◆ Design partner previewCompliance Vault — sample data, real contracts.
Schedule a DP conversation →Layout, contracts, and data shapes are real. The data shown is illustrative. Real customer data flows live with our first design partner.
Annex IV — fixture preview.
EU AI Act · Annex IV §1–§9
Foundation Model A
mdl-foundation-A
mime: text/plainsize: 3009 bytesstatus: active
raw →◆ PDF backend pending (BE6 swap)Annex IV pack rendered as plain text. The real PDF templater lands when silver-annex-iv-v2's BE6 backend ships; the UI will branch on mime_type() and switch to embedded PDF preview without a contract change.
MOCK PDF — Annex IV §1–§9 placeholder
=====================================
NOTE: this is a text-fixture stand-in for the binary PDF that will be
emitted by the real Annex IV templater. MIME type returned by the
mock is `text/plain`, not `application/pdf`. See
`crates/silver-annex-iv-v2/MOCK.md` for the swap-when-ready plan.
Model: fixture-foundation-model-A
Generated: 2026-05-04 (mock; real impl will stamp real datetime)
Spec ref: EU AI Act Art. 11 + Annex IV
Format: Annex IV §1–§9, AI Office template alignment
§1 — General description of the AI system
-----------------------------------------
Foundation model "fixture-foundation-model-A" is a general-purpose
language model trained on a multi-source web corpus. Intended use:
downstream fine-tuning by deployers. Not directly user-facing in
production.
§2 — Detailed description
-------------------------
§2(a) Hardware: 4 × H100 nodes, 256 GPU-days.
§2(b) Architecture: 13B-parameter decoder-only transformer.
§2(c) Logic and key design choices: pre-norm, RoPE positional encoding,
SwiGLU feed-forward, grouped-query attention.
§2(d) Datasheets: see attached Croissant 1.1 manifest
`pii-in-training-corpus.json`. Corpus composition, license
attribution, and TDM-opt-out compliance enumerated there.
§2(e) Validation/testing: held-out perplexity 8.4; harm-bench refusal
rate 96.1%; copyrighted-text regurgitation rate 0.07%.
§2(f) Cybersecurity measures: SOC2 Type-II hosted training environment;
dataset access ACLed by role.
§3 — Monitoring, functioning and control
----------------------------------------
Per-deployment logging via standard output-capture; weekly drift
review; quarterly red-team review on novel jailbreak vectors.
§4 — Performance metrics
------------------------
Held-out perplexity: 8.4
Refusal rate (harm-bench): 96.1%
Copyrighted regurgitation rate: 0.07%
DSAR completion latency p99: 14 days
§5 — Risk-management system
---------------------------
Per ISO/IEC 23894:2023 + AI Office guidance. Risk register
maintained in `model-card-fixture-foundation-model-A.md` linked to
this pack.
§6 — Lifecycle changelog
------------------------
2026-04-01 Initial training run completed.
2026-04-22 Post-market patch v1.1: bias-mitigation fine-tune.
2026-05-01 GDPR Art. 17 erasure event for user 4192813 propagated;
partial — see `dsar-trace-4192813.json`.
§7 — Harmonised standards applied
---------------------------------
ISO/IEC 23894:2023 (AI risk management)
ISO/IEC 42001:2023 (AI management system)
NIST AI 600-1
ISO/IEC 5259-3 (Data quality for AI)
§8 — Declaration of conformity
------------------------------
DoC reference: DOC-2026-MOCK-001 (mock; real signed PDF lands when
the conformity-assessment workflow ships).
§9 — Post-market monitoring plan
--------------------------------
Quarterly auditor re-attestation; continuous incident-replay ledger
(see `/vault/incident/<id>` surface); annual full Annex IV refresh.
— END MOCK PDF —