Approvals and Compliance
Maestro-1 works (Part 12) and is defended (Part 13). Before anyone outside the company uses it, Gopala has to answer to three audiences: regulators, customers' security and legal teams, and his own release gate. This part maps what the EU AI Act asks of a general-purpose model provider and where Maestro-1 sits (with a script that does the arithmetic), the voluntary frameworks customers ask about, the documents every release needs, customers' security reviews, and the release decision itself, which in the story first comes back no.
ai_act_check.py and release_check.py were run (real output below); the release manifest is the scenario. model_card.md is an outline, not executed. This is not legal advice; rules change and have exceptions, so a real company needs counsel. Disclosure: Claude, made by Anthropic, helps write this site.
1. Mission log: weeks 50โ60
- Maps obligations with counsel; writes
ai_act_check.py; starts the documents.
- Model card, data summary, copyright policy, security questionnaires from three partners.
- First release review: no-go (section 10). Second, a week later: go, for design partners first.
2. Why approvals are a product feature
- Enterprise customers buy evidence, not only capability.
- Their security teams ask where data goes, who can access it, how the model was tested and what happens in an incident.
- A start-up that answers in days instead of months closes deals faster.
- Regulation sets a floor.
- The EU AI Act applies to providers placing general-purpose models on the EU market, wherever they are based.
- Other jurisdictions have their own rules; Gopala starts with the EU because its obligations for model providers are the most specific today and his partners operate there.
Three audiences, three kinds of evidence
Much of the same evidence serves all three.
3. The EU AI Act and general-purpose models
- The key dates.
- Obligations for general-purpose AI (GPAI) model providers have applied since 2 August 2025.
- Models already on the market before that date have until 2 August 2027.
- The Commission's enforcement powers for GPAI start on 2 August 2026.
- Duties for every GPAI provider.
- Technical documentation of the model for the AI Office and national authorities.
- Information for downstream providers who build on it.
- A policy to comply with EU copyright law, including machine-readable opt-outs.
- A public summary of the content used for training, using the Commission's template.
- Extra duties above a compute threshold.
- A model is presumed to have "systemic risk" if its cumulative training compute exceeds 10^25 FLOPs; the presumption can be rebutted and the Commission can designate models below it.
- Those providers must also evaluate and adversarially test the model, assess and mitigate systemic risks, report serious incidents and ensure cybersecurity, and notify the Commission.
- The voluntary Code of Practice.
- Finalised on 10 July 2025, with chapters on transparency, copyright, and safety and security; signing it is one way to show compliance.
4. Where Maestro-1 sits
"""ai_act_check.py โ where does Maestro-1 sit under the EU AI Act's
general-purpose AI (GPAI) rules? A planning aid, NOT legal advice.
Facts used: GPAI obligations apply from 2 August 2025; a model is
presumed to have systemic risk above 1e25 FLOPs of cumulative training
compute (rebuttable; the Commission can also designate models). Training
compute uses 6 x active params x tokens (Part 3); post-training adds a
little. Scale-up scenarios are hypothetical.
"""
THRESHOLD = 1e25
RUNS = { # name: (active params, tokens)
"Maestro-1 pre-training": (11.7e9, 10e12),
"post-training + loop cycles (our estimate)": (11.7e9, 0.3e12),
}
BASE = ["technical documentation for the AI Office and downstream providers",
"information and documentation for downstream providers integrating the model",
"a policy to comply with EU copyright law, including opt-outs",
"a public summary of the content used for training (Commission template)"]
SYSTEMIC = ["model evaluations including adversarial testing", "assess and mitigate systemic risks",
"track and report serious incidents", "adequate cybersecurity protection"]
def main():
total = 0
for name, (n, d) in RUNS.items():
f = 6 * n * d; total += f
print(f"{name:44s} {f:.2e} FLOPs")
print(f"{'cumulative':44s} {total:.2e} FLOPs = {total/THRESHOLD:.1%} of the 1e25 presumption threshold")
print("obligations as a GPAI provider placing the model on the EU market:")
for o in BASE: print(" -", o)
for mult in (5, 10, 15):
f = total * mult
print(f"if a future run used {mult:2d}x the compute: {f:.1e} FLOPs -> "
f"{'presumed systemic risk: add ' + str(len(SYSTEMIC)) + ' more duties and notify the Commission' if f > THRESHOLD else 'still below the threshold'}")
if __name__ == "__main__":
main()Maestro-1's training compute vs the systemic-risk presumption
Real output of ai_act_check.py (6ยทNยทD estimate, 11.7B active parameters).
- What it means.
- Maestro-1 used about 7% of the compute that triggers the systemic-risk presumption, so it carries the four baseline duties, not the extra four.
- A future model with about 15 times the compute would cross the line; that is a planning fact for Maestro-2, not a distant one.
- The 7.02e23 here differs slightly from Part 3's 7.2e23 because Part 9 fixed the active size at 11.7B rather than 12B.
- The baseline duties still take real work.
- The training-content summary depends on Part 8's manifest; without it, this would take months.
5. Voluntary frameworks customers ask about
| Framework | Who publishes it | What it is | How Gopala uses it |
|---|---|---|---|
| AI Risk Management Framework 1.0 | NIST (US), January 2023 | Govern, map, measure, manage AI risks | Structure for the risk register and questionnaires |
| ISO/IEC 42001 | ISO/IEC, December 2023 | Certifiable AI management system standard | Planned for year two; certification is costly for a start-up |
| SOC 2 Type II | AICPA | Audit of security controls over time | Partly inherited from the GPU provider (Part 4); own audit in year two |
| OWASP LLM Top 10 | OWASP | Application security risks for LLM apps | Checklist behind Part 13 |
- Map once, answer many times.
- Each control (for example "access to training data is logged") is written once with its evidence, then referenced from every framework and questionnaire that asks.
6. The documents every release needs
- Model card.
- Mitchell et al. (2019) proposed model cards: intended use, limitations, evaluation results across conditions.
- His version adds confidence intervals from Part 12 and known safety weaknesses from Part 13.
# Maestro-1 โ model card (outline) ## Model - Developer: Maestro (start-up) ยท Version: 1.0 ยท Date: month 13 - Architecture: mixture-of-experts, 65.2B total / 11.7B active parameters - Training compute: ~7.2e23 FLOPs (6ยทNยทD estimate) ยท Context: 64K tokens ## Intended use - Software-engineering assistance inside customers' own repositories, through agents with a tool guard - Not intended for: unsupervised production changes, medical/legal advice, decisions about people ## Training data - ~10T tokens: filtered web, permissively licensed code, maths, books, multilingual, verified skills - Public summary: see data-summary.pdf (EU template) ยท opt-outs honoured ## Evaluation (harness v3, contamination-checked, 95% CIs) - Partner held-out tasks: pass@1 ... [CI] vs Maestro-0 ... [CI] - Public benchmarks: ... (with contamination notes) ## Safety - Red-team attack success by category with Wilson intervals; known weaknesses listed ## Limitations - Below frontier models on general knowledge; English-first; can still write insecure code ## Contact - security@ and incidents@ addresses, response times
- The other documents.
- Training-content summary (EU template) and the internal data manifest (Part 8).
- Copyright and opt-out policy.
- Evaluation and safety reports.
- The loop's gate log since the last release (Part 11).
- Incident process and contacts.
7. Customer security reviews
- Every partner sends a questionnaire; the questions repeat.
- Where is our code processed and stored? (Named regions, the sandbox design.)
- Is our data used to train your models? (Only verified skills, abstracted, under contract; never raw code.)
- Who can access it? (Named roles, logged, break-glass only.)
- How do you handle incidents? (Part 13's process with notification times.)
- The contract matters as much as the controls.
- Data processing terms, a clear statement of what is and is not trained on, and the right to have skills derived from their tasks removed.
8. The release review
- A meeting with a checklist and a decision.
- Attendees: Gopala (founder and safety lead), the evaluator hat's report, outside counsel, one partner representative.
- The release check runs first; if it says no-go, the meeting is about the gaps.
"""release_check.py โ the release gate for a model version.
Every item must exist and be signed before Maestro-1 ships to anyone.
The manifest below is the SCENARIO state at the first review: two items
are missing, so the answer is no.
"""
REQUIRED = {
"model_card": "model card: intended use, limits, evaluation results with CIs",
"eval_report": "evaluation report from the harness (Part 12), contamination checked",
"safety_report": "safety evaluation and red-team summary (Part 13)",
"data_summary": "public training-content summary (EU template) and data manifest",
"copyright_policy": "copyright and opt-out policy",
"gate_log": "loop gate decisions since the last release (Part 11)",
"incident_process": "incident contact, severity levels, response times",
"partner_signoff": "design partners reviewed on their own data",
"rollback_plan": "previous version kept serving; rollback tested",
}
MANIFEST = {
"model_card": {"done": True, "signed": "Gopala"},
"eval_report": {"done": True, "signed": "Gopala"},
"safety_report": {"done": True, "signed": "external red team + Gopala"},
"data_summary": {"done": False, "signed": ""},
"copyright_policy": {"done": True, "signed": "counsel"},
"gate_log": {"done": True, "signed": "Gopala"},
"incident_process": {"done": True, "signed": "Gopala"},
"partner_signoff": {"done": True, "signed": ""},
"rollback_plan": {"done": True, "signed": "Gopala"},
}
def main():
missing = []
for k, desc in REQUIRED.items():
m = MANIFEST.get(k, {})
ok = m.get("done") and m.get("signed")
print(f"{'OK ' if ok else 'MISS'} {desc}" + ("" if ok else f" ({'not done' if not m.get('done') else 'not signed'})"))
if not ok: missing.append(k)
print("RELEASE: " + ("GO" if not missing else f"NO-GO, {len(missing)} item(s) open: {', '.join(missing)}"))
if __name__ == "__main__":
main()- The first review says no.
- The public training-content summary was not finished, because two data sources in the manifest still lacked licence details.
- One design partner had not yet reviewed the model on their own data.
- What happened next.
- The two sources were removed from the summary's scope after confirming they had been excluded from training; the summary was finished.
- The partner ran their review in three days. The second review a week later said go, for design partners first.
9. Try the release gate
Release checklist (same items as release_check.py)
10. A staged release
Who gets Maestro-1, and when
Each stage has its own gate; any severity-1 incident pauses the next stage.
- Why staged.
- Real usage finds problems no harness does; finding them with five partners is cheaper than with five hundred customers.
- Phase 4 (Parts 15โ18) covers serving, products and operations through these stages.
11. Personal data in customers' code
- If client repositories contain personal data, data-protection law applies to processing it.
- In the EU, the GDPR requires a written data processing agreement when a service provider processes personal data on a customer's behalf (Article 28).
- The customer is usually the controller; Maestro acts as a processor for task execution.
- What that means in practice.
- A data processing agreement in every contract, listing sub-processors such as the GPU provider (Part 4).
- Processing kept in agreed regions; deletion after the retention period; help with data-subject requests.
- Training on derived skills needs its own legal basis and contract terms.
- That is why skills are abstracted, scanned and contractually agreed (Parts 8 and 13), and why partners can ask for removal.
12. One evidence library
- Every claim in every document points to a record.
- "Benchmarks were removed from training data" links to Part 8's decontamination log.
- "Updates are human-approved" links to Part 11's gate log.
- "Access is logged" links to access-log samples.
- Why it pays.
- A questionnaire answered in hours, an auditor's request answered in a day, a regulator's question answered with records instead of reconstruction.
From records to documents
Write the record when the work happens; documents become assembly, not archaeology.
13. The compliance calendar
| When | What |
|---|---|
| Every release | Release check, model card update, data summary update if training data changed |
| Every loop cycle (2 weeks) | Gate log entry with sign-off (Part 11) |
| Quarterly | Safety drill, red-team refresh, review of new regulatory guidance |
| Yearly | External security audit; review of the obligation map with counsel |
| Before any larger training run | Re-run ai_act_check.py; plan for systemic-risk duties if near the threshold |
14. Clauses that carry the promises
| Clause | What it says | Backed by |
|---|---|---|
| Data use | Raw customer code is never used for training; abstracted skills only with consent | Part 8 pipeline and manifest |
| Removal | Skills derived from a customer's tasks are removed on request from future training | Manifest and skill-library ids |
| Residency | Processing in named regions | Sandbox and provider configuration |
| Incidents | Notification within stated hours | Part 13 incident process |
| Sub-processors | GPU and cloud providers listed; changes notified | Part 4 contracts |
| Model changes | Notice before model updates; rollback available | Part 11 gate, Part 15 rollout |
- Never promise what the system cannot show.
- Each clause has a row in the evidence library; legal and engineering review it together.
15. Beyond the EU
- Rules differ by country and change often.
- Many jurisdictions regulate how AI is used (deployers) more than how models are built, and sector rules (finance, health) add their own duties.
- Gopala keeps one obligation map per market his customers operate in, reviewed yearly with local counsel.
- The series stays global for now.
- Localised guidance (for example for India) is planned as a separate section later in the site.
16. Who owns compliance as the company grows
- Today: Gopala, with counsel and agents drafting.
- At about 15 people: a part-time compliance and security lead.
- Owns the evidence library, questionnaires and audits; the founder still signs releases.
- At larger scale: separate safety, security and legal functions.
- The release gate gains independent reviewers who did not build the model.
17. Seven approval mistakes
- Treating compliance as paperwork after launch.
- Assuming a small model has no obligations.
- Writing the training-data summary from memory instead of a manifest.
- Answering each customer questionnaire from scratch.
- Promising "we never train on your data" without checking the pipeline.
- Releasing to everyone at once.
- Taking legal conclusions from a blog post (including this one).
18. Delegation box
- The release decision
- Positions on data use in contracts
- The obligation map with counsel
- Drafting documents from the manifest and reports
- First drafts of questionnaire answers
- Tracking regulatory updates
- Outside counsel (EU and contracts)
- GPU provider's audit reports
- A compliance tool for questionnaires
19. Budget and risk tracker
| Item | Status after Part 14 | Note |
|---|---|---|
| Spend to date | ~$7.4M (scenario) | Legal, documentation, reviews |
| Status | Released to design partners (scenario) | Second review: go |
| Time | Week 60 (month ~14) | Phase 3 complete |
| Regulatory risk | Managed | Below 1e25; baseline GPAI duties met |
| Customer-trust risk | Watch | Questionnaires answered; audits in year two |
| Future-scale risk | Plan | Maestro-2 at ~15ร compute would cross the threshold |
20. The series
The Maestro: 18 parts in five phases
- Phase 3 ends with evidence: a harness that can say no, six safety layers, a document set built from records rather than memory, and a release gate that did say no once before saying yes.
- Maestro-1 now goes to design partners.
- Phase 4 is shipping: Part 15, serving at scale.
- Follow it on Maestro Mission Control.
21. Reading the numbers
- Primary / standard
- EU AI Act GPAI obligations from 2 Aug 2025; legacy models until 2 Aug 2027; enforcement from 2 Aug 2026; 10^25 FLOP presumption; Code of Practice finalised 10 Jul 2025 (European Commission). NIST AI RMF 1.0 (Jan 2023). ISO/IEC 42001 (Dec 2023). Model cards (Mitchell et al., 2019).
- Secondary โ ๏ธ
- Some secondary summaries of the Code of Practice were used to check dates (โ ๏ธ); the Commission pages are the reference.
- Our assumptions
- The release manifest, review outcomes, timelines, costs and scale-up scenarios.