TIER B BUILT · GATE NOT CLEARED

AI Act Compliance Evidence Pipeline

If a regulator asks for the evidence behind your AI, can you hand it over? This pipeline generates the evidence a regulator asks for: Annex IV documentation and a tamper-evident audit trail, built from the run itself, not written from memory.

scaffold from €2,490

make run - CREDIT-SCORING MODEL (ANNEX III 5b)

Pipeline

✓train + calibrate AUC .79 · ECE .04

✓bias audit four-fifths 0.87 - passes

✓drift baseline PSI 0.04 · thresholds set

✓explanations SHAP per decision

✓500 decisions logged, hash-chained

✓annex_iv.md + evidence.json written

⛓-⛓-⛓-⛓-⛓-⛓-⛓-⛓
AuditLog.verify() → TRUE · tamper-evident · replayable

Annex IV technical documentation - generated

ANNEX IV - TECHNICAL DOCUMENTATION

per Regulation (EU) 2024/1689, Article 11 · generated 2026-07-14 03:12 UTC

§1 General description of the AI system✓ GENERATED

§2 Intended purpose & risk classification✓ GENERATED

§3 Data & data governance✓ GENERATED

§4 Metrics: accuracy, calibration, bias-by-group✓ GENERATED

§5 Human oversight measures✓ GENERATED

§6 Monitoring, drift & post-market plan✓ GENERATED

§7 Logging & record-keeping (Art. 12)✓ GENERATED

§8 Provider details & conformity route« OPERATOR »

EVIDENCE PACKAGE
ASSESSOR-REVIEWABLE

⚖ Produces evidence & documentation - a qualified assessor reviews before any conformity declaration.

How it works

Scope & risk class - with a human

What does the system decide, who does it affect, is it Annex III? The pipeline starts with operator checkpoints, not assumptions - risk class is never asserted by a machine.

One run generates the evidence

Calibration (AUC, Brier, ECE), bias-by-group with the four-fifths rule, SHAP explanations, drift baseline with alert thresholds - measured on your model, your data.

Three artifacts, regulator-shaped

annex_iv.md (Art. 11 technical file), audit_log.jsonl (Art. 12, hash-chained - tampering breaks the chain verifiably), evidence.json (every metric, replayable).

It stays current by itself

Nightly monitoring re-runs drift checks; every model upgrade regenerates the dossier. Compliance as a pipeline, not a PDF that rots in a drawer.

Methodology reference

The published method this build is based on, and what it did in production.

Google introduced Model Cards, now an industry standard for documenting model performance and limits.

Structured model documentation · as reported by arXiv ↗

Referenced for methodology only. AimRank is not affiliated with, endorsed by, or a customer of these companies; each links to its own public source. Sources last checked September 2026.

The Four Guarantees™ - this build

The scaffold wires all four in. The PoC proves them on your data, as of that day. Only Implementation keeps them true over time: watched, drift-checked, with recourse.

Measured value

A defensible conformity-evidence package in 3-6 weeks instead of ~6 months of manual documentation work - the agreed outcome (conservative).

Defensible

This IS the defensibility machine: Art. 11 technical file, Art. 12 tamper-evident log, Art. 13 explanations - generated from the actual system, not described from memory.

Self-improving

PSI/KS drift with WARN/ALERT thresholds (Art. 15, 72) feeding champion/challenger retraining; the dossier regenerates on every change.

Yours & everywhere

One container - GCP/AWS/Azure/on-prem. Full source. MCP tools (predict / explain / evidence / healthcheck) so your agents query compliance status directly.

The number, sized honestly

Reference run: a credit-scoring model (canonical Annex III 5(b) high-risk case) on the public German Credit dataset - the full pipeline you can watch execute before buying anything. Its own gate fails that reference model on the four-fifths rule (0.73 against the 0.80 floor on age group, run of 8 Sep 2026), so the pipeline is built and evaluated and does not call itself cleared: the dossier says so instead of hiding it, which is the product working.

3-6 wks to an assessor-reviewable evidence package - vs ~6 months manual

3 artifacts per run: Annex IV file · hash-chained log · evidence.json

Dec 2027 when Annex III high-risk obligations land, phased by the Digital Omnibus. Art. 50 transparency applies now.

100% of logged decisions replayable - tampering breaks the chain, verifiably

Three ways to own it

Tier	What you get	Price
Scaffolding ★	The full repo + templates - pipeline, Annex IV generator, audit chain, Terraform, MCP server. Your team points it at your model.	€2,490
PoC ★★RECOMMENDED	Run against one of your real AI systems - the Annex IV evidence pack generated end-to-end, gaps surfaced honestly. Code + quality report yours. No guarantee at PoC.	from€7,500
Implementation ★★★	Production: pipeline wired into your ML lifecycle, evidence auto-refresh, audit-chain monitoring, regulation-update maintenance, assessor-ready exports - the agreed number (conservative) attaches here.	from €22,500+ from €900/month assurance

★ = engagement depth. PoC is the recommended path: quality proven on your data before production money. The PoC carries no performance guarantee by design; the agreed number (conservative) attaches at Implementation, informed by the PoC report.

What we don't promise

Annex IV is a legal document - your compliance officer signs it, not us, and never a machine. This pipeline produces the evidence and the format; a qualified assessor or notified body reviews before any conformity declaration. We never assert your risk class - that's a checkpoint with a human and, where it's ambiguous, your counsel. Not legal advice, by design.

Ready to see your own number?

Request the build: within 48h you get a personal reply with the value sized to your volume.

No commitment · reply within 48h · your data stays in the EU