@fortress/provenance-sdk

EU-sovereign, offline invisible watermarking + content provenance. One package for:

• Resilient marks → forensic leak-tracing / anti-piracy (survive the distribution pipeline).
• Fragile marks → authenticity / tamper-evidence (break on any edit).
• Text/code copy detection → robust provenance fingerprinting (no hidden bits to strip).

No cloud, no API, no GPU. Every number below is reproduced by npm test. Boundaries are stated, not hidden (Rule 9 / Documentation/SSOT/BENCHMARK_TRUTH.md).

Eleven per-use-case facades (pick the right one — never mix)

const {
  // RESILIENT — survive distribution → attribution / leak-tracing
  ContentProvenance,   // AI-content image attribution (EU AI Act Art.50)
  AntiPiracy,          // studio leak-tracing, crop-hardened stills
  VideoProvenance,     // per-frame video + temporal majority (H.264)
  AudioProvenance,     // audio through MP3 128/64 kbps
  TextProvenance,      // text/code copy + paraphrase (fingerprint, no hidden bits)
  DataProvenance,      // JSON/CSV rows (float jitter + row-permutation codec)
  // FRAGILE — shatter on edit → authenticity / tamper-evidence
  LivenessAuthenticity,// KYC injection/replay
  ScienceIntegrity,    // scientific-image integrity
  CaptureAttestation,  // creator "captured live & unaltered" badge — LIVE-CAPTURE ONLY, NOT "not AI"
  TamperSeal,          // bit-exact signed-hash tamper-evidence — ANY modality, ANY 1-bit change
  // TOOLING
  LeakScanner,         // crawled suspect media → recipient (correct DC-QIM, not LL-STDM)
} = require('@fortress/provenance-sdk/markets');

Each facade hardcodes its validated profile so you cannot pick the wrong engine for a use case. Decision table + params: Documentation/SSOT/USE_CASE_TO_SDK_MAP.md. Crawler wiring: CRAWLER_INTEGRATION.md.

The "is it AI?" badge — honest scope (Rule 9)

CaptureAttestation proves "this exact file is unaltered since we sealed it" — NOT "not AI-generated". Software cannot prove sensor origin; true camera-origin needs hardware (L3 Play Integrity / L4 on-sensor). The SDK refuses to assert it (attestation.assertsCameraOrigin === false). See Documentation/STRATEGY/Creator_Platform_Shipping_Kit.md §3.

Install & use

const { FortressProvenance } = require('@fortress/provenance-sdk');
const fp = new FortressProvenance({ merchant: 'did:ohm:acme-media' });

// 1) Per-recipient leak-tracing (image; grayscale 1024x1024 Uint8/Float array)
const { pixels, keyHex } = fp.embedImage(gray, 'subscriber-0042');            // profile: 'stealth' (default)
const hit = fp.extractImage(suspectGray);                                     // → { found, keyHex }  (look keyHex up in your registry)

// crop-hardened profile (survives crop up to ~50%, slightly more visible)
const c = fp.embedImage(gray, 'subscriber-0042', { profile: 'crop-hardened' });
const hit2 = fp.extractImage(croppedSuspect, { profile: 'crop-hardened' });   // → { found, keyHex, via:'crop' }

// 2) Authenticity / tamper-evidence (fragile)
const a = fp.embedAuthenticity(gray, 'asset-001');
fp.verifyAuthenticity(suspect, a.keyHex);                                     // → { authentic, tamperDetected }

// 3) Text / code copy detection (paragraph-length text; any code)
const sig = fp.fingerprintCode(originalCode);
fp.matchProvenance(sig, fp.fingerprintCode(suspectCode));                     // → { similarity, isCopy }

// 4) Evidence package (DSR / provenance report)
fp.generateEvidence({ recipientId, assetId, keyHex, suspectUrl });

Validated capability scorecard (reproduced)

| Capability | Result | Profile | |---|---|---| | Image vs JPEG/recolor/noise/≤75% resize | 100% | stealth (~32 dB, margin 250) | | Image crop 10–50% (centered + off-center, ±H.264) | 100% (crop+H.264 80–100%) | crop-hardened (24 dB) | | Video — real H.264 CRF 18–35 | BER 0% | (backend video engine) | | Audio — real MP3 128/64 kbps | recovered, conf 0.97 | (backend audio engine) | | Fragile: reads original / detects screenshot+AI-edit | 100% / 100% (40 dB) | fragile | | Code copy: reformat + rename + comments | 100% | fingerprint | | Text copy: whitespace/case 100%, sentence-reorder 75%, light-reword 40% | — | fingerprint |

Profiles

• stealth — most invisible (~28–30 dB); full distribution-pipeline survival; no crop recovery.
• crop-hardened — full-key 4×4 tiling (24 dB); adds centered/off-center crop ≤50% + crop-after-H.264.
• fragile — ~40 dB; reads the untouched original, breaks on any manipulation (tamper-evidence / liveness).

Honest boundaries (do NOT claim past these)

• Generative AI redraw destroys the resilient mark (information-theoretic) — that destruction is the basis of the fragile product, not a defect.
• Crops >50%, blind crop+rotation combined, full 6-DoF affine — open R&D (precisely bounded in BENCHMARK_TRUTH.md).
• Text/code: full semantic rewrite / refactor escapes the fingerprint (info-theoretic). Text needs paragraph-length input (≥~30 words) for a reliable fingerprint. Heavy paraphrase is weak (≈1%); pair with a semantic-embedding layer for that.
• keyHex → recipient mapping is the integrator's responsibility (keep a registry; LeakScanner does this for you). Image API expects a 1024×1024 grayscale array. Video and audio use the same validated DC-QIM engine as images (NOT the LL-STDM stream engine — that one is JPEG-fragile and is for live-stream screenshots only).

Video, audio & CLI

// VIDEO — per grayscale frame (validated through real H.264 CRF28, 31 dB)
const v = fp.embedVideoFrame(grayFrame, 'subscriber-0042');
const hits = frames.map(f => fp.extractVideoFrame(f));
fp.extractVideoMajority(hits);                 // → { found, keyHex, votes }   (temporal majority vote)

// AUDIO — mono PCM Float32 [-1,1] (validated through real MP3 128 kbps)
const a = fp.embedAudio(pcm, 'subscriber-0042');   // a.pcm  (SNR ~20 dB — tune _audioEngine margin for your loudness target)
fp.extractAudio(suspectPcm);                       // → { found, keyHex }   (half-block sliding window handles codec delay)

# CLI (image I/O via ffmpeg)
fortress mark  cover.png subscriber-0042 --profile crop-hardened --out cover.marked.png
fortress trace suspect.png --profile crop-hardened     # → recovered key → look up recipient
fortress match original.js suspect.js                   # → code/text copy similarity
fortress key   subscriber-0042

Media validated: npm run test:media (real ffmpeg H.264 + MP3) → 4/4 PASS. Caveat: pure black/white synthetic patterns (e.g. color-bar test charts) clip at 0/255 and weaken the mark — real photographic/video content is unaffected. Audio SNR ~20 dB is a pilot default; lower the margin for inaudibility (trades codec-robustness).

Reproduce

npm test           # smoke 7/7 + market 22/22 + signer 4/4  (10 facades incl. badge + leak-scanner + data)
npm run test:media # media-test.cjs  → 4/4 PASS  (real H.264 + MP3 via ffmpeg)
npm run demo       # 7 demos exit 0  (incl. demos/data.cjs + demos/creator-platform.cjs)

Engines are faithful re-expressions of fortress-dwt-v3.engine.ts, cross-validated bit-for-bit (Δ 0.00 dB). Full truth registry: Documentation/SSOT/BENCHMARK_TRUTH.md; engine map: ENGINE_REGISTRY.md.