7 @@ export default function BurstSimulator() {
  const estimatedTokensM = Number(((totalTurnsAttempted * 21000) / 1000000).toFixed(1));
  const smoothCount = simulationResults.filter((r) => r.status === 'smooth').length;
  const queuedCount = simulationResults.filter((r) => r.status === 'queued').length;
+  const overageCount = simulationResults.filter((r) => r.status === 'overage').length;
  const blockedCount = simulationResults.filter((r) => r.status === 'blocked').length;

  const filteredResults = useMemo(() => {
@@ -489,7 +493,7 @@ export default function BurstSimulator() {
                        <div className="flex justify-between gap-4 capitalize">
                          <span className="text-text-muted">Status:</span>
                          <span className={`font-semibold ${
-                            data.status === 'smooth' ? 'text-emerald-400' : data.status === 'queued' ? 'text-amber-400' : 'text-rose-400'
+                            data.status === 'smooth' ? 'text-emerald-400' : data.status === 'queued' ? 'text-amber-400' : data.status === 'overage' ? 'text-orange-400' : 'text-rose-400'
                          }`}>
                            {data.status}
                          </span>
@@ -509,6 +513,8 @@ export default function BurstSimulator() {
                        ? '#10b981'
                        : entry.status === 'queued'
                        ? '#f59e0b'
+                        : entry.status === 'overage'
+                        ? '#f97316'
                        : '#ef4444'
                    }
                  />
@@ -565,6 +571,17 @@ export default function BurstSimulator() {
              <AlertTriangle className="w-3.5 h-3.5" />
              Queued ({queuedCount})
            </button>
+            <button
+              onClick={() => setStatusFilter('overage')}
+              className={`px-3 py-1 rounded font-medium transition-colors flex items-center gap-1 ${
+                statusFilter === 'overage'
+                  ? 'bg-orange-950/80 text-orange-300 border border-orange-600/40'
+                  : 'text-text-muted hover:bg-surface hover:text-orange-400'
+              }`}
+            >
+              <DollarSign className="w-3.5 h-3.5" />
+              Overage ({overageCount})
+            </button>
            <button
              onClick={() => setStatusFilter('blocked')}
              className={`px-3 py-1 rounded font-medium transition-colors flex items-center gap-1 ${
@@ -583,7 +600,12 @@ export default function BurstSimulator() {
        <div className="divide-y divide-steel-800/80">
          {filteredResults.map((result) => {
            const isSelected = activeDetail?.profile.id === result.profile.id;
-            const statusConfig = {
+            const statusConfigTable: Record<ThrottleStatus, {
+              label: string;
+              badge: string;
+              icon: typeof ShieldCheck;
+              color: string;
+            }> = {
              smooth: {
                label: 'Smooth Sprint',
                badge: 'bg-emerald-950/60 text-emerald-300 border-emerald-600/40',
@@ -596,13 +618,20 @@ export default function BurstSimulator() {
                icon: AlertTriangle,
                color: 'text-amber-400',
              },
+              overage: {
+                label: 'Paid Overage',
+                badge: 'bg-orange-950/60 text-orange-300 border-orange-600/40',
+                icon: DollarSign,
+                color: 'text-orange-400',
+              },
              blocked: {
                label: 'Session Blocked',
                badge: 'bg-rose-950/60 text-rose-300 border-rose-600/40',
                icon: XCircle,
                color: 'text-rose-400',
              },
-            }[result.status];
+            };
+            const statusConfig = statusConfigTable[result.status];

            const StatusIcon = statusConfig.icon;

@@ -654,7 +683,7 @@ export default function BurstSimulator() {
                  </div>

                  {/* Turns Throughput Breakdown */}
-                  <div className="grid grid-cols-3 gap-3 text-center min-w-[200px]">
+                  <div className="grid grid-cols-4 gap-3 text-center min-w-[240px]">
                    <div className="p-1.5 rounded bg-void-950/40 border border-steel-800">
                      <div className="text-[10px] uppercase text-text-muted">Fast</div>
                      <div className="text-xs font-mono font-bold text-emerald-400">
@@ -667,6 +696,12 @@ export default function BurstSimulator() {
                        {result.slowTurnsCompleted}
                      </div>
                    </div>
+                    <div className="p-1.5 rounded bg-void-950/40 border border-steel-800">
+                      <div className="text-[10px] uppercase text-text-muted">Overage</div>
+                      <div className="text-xs font-mono font-bold text-orange-400">
+                        {result.overageTurns}
+                      </div>
+                    </div>
                    <div className="p-1.5 rounded bg-void-950/40 border border-steel-800">
                      <div className="text-[10px] uppercase text-text-muted">Blocked</div>
                      <div className="text-xs font-mono font-bold text-rose-400">
@@ -6,8 +6,10 @@ import { LoadingSpinner } from '../components/shared/LoadingSpinner';
 import {
  HARDWARE_PRESETS,
  computeHardwareEconomics,
+  SALVAGE_CLAMP_MAX,
  type HardwarePreset,
 } from '../lib/hardware';
+import { numParam } from '../lib/params';
 import { formatPrice } from '../lib/pricing';
 import {
  Cpu,
@@ -24,6 +26,7 @@ import {
  Layers,
  Sparkles,
 } from 'lucide-react';
+
 import {
  ResponsiveContainer,
  LineChart,
@@ -43,13 +46,13 @@ export default function HardwareBreakeven() {

  // Read initial params
  const initialPresetId = searchParams.get('preset') || 'mac-mini-m4-pro-64gb';
-  const initialAmortization = Math.min(48, Math.max(6, Number(searchParams.get('amortization')) || 24));
-  const initialSalvage = Math.min(50, Math.max(0, Number(searchParams.get('salvage')) || 20));
-  const initialKwhCost = Math.min(1.0, Math.max(0.01, Number(searchParams.get('kwh')) || 0.16));
-  const initialDailyInference = Math.min(24, Math.max(1, Number(searchParams.get('infHours')) || 3));
-  const initialDailyIdle = Math.min(24, Math.max(0, Number(searchParams.get('idleHours')) || 9));
-  const initialVolumeM = Math.min(100, Math.max(0.5, Number(searchParams.get('volume')) || 6));
-  const initialInputRatio = Math.min(0.95, Math.max(0.5, Number(searchParams.get('inputRatio')) || 0.8));
+  const initialAmortization = numParam(searchParams.get('amortization'), 24, 6, 48);
+  const initialSalvage = numParam(searchParams.get('salvage'), 20, 0, SALVAGE_CLAMP_MAX);
+  const initialKwhCost = numParam(searchParams.get('kwh'), 0.16, 0.01, 1.0);
+  const initialDailyInference = numParam(searchParams.get('infHours'), 3, 1, 24);
+  const initialDailyIdle = numParam(searchParams.get('idleHours'), 9, 0, 24);
+  const initialVolumeM = numParam(searchParams.get('volume'), 6, 0.5, 100);
+  const initialInputRatio = numParam(searchParams.get('inputRatio'), 0.8, 0.5, 0.95);
  const initialApiId = searchParams.get('api') || 'anthropic/claude-3.7-sonnet';
  const initialPlanId = searchParams.get('plan') || 'cursor-pro';

@@ -574,6 +577,28 @@ export default function HardwareBreakeven() {
            </div>
          </div>

+          {/* Workload feasibility warning (VULN-04) */}
+          {economics.infeasibleWorkload && (
+            <div className="p-3 rounded-lg bg-amber-950/50 border border-amber-700/40 text-xs text-amber-200 flex items-start gap-2">
+              <AlertTriangle className="w-4 h-4 text-amber-400 shrink-0 mt-0.5" />
+              <div>
+                <span className="font-bold">Infeasible workload for this duty cycle.</span> Generating{' '}
+                {monthlyTokenVolumeM}M tokens/mo needs ~{economics.requiredInferenceHours}h of pure decode at{' '}
+                {currentPreset.recommendedModels[0]?.tokensPerSec ?? 0} tok/s, but your configured schedule
+                allows at most {economics.feasibleMonthlyVolumeM}M tok/mo ({dailyInferenceHours}h/day × 30.4).
+                Either raise daily inference hours, lower the volume target, or pick a faster preset —
+                the crossover verdict above cannot be physically realized at this setting.
+              </div>
+            </div>
+          )}
+          {!economics.infeasibleWorkload && (
+            <p className="text-[11px] font-mono text-steel-500">
+              Feasibility check: {economics.requiredInferenceHours}h decode required vs{' '}
+              {(dailyInferenceHours * 30.4375).toFixed(1)}h available — workload fits the configured duty cycle.
+              Payback is quoted against full CapEx; headline savings use the amortized basis.
+            </p>
+          )}
+
          {/* Interactive Chart Container */}
          <div className="border border-steel-700/60 rounded-xl p-5 bg-surface space-y-4">
            <div className="flex flex-col sm:flex-row sm:items-center justify-between gap-3 border-b border-steel-700/60 pb-3">
@@ -8,6 +8,7 @@ import {
  calculateReasoningCost,
  calculateMonthlyReasoningWorkload,
  type ReasoningEffort,
+  type ReasoningTurnCost,
 } from '../lib/reasoning';
 import type { CacheRate } from '../lib/types';
 import { formatPrice } from '../lib/pricing';
@@ -109,6 +110,16 @@ export default function ReasoningExploder() {
    return calculateMonthlyReasoningWorkload(activeModel, turns, cacheRate, effort);
  }, [activeModel, turns, cacheRate, effort]);

+  // Feasibility per effort level for the active model (VULN-06)
+  const effortFeasibility = useMemo(() => {
+    if (!activeModel) return null;
+    const map = {} as Record<ReasoningEffort, ReasoningTurnCost>;
+    for (const eKey of ['off', 'low', 'medium', 'high', 'max'] as ReasoningEffort[]) {
+      map[eKey] = calculateReasoningCost(activeModel, cacheRate, eKey);
+    }
+    return map;
+  }, [activeModel, cacheRate]);
+
  // Popular reasoning benchmark models for side-by-side comparison
  const comparisonModels = useMemo(() => {
    const targets = [
@@ -215,22 +226,38 @@ export default function ReasoningExploder() {
          {(['off', 'low', 'medium', 'high', 'max'] as ReasoningEffort[]).map((eKey) => {
            const spec = REASONING_EFFORT_SPECS[eKey];
            const isActive = effort === eKey;
+            const infeasible = effortFeasibility != null && !effortFeasibility[eKey].feasible;
+            const infeasibleTitle = effortFeasibility?.[eKey].infeasibilityReason ?? '';
            return (
              <button
                key={eKey}
-                onClick={() => handleEffortChange(eKey)}
+                onClick={() => { if (!infeasible) handleEffortChange(eKey); }}
+                disabled={infeasible}
                className={`px-3.5 py-1.5 rounded-full text-xs font-medium border transition-all ${
-                  isActive
+                  infeasible
+                    ? 'border-steel-800 bg-void-950/60 text-steel-600 line-through cursor-not-allowed'
+                    : isActive
                    ? 'border-blood-500/80 bg-blood-950/60 text-blood-200 shadow-[0_0_12px_hsl(0_70%_40%_/0.25)]'
                    : 'border-steel-700/60 bg-void-950/40 text-steel-300 hover:border-steel-500 hover:text-text'
                }`}
-                title={spec.description}
+                title={infeasible ? `Infeasible: ${infeasibleTitle}` : spec.description}
              >
                {spec.label}
              </button>
            );
          })}
        </div>
+
+        {/* Infeasible effort warning (VULN-06) */}
+        {effortFeasibility && !effortFeasibility[effort].feasible && (
+          <div className="mt-3 p-3 rounded-lg bg-amber-950/50 border border-amber-700/40 text-xs text-amber-200 flex items-start gap-2">
+            <AlertTriangle className="w-4 h-4 text-amber-400 shrink-0 mt-0.5" />
+            <div>
+              <span className="font-bold">Impossible combination.</span>{' '}
+              {effortFeasibility[effort].infeasibilityReason}
+            </div>
+          </div>
+        )}
      </div>

      {/* Interactive Controls & Cost KPIs */}
@@ -21,6 +21,7 @@ import {
  RotateCcw,
  Zap,
  TrendingDown,
+  AlertTriangle,
 } from 'lucide-react';

 export default function SessionReceipt() {
@@ -85,6 +86,7 @@ export default function SessionReceipt() {
      `================================================`,
      `          TOKEN-MAX AGENT SESSION RECEIPT       `,
      `================================================`,
+      receiptReport.tokensMeasured ? null : `*** ESTIMATED: transcript had no token accounting. ***`,
      `Session: ${activeSession.sessionTitle}`,
      `Turns Executed: ${activeSession.totalTurns}`,
      `Tool Invocations: ${activeSession.toolInvocationsCount}`,
@@ -212,6 +214,18 @@ export default function SessionReceipt() {
            {errorMsg}
          </div>
        )}
+
+        {!receiptReport.tokensMeasured && (
+          <div className="mt-4 p-3 rounded-lg bg-amber-950/60 border border-amber-600/40 text-xs text-amber-200 flex items-start gap-2">
+            <AlertTriangle className="w-4 h-4 text-amber-400 shrink-0 mt-0.5" />
+            <div>
+              <span className="font-bold">Estimated receipt.</span> This transcript contained no token
+              accounting — token counts and costs below are synthesized from the standard 21K agent
+              request shape, not measured from your session. Treat every dollar figure as a rough
+              planning estimate.
+            </div>
+          </div>
+        )}
      </div>

      {/* Session KPI Summary Cards */}