Was this work used to train that model?

A signed seal is a verifiable record of provenance. It is not a substitute for the chain-of-custody discipline maintained by counsel. Specific cryptographic primitives and version numbers are disclosed to counsel and opposing experts under MNDA.

Holdings summarized for marketing, accurate but compressed. This is product capability, not legal advice.

1. STEP 01

   Submit work + suspect model

   Upload your copyrighted content and select the target LLM, image, audio, or video model. We never retain plaintext beyond the engagement window.

   iplyr ▸ ingest

   bash

   1$ iplyr submit \
   2    --work ./novel.epub \
   3    --target anthropic/claude-3.7 \
   4    --jurisdiction US-9 \
   5    --retention 30d
   6  ✔ work hashed     sha256:b2f0…d41a
   7  ✔ AES-256 sealed  envelope_id env_01J…
   8  → matter_id mat_01J9X8K2QV

2. STEP 02

   Probe & extract

   IPLYR runs multi-turn extraction probes, MIA ensembles, and cross-modal alignment against the target model with adversarial-robustness controls.

   iplyr ▸ probe

   bash

   1[probe 01]  divergence-attack         seed=… budget=2048t
   2[probe 02]  prefix-completion         k=64  temp=0.0
   3[probe 03]  paraphrase-canary         n=128
   4[probe 04]  style-fingerprint (CSD)   d=512
   5[probe 05]  watermark-residual        FFT-band=π/4
   6  → 16 probes complete · 14 positive · 0 inconclusive

3. STEP 03

   Score & classify

   TQ + nv-recall + concept memorization + RAG two-stage analysis produce per-block scores with confidence intervals.

   findings.exhibit-A · per-block evidenceTier A

   block 0007match 0.94

   high

   block 0011match 0.88

   high

   block 0014match 0.82

   high

   block 0019match 0.61

   moderate

   block 0022match 0.41

   low
4. STEP 04

   Evidentiary classification

   Results are mapped to SaTML 2025 evidence tiers with Daubert compliance flags and jurisdiction-specific thresholds.

   A

   Daubert-ready

   Reproducible, peer-reviewed methods, error rates published.

   B

   Pre-filing diligence

   Strong indicators; supports complaint and discovery.

   C

   Investigative

   Signal warrants further extraction probes.

   D

   Insufficient

   No memorization detected at adversarial threshold.
5. STEP 05

   Court-admissible package

   Signed PDF with chain of custody, expert-declaration template, exhibit list, and blockchain-anchored hash. Ready for filing.

   package.iplyr

   json

   1{
   2  "package_id": "pkg_01J9X8K2QV",
   3  "tier": "A",
   4  "daubert_compliant": true,
   5  "chain_of_custody": "sha256:9c3a…f201",
   6  "btc_anchor": { "block": 832914, "txid": "5fa1…" },
   7  "exhibits": ["A-Findings.pdf", "B-PerBlock.pdf", "C-Methods.pdf"],
   8  "expert_declaration_template": "decl-rev-2025.03.docx",
   9  "reproducible_from_hash": true
   10}

Forensic evidence is a category none of them serve. Plaintiffs in Kadrey, Getty, and Bartz-style matters are losing for lack of exactly this.

{
  "model": "anthropic/claude-3.7",
  "work_blocks": ["sha256:b2f0…d41a", "sha256:9c3a…f201"],
  "k_percent": 20
}

…

40+ attribution & forensic endpoints · 8 TDPI marketplace endpoints · 6 memorization-risk endpoints. Full OpenAPI 3.1 spec at /docs/api.