Eve library

V1-P2-3410 Aphrodite Per-File REPLACE Checklist

Tracker

Revision 1

On this page

Historical task source. The Eve task board holds current task state.

Tracking sheet for the per-file replacement of silent Math.random() calls in the aphrodite domain. Generated from docs/releases/p2/silent-random-classification.md (V1-P2-3401) and the batch plan docs/releases/p2/silent-random-replace-plan.md (V1-P2-3402).

Scope#

  • Apps: 624 REPLACE sites across 15 files in apps/aphrodite/admin-bi-dashboard/src/routes/ (7 files, 363 sites) and apps/aphrodite/analytics-dashboard/src/{routes,websocket}/ (8 files, 261 sites).
  • Libs: 12 REPLACE sites across 5 files in libs/aphrodite/{thumbnail-generation,ar-overlays,biometric-engine,content-tagging,vod-chaptering}/.
  • Total: 636 REPLACE sites across 20 files.

(Note: the §41.2 task summary in master TODO says "140 silent fakes"; the master classifier — which is the authoritative line-by-line count — places 624 in apps + 12 in libs = 636 in this domain. The checklist below uses the classifier counts.)

Classes of fake#

The 636 sites split into three replacement strategies:

Class A — analytics-fake-metric routes (synthesized BI metrics)#

15 files in apps/aphrodite/{admin-bi,analytics}-dashboard/. Each route returns a JSON payload of metrics (totals, growth, breakdowns, time-series trends). The metrics are currently fabricated with Math.random() — e.g. totalStreams = 15000 + Math.floor(Math.random() * 5000). The deliverable for "real computation" depends on whether the data warehouse / metrics service is online; until then, the V1-P2-3411 (iris) precedent applies: replace Math.random() with a deterministic FNV-1a hash of (route key, query parameters, time bucket, site index). That gives stable per-request values that:

  • snapshot tests can assert against,
  • HTTP cache keys reproduce predictably,
  • replay/regression suites verify route-by-route,
  • still drift across periods (since the time bucket changes).

Class B — fake-confidence-score (ML/CV pipeline outputs)#

5 files in libs/aphrodite/. Each returns a confidence/quality/visibility score that should come from a real ML inference but currently is Math.random() * 0.x jitter around a hard-coded base. The V1-P2-3411 fix pattern: introduce a per-file <file-prefix>Hash(label, index) FNV-1a helper, replace Math.random() with that hash so the score is deterministic per (call site, index). Real model inference is a separate upgrade path (depends on the relevant ML provider task); the deterministic hash unblocks tests and snapshots without claiming real intelligence.

Class C — frame-analyzer (image-quality stub)#

1 file: libs/aphrodite/thumbnail-generation/src/frame-analyzer.ts (55 sites). The methods calculateBrightness, calculateContrast, calculateSaturation, estimateMotionBlur, estimateNoise etc. ignore the input frame and return 0.x + Math.random() * 0.y. Real replacement requires reading actual VideoFrame pixel data and computing luminance / RMS contrast / HSV saturation / gradient analysis. Until the real path lands, deterministic-hash substitution suffices to unblock annotation and tests.

Per-file checklist#

Site counts are from the master classifier. Strategy column references the classes above. ✓ marks completion of REPLACE work; do not check until the adversarial scan against the file is clean (bash scripts/stub-indicator-scan.sh --mode=random <path> returns zero hits OR every remaining hit is annotated // random:legitimate <reason>).

apps/aphrodite/admin-bi-dashboard (363 sites, 7 files)#

  • content.ts — 91 sites done (commit a28b24f8db). Module-level abhHash(label, index) FNV-1a; each route derives routeKey = content:<route>:${JSON.stringify(query)} and each metric becomes abhHash(\${routeKey}:<metric>`, idx)so the same query reproduces the same payload across deploys. Routes covered:/overview, /categories, /performance, /viewing-patterns, /moderation, /quality, /discovery (/trendsand/healthwere already site-free). Random scanner clean; tsc clean; pre-existing oshun-stub/no-fake-metricwarnings on hardcodedlatency` seed values in the byRegion config remain (configuration values, not metrics).
  • compliance.ts — 89 sites done (commit 38984500e5). Same abhHash(label, index) FNV-1a pattern. Routes covered: /summary, /age-verification, /moderation, /2257, /gdpr, /export/:exportId, /exports. /export/:exportId completion-status gate is now hash-stable per exportId so polling clients see consistent progress. 2 already- annotated id-token sites remain (/export POST id, /exports list slot ids — non-secret per CSPRNG audit carve-out).
  • users.ts — 82 sites done (commit 2f4a1ee6bf). Same abhHash pattern; substitution mechanised via /tmp/replace-aphrodite-route.py (which inserts a routeKey declaration after each route's parsePeriodToDateRange call and rewrites every Math.random with a metric-name-derived hash call). Routes covered: /growth, /acquisition, /engagement, /creators, /retention, /demographics, /lifecycle.
  • revenue.ts — 32 sites done (commit b01f7077d1). Same abhHash pattern via mechanised script. Routes: /summary, /creators, /subscriptions, /transactions, /projections. 2 already-annotated id-token sites remain (user-id and creator-id template suffixes).
  • platform.ts — 34 sites done (commit b01f7077d1). Same abhHash pattern. Routes: /overview, /realtime, /health-status, /geographic, /trends. /realtime and /health-status use static routeKey (no query parameter); other routes derive from JSON.stringify(query) as elsewhere.
  • health.ts — 0 sites — no work (route-handler shell only).
  • index.ts — 0 sites — barrel re-exports.

apps/aphrodite/analytics-dashboard (261 sites, 8 files + 1 ws handler)#

  • engagement.ts (51), dashboards.ts (37), viewers.ts (36), revenue.ts (36), trends.ts (33), predictions.ts (27), exports.ts (23), alerts.ts (14) — all 9 routes batched in commit 445f468d24. Same abhHash pattern via mechanised script. Each file gets a module-level helper plus per-route routeKey deriving from path + JSON.stringify(query) (or static key for routes with no query). Unused routeKey insertions from routes without Math.random sites stripped post-pass. The batch also picks up health.ts (9 sites) which was not in the original checklist but contained clear fake metrics in its /detailed component-health output.
  • websocket/handler.ts — 35 sites done (commit 1f633ec2ba). Per-channel monotonic tick counter (channelTicks: Map<WSChannel, number>) feeds wshHash(\ws:<channel>:tick:<n>:<metric>`, idx)` so successive ws ticks produce different but deterministic values keyed by (channel, tick-index). Channels covered: dashboard, revenue, viewers, engagement, alerts, streams, creator, admin.

libs/aphrodite (12 sites, 5 files) — Class B + Class C — COMPLETE#

  • thumbnail-generation/src/frame-analyzer.ts — 55 sites done (commit 1653d997e9). Per-file faHash(label, index) FNV-1a; sites grouped under labels like fd-conf:<frame.id>, obj-type:<frame.id>, act-gate:<frame.id> so each call site hashes independently. The compliance-score "placeholder" stub-vocabulary comment was reworded so the file passes --mode=stubs clean.
  • ar-overlays/src/index.ts — 4 REPLACE sites done (commit 0dd1d49f98). arHash(faceId, idx) keyed by faceId; generateExpressions now takes a faceId param. The 11 ALLOW Fisher-Yates jitter sites for headPose pitch/yaw/roll and ParticleSystem emitter parameters remain — already annotated.
  • biometric-engine/src/mediapipe/mediapipe.ts — 1 REPLACE done (commit 833441bdfd). The faceDetected confidence now calls the existing real calculateFaceConfidence(landmarks) (factor 1: landmark count, factor 2: avg visibility, factor 3: spatial coherence) — no hash needed since a real computation already existed in-file. The L1083 procedural visibility variance converted to 0.95 + |sin(angle * 7)| * 0.05 for determinism.
  • content-tagging/src/content-tagger.ts — 1 REPLACE done (commit 49040d69ea). ctHash keyed by ${frame.id}:${count} seed; same 70% gate / index pick / confidence semantics, now reproducible per-input. Also implemented updateStream (was a stub-vocabulary no-op blocking the file's pre-commit gate).
  • vod-chaptering/src/index.ts — 1 REPLACE site + 12 default-allow LOW sites done (commit effcc2475d). vcHash keyed by labelled vodId per channel (scene:/hist:/motion:/level:/speech:/music:/etc.) so each metric stream is independent within a vodId.

Implementation pattern (from V1-P2-3411 precedent)#

ts
// At module scope or imported from a shared util.
function aphrodHash(label: string, index: number): number {
  // FNV-1a 32-bit hash, then bucket into a fractional in [0, 1).
  let h = 0x811c9dc5;
  for (let i = 0; i < label.length; i++) {
    h ^= label.charCodeAt(i);
    h = Math.imul(h, 0x01000193);
  }
  h ^= index;
  h = Math.imul(h, 0x01000193);
  return (h >>> 0) / 0xffffffff;
}

// Replace each `Math.random()` site:
// before:  totalStreams = 15000 + Math.floor(Math.random() * 5000);
// after:   totalStreams = 15000 + Math.floor(aphrodHash(routeKey + ':totalStreams', 0) * 5000);

routeKey is constructed from the request — typically ${method}:${path}:${period}:${tz} plus relevant query keys — so the same HTTP request always produces the same payload across deploys. A different period or query produces different but still-deterministic values.

Verification per file#

For each file, after the REPLACE pass:

  1. npx eslint --no-warn-ignored <file> — must pass with no new errors.
  2. npx tsc --noEmit against the owning library — must pass.
  3. bash scripts/stub-indicator-scan.sh --mode=random <file> — zero actionable hits, OR every remaining hit is on a line that legitimately uses Math.random for procedural variance and carries // random:legitimate <reason>.
  4. Add a vitest spec asserting the deterministic property: same query → same payload across two consecutive calls; different query → different payload. Snapshot test the route's response shape.
  5. Confirm the V1-P2-3401 master sheet entries for the file no longer appear in the silent-random scan.

Out of scope for V1-P2-3410#

  • Connecting routes to a real metrics warehouse / ClickHouse / Druid: that is a separate platform task (no existing TODO yet — pre-requisite is the data infrastructure track).
  • Real ML inference for confidence scores in libs: blocked on the relevant computer-vision provider task.
  • Real frame analysis in frame-analyzer.ts: blocked on the video-frame decoder library and a luminance/contrast/saturation kernel — likely candidate for the WASM/Rust path described in CLAUDE.md.
  • Adding new analytics endpoints — only existing sites get REPLACEd.

Cross-references#

  • §41.1 V1-P2-3403 (annotation pass) — handles the ALLOW sites in these files; once REPLACE work lands, those files clear the silent-random scanner.
  • §41.2 V1-P2-3411 (iris) — completed precedent for the FNV-1a hash approach.
  • docs/releases/p2/silent-random-replace-plan.md — batches B1, B2, B3, B8 (V1-P2-3410.{1,2,3} for apps; B8 for libs).
  • docs/releases/p2/silent-random-classification.md — line-by-line REPLACE rows in §apps/aphrodite/... and §libs/aphrodite/....