Ctrl AI doesn’t answer your question and then write a log file about it. The answer is a sequence of signed Ctrl invocations, each producing an Ed25519 receipt anchored in an append-only transparency log. Article 12 record-keeping, Article 13 transparency, and Article 14 human oversight come for free with the architecture. The EU AI Act Starter Pack adds the typed Ctrls that encode the Act’s specific decisions — risk-tier classification, oversight gating, retention checks.
Fines up to €35M or 7% of global annual turnover
High-risk AI systems must automatically record events for traceability. Records must be appropriate for the system's intended purpose; deletable only after at least six months.
Systems must be designed for sufficient transparency so deployers can interpret outputs. Instructions for use, characteristics of training data, expected accuracy, and human oversight measures must be documented.
Effective oversight by natural persons. The system must allow the human to override, intervene, or interrupt. Automated decisions affecting natural persons require a verified path to human review.
The Act’s structural obligations — record-keeping, transparency, human oversight, robustness — are satisfied by how every Ctrl runs in Ctrl AI. You don’t install them; they are the platform. The Pack is the layer above: typed Ctrls that encode the Act’s specific decisions, so when your team asks an AI question, the agent has signed functions to call instead of opinions to render.
Every Ctrl invocation writes a receipt row and a transparency-log entry. The log is hash-chained — past entries cannot be rewritten without breaking every entry that follows. Six-month retention is a config minimum, not a feature you have to remember to turn on.
A Ctrl is a typed function — explicit inputSchema, outputSchema, post-conditions, human-signed rationale. The deployer reads exactly what will run. The LLM does not paraphrase prose into behaviour; it picks a Ctrl and calls it.
Destructive Ctrls refuse to invoke without an approved gate. The dispatcher enforces this — the model cannot bypass it. Delegations from a human to an agent or external API client are scoped by Ctrl, jurisdiction, condition, and expiry.
Post-conditions run after every invocation; a failure degrades the receipt's trust level (verified → partial → degraded). Typed inputs and outputs are validated at the boundary — malformed calls are rejected before they reach the runtime.
Below is what happens when a compliance officer at a mid-sized European bank (the fictional acme-bank) asks Ctrl AI a real question. Each step is a Ctrl invocation; each invocation produces a receipt; each receipt links back to the EU AI Act clause it satisfies. Nothing is paraphrased.
“Can we deploy this credit-scoring model to score retail mortgages for our French branch?”
Agent searches the installed catalog with “credit scoring EU AI Act risk classification”. Because acme-bank installed eu-ai-act-starter@v1 and the org is jurisdiction-tagged [“EU”], three Ctrls rank highest, sorted by clause match.
classify-ai-use-case-risk-tier [appliesToJurisdictions: ['EU']] human-oversight-required [satisfiesClauses: article_14] log-ai-action-decision [satisfiesClauses: article_12]
Agent calls classify-ai-use-case-risk-tier with the use-case description and French retail context. The grounded-LLM runtime cites Annex III as a source and returns JSON-schema-validated output. Two post-conditions pass.
{ riskTier: 'high',
reasoning: 'Annex III §5(b) covers AI used to evaluate
creditworthiness of natural persons.',
citedAnnexClauses: ['Annex III §5(b)'] }
trust = grounded · receipt rcp_a91f… signed · log seq #18,402Agent chains the result into human-oversight-required using {{steps.02.outputs.riskTier}}. The deterministic runtime returns a typed boolean — no model in the loop.
{ oversightRequired: true,
rationale: 'Article 14 mandates meaningful human oversight
for this risk tier.',
citedArticles: ['eu_ai_act:article_14'] }
trust = verified · receipt rcp_b73c… signed · log seq #18,403Agent records the decision via log-ai-action-decision so the determination itself is recorded under Article 12 — not just the question that triggered it. The record id becomes the audit anchor.
{ recordId: 'rec_8h2k4nq1',
loggedAt: '2026-05-13T09:42:14.221Z',
satisfies: ['eu_ai_act:article_12',
'eu_ai_act:article_13'] }
trust = verified · receipt rcp_c04e… signed · log seq #18,404This deployment is classified high-risk under EU AI Act Annex III §5(b) [ctrl:classify-ai-use-case-risk-tier#a91f]. Article 14 therefore requires meaningful human oversight before deployment [ctrl:human-oversight-required#b73c]. The determination has been recorded under Article 12 as record rec_8h2k4nq1 [ctrl:log-ai-action-decision#c04e]. You need an approved oversight gate from the DPO before this can ship.
Every [ctrl:slug#id] chip is clickable. Each opens the receipt, shows the Ed25519 signature, and links to the transparency-log entry on the public Trust Portal. None of the factual claims above can exist without a signed receipt to back them.
What the regulator sees if they ask: an append-only chain, four receipts, each signed, each tied to the article it satisfies, each independently verifiable against a publicly published public key. No screenshot evidence. No after-the-fact log reconstruction. The audit artifact is produced by the system, not collected from it.
The v1 Starter Pack ships the four highest-leverage decisions for early deployers. Each Ctrl has a typed input schema, a typed output schema, executable post-conditions, and a cited regulatory clause. Install materialises them into your org’s library, where your compliance officer refines and signs each one. The pack grows over time — v2 adds Annex IV technical documentation, conformity-assessment workflow, and GPAI obligations.
The clock is the regulator’s, not ours. Install the Pack today; customise the Ctrls to your context; ship signed AI before Aug 2026.