Skip to main content

The conformity file: assembling the evidence for the system you shipped in Module 3

The short answer

The file is assembled, not authored

You produced almost every page of the conformity file while building and testing the system in Modules 1 through 4. This topic is about putting those artifacts into the nine Annex IV slots in the order a reviewer reads, not generating new evidence. Treating the file as an end-of-project writing task is the mistake that leaves you reconstructing facts from memory.

What you will be able to do

  • Assemble a conformity file for a high-risk AI system by mapping the artifacts you already built (the trained model and its failure explanation, the data provenance file, the evaluation report, the incident log) into the technical documentation structure the EU AI Act requires in Annex IV.
  • Distinguish your role as provider or deployer of the system you shipped, and identify which evidence pack each role owns under Articles 16 and 26 of the EU AI Act (Regulation (EU) 2024/1689).
  • Name the nine elements of Annex IV technical documentation and place each of your existing artifacts into the correct slot.
  • Explain the difference between an internal self-assessment (Annex VI) and a notified-body assessment (Annex VII), and identify why a biometric system may trigger the notified-body route under Article 43.
  • Draft the EU declaration of conformity (the short signed statement required by Article 47 and Annex V) that sits at the front of the file, and state how long it must be retained (ten years, Article 18).
  • Attack your own file the way a regulator will: find the missing necessity justification, the undated version, the untraceable claim, and close each gap before it is found for you.
  • State the current application dates so you build to the real deadline, not a stale one: prohibitions and AI literacy apply since February 2025, general-purpose AI model rules since August 2025, and stand-alone Annex III high-risk obligations are deferred to 2 December 2027 after the 2026 Digital Omnibus amendments.

The ideas, one by one

Role decides the file

Provider or deployer is the first question, because it decides who owns the technical documentation, who signs the declaration, and what the file must contain (Articles 16 and 26). You shipped the Module 3 system, so you are its provider and you own the full file. Buying a system does not outsource responsibility; a deployer still owns oversight, monitoring, and impact-assessment records.

Annex IV is your table of contents

Nine slots: general description, development and data, capabilities and limitations, metric justification, risk management, lifecycle changes, standards applied, the declaration, and post-market monitoring. Seven you fill from existing artifacts, one is fresh (the declaration), one is a living commitment (post-market monitoring).

Self-assessment is the default; biometrics is the exception

Most Annex III high-risk systems self-assess under Annex VI, with no outside body reading the file before market. Biometric systems can trigger the notified-body route under Annex VII and Article 43 when harmonized standards are not applied. Assemble every file to the standard of a competent stranger, because you rarely know which route you will end on.

The necessity case is the heart of the file

The risk-management slot (Article 9) is where you justify building the system at all: the aim, the less intrusive alternatives, the harm to non-targets, and why the benefit outweighs it. This is the exact evidence the Bunnings determination found missing. A file that documents only how the system works, not why it should exist, cannot defend the decision to run it.

Sign last, and honestly

The declaration of conformity (Article 47) is issued under your sole responsibility and retained ten years (Article 18). Sign only after reading the file. An honest, dated late document is defensible; a fabricated contemporaneous one turns a compliance gap into a fraud finding. Mark your own gaps rather than hiding them.

Build to the real deadline

Prohibitions and AI literacy apply since February 2025; GPAI rules since August 2025; stand-alone Annex III high-risk obligations are deferred to 2 December 2027 and product-embedded high-risk to 2 August 2028 after the 2026 Digital Omnibus. The file is a living folder to have complete before the obligation bites, not a scramble at the deadline.

A chain of custody, not a binder

The expert does not read your file to admire it; they read it to break it, by pulling one claim and seeing if the trail goes cold. A claim that traces to a dated, versioned artifact survives; a claim that floats free, or points to a test that predates the live model, is a broken chain. This is why assembly from dated artifacts beats writing a polished report at the end: the report asserts, the assembled file proves. Every dated version stamp you add while shelving is one link an attacker cannot break later.

The file is a living folder, not a signing-day snapshot

Two of the nine slots are not retrospective: the declaration you sign last, and post-market monitoring (Article 72), which keeps the file true after signing. A file whose performance claims describe a model that has since drifted is already false, even if it was accurate the day it was signed. (see Topic 4.5) Seed post-market monitoring from your Module 3 incident process, name what you watch and how a field signal reopens the risk assessment, and review the file on a fixed cadence whether or not anyone asks. The file is complete only in the sense that a garden is complete: it stays true only if it is tended.

Mark your gaps; do not hide them

A gap you flag with an owner and a remediation date reads as a governance process working; a gap you hide reads, the moment it is found, as concealment that taints the whole file. A regulator will find material gaps during an inspection regardless, so the only thing you control is whether they are found already marked and being fixed. Marking converts a weakness into evidence of diligence.

The file is built to be attacked

Your conformity file is opened by a hostile examiner in Module 11, rebuilt after, linked by the evidence annex in Module 10, and inspected by the board in the capstone. A file that survives contact is real governance. Find and close your own gaps now, because every gap you leave is one the attack will use.

Listen to it

The same lesson, talked through by two hosts as episode 39 of the podcast.

Real cases

These are real, documented cases. Each shows what a conformity file is for by showing what happens when the evidence is not assembled in advance.

Example 1: Bunnings facial recognition (Australia, OAIC determination 2024). This is the anchor. Bunnings ran facial recognition across 63 stores in Victoria and New South Wales from 2018 to 2021. The system captured every visitor's face, converted it to a mathematical vector, and matched it against a database of persons of interest; non-matches were deleted almost immediately and matches were escalated to a trained team for a human decision (OAIC, 19 November 2024). Privacy Commissioner Carly Kind found breaches of the Privacy Act 1988: collecting sensitive information without consent, failing to take reasonable steps to notify people, failing to have practices and procedures to ensure compliance, and insufficient transparency in the privacy policy. The Commissioner did not impose a fine, citing Bunnings' cooperation, but ordered it to cease the practice and publish a statement of its failures. The lesson for the conformity file is exact: the company may have had operational reasons and even a genuine safety problem, but when the regulator asked for the evidence of necessity, proportionality, notice, and assessment, in one place, it did not exist in the required form. A conformity file is that evidence, assembled before the question is asked. (Note for currency: the Administrative Review Tribunal later reviewed aspects of the matter; the teaching point about assembling evidence in advance stands regardless of the review's outcome.)

Example 2: Rite Aid facial recognition ban (United States, FTC 2023). The US Federal Trade Commission banned the pharmacy chain Rite Aid from using facial recognition for five years after finding it deployed the technology without reasonable safeguards, generating false-positive matches that led staff to stop, search, and accuse shoppers, disproportionately affecting people of color (FTC, December 2023). The failure was not only the technology; it was the absence of documented testing, accuracy assessment, and human-oversight design that a conformity file would have forced into existence before deployment. This case is the owned anchor of another topic and is referenced here only to show the pattern. (see Topic 4.4)

Example 3: The EU AI Act's own text as the specification. The clearest "example" of what belongs in the file is the law itself. Annex IV (nine documentation elements), Article 9 (risk management), Article 47 and Annex V (declaration of conformity), and Article 43 (assessment route) are not commentary; they are the checklist. A provider who reads Annex IV as a table of contents and fills each slot from existing artifacts has a file. A provider who treats the law as background reading has a pile. The globally relevant point: even outside the EU, this structure has become the de facto template that serious buyers, insurers, and courts recognize, which is why assembling it is worth doing even where it is not yet legally compelled.

Example 4: A deployer's parallel pack (illustrative of Article 26). Consider a hospital in Ireland that buys a diagnostic support model from a vendor. The hospital is a deployer. Its file is not the vendor's Annex IV documentation; it is the vendor's declaration of conformity plus the hospital's own records: proof that clinicians using it were trained, that human oversight is real (a doctor signs the decision, the model advises), that the system is used only for its stated purpose, that logs are retained, and that a fundamental rights impact assessment was completed where required (Article 27). (see Topic 10.4) The lesson: the file follows the role. A deployer who assumes the vendor's file covers everything discovers, when a patient is harmed, that the oversight and impact-assessment records were theirs to keep and are missing.

The pattern across these cases. Read together, the examples show one recurring shape, and naming it makes the file's purpose concrete:

  • The technology often worked well enough that the failure was not primarily technical.
  • What was missing was assembled evidence that the deployment was necessary, proportionate, and disclosed, produced before the harm, not after the question.
  • The performance failures that mattered were uneven ones (false positives falling on a group), which a headline average hid and a disaggregated file would have surfaced.
  • The role (provider or deployer) decided who owed what, and buying the system never outsourced the duty to prove it was thought through.
  • In every case, a conformity file assembled in advance would have converted an investigation into a document handover.

That shape is why this topic exists: the file is the difference between "we had reasons" and "here is the evidence of our reasons, dated and traceable, before you asked."

Example 5: ISO/IEC 42005:2025, the international standard for the assessment inside your file. In 2025 the International Organization for Standardization and the International Electrotechnical Commission published ISO/IEC 42005:2025, "Information technology, Artificial intelligence, AI system impact assessment" (ISO, published 2025). It is a guidance standard, not a legal mandate: it gives organizations a repeatable structure for assessing an AI system's impact on individuals and society across the lifecycle, and its Annex A maps that assessment back to the AI management-system standard ISO/IEC 42001:2023 so a team already running a management system does not duplicate the work. The reason this matters for the conformity file is direct: the necessity and proportionality reasoning that lives in your Article 9 risk-management slot is exactly the kind of impact assessment ISO/IEC 42005 structures. A provider who conducts the impact assessment to a recognized standard, and cites it in slot 7 (standards applied), turns a slot that would otherwise say "no harmonized standard exists yet" into a slot that says "we applied ISO/IEC 42005 to structure the impact assessment." That is a stronger file. (Marker: emerging. ISO/IEC 42005 is newly published and is not yet a harmonized standard under the EU AI Act, so it does not by itself create a presumption of conformity; treat it as a credible engineering framework you chose, not as legal cover, and re-verify its status before relying on it.) The globally relevant point is that the "assemble the evidence" discipline this topic teaches is now codified in an international standard, which is why serious buyers and insurers increasingly expect it even where no law compels it. (see Topic 6.3)

Where people go wrong

  • "The conformity file is paperwork you write at the end." Wrong, and expensively so. The file is the assembly of evidence you generated while building and testing the system. If you treat it as an end-of-project writing task, you will be reconstructing facts from memory, which is exactly the position that turns a compliance gap into an indefensible one. Assemble as you build; the file is a folder that grows, not a report you draft.
  • "If we self-assess, no one will read the file, so it can be rough." Dangerous. Self-assessment under Annex VI means no outside body reads it before market, but a regulator, a court, an insurer, or a claimant's lawyer can demand it later, and a biometric system may force the notified-body route (Annex VII) where a stranger reads it up front. Assemble every file to the standard of a competent outsider, because you do not control who opens it or when.
  • "We bought the system, so the vendor's file covers us." Only partly. A deployer inherits real obligations under Article 26: oversight, monitoring, logging, and in defined cases a fundamental rights impact assessment (Article 27). The Bunnings case shows a deployer of a third-party capability still carrying heavy duties of consent, notice, and assessment. The file follows the role; buying does not outsource responsibility.
  • "A signed declaration is a formality." The declaration of conformity (Article 47) is issued under the provider's sole responsibility. Signing it is asserting the file behind it is true. Sign last, after reading the file, never first. A signature on an unread file is not a formality; it is a liability with your name on it.
  • "Overall accuracy is enough for the performance slot." No. Annex IV asks for the appropriateness of the metrics and the accuracy for specific persons or groups. A headline accuracy figure that hides a high false-positive rate on one demographic is precisely the failure the Rite Aid and Bunnings cases turned on. The file must show disaggregated performance, not a flattering average.
  • "An undated file is fine as long as the facts are right." An undated file cannot be trusted because a reviewer cannot tell which version of the system the evidence describes. Every artifact needs a version and a date, and the lifecycle-changes slot (Annex IV point 6) must record what changed and when. A substantial modification (Article 43(4)) can even require a fresh assessment; without dated change records you cannot prove you triggered or avoided that.
  • "We will build the file to the August 2026 high-risk deadline." Stale. The 2026 Digital Omnibus deferred stand-alone Annex III high-risk obligations to 2 December 2027 and product-embedded high-risk to 2 August 2028 (Council of the EU, 29 June 2026). Building to the old date wastes effort on the wrong schedule. Note also that prohibitions and literacy already apply, so parts of your obligation are live now.
  • "If we cannot prove necessity, we can add it later quietly." The necessity and proportionality reasoning is the heart of the risk-management slot and the exact thing Bunnings could not produce. If it does not exist, write it honestly and dated, and say when it was written. Backdating a fabricated version converts a documentation gap into a fraud, which regulators treat far more harshly than an honest late document.
  • "The foundation model's documentation is our documentation." Wrong, and it is the most common gap in files built on a licensed large model. The upstream general-purpose AI provider owes its own technical documentation and training-content summary under the Act (applicable since 2 August 2025), but that describes the general model, not your specific high-risk system built on it. (see Topic 5.5) Your file must still document how you integrated, configured, constrained, and tested that model for your purpose, and must record honestly where the upstream provider would not disclose what you needed. The fix: treat the vendor model card as an input to slot 2, then write your own integration and testing evidence; where a disclosure gap exists, name it and show how your own testing compensated, rather than leaving a silent hole.
  • "We scoped the file to how the system is supposed to be used." Incomplete. Article 9 requires the risk-management system to address reasonably foreseeable misuse, not only intended use. A watch-list tool designed for human-reviewed flags is foreseeably misused if a manager treats a flag as an automatic ban, and a file that ignores that path has a hole exactly where real harm happens. The fix: in the risk-management slot, list the two or three ways a rushed or incentivized human will actually misuse the system, and show the control (human-in-the-loop policy, override logging, reviewer training) you built against each. Scope to the path people take, not the path the design assumes.
  • "Post-market monitoring is next year's problem, so slot 9 can be a placeholder." Dangerous, because the file's truth decays the day you sign it. Slot 9 (post-market monitoring, Article 72) is what carries a field problem back into the risk-management system and keeps the whole file current; a placeholder there means that by the time a regulator opens the file, its performance claims describe a model that has since drifted. (see Topic 4.5) The fix: seed slot 9 from your Module 3 incident log and post-incident review (see Topic 3.4), name concretely what you watch (error rates by group, override frequency, new failure reports), how often, and the route by which a field signal reopens the risk assessment. A living slot 9 is what makes the quarterly review in the scenario real rather than aspirational.
  • "A high-risk classification means the harmonized standards will tell us what to do." Not yet. As of 2026 the harmonized standards for the AI Act are still being finalized, so most files cannot claim a presumption of conformity from a harmonized standard and must instead describe the other solutions applied (for example ISO/IEC 42001 for the management system, ISO/IEC 42005 to structure the impact assessment) and justify the engineering choices. The fix: in slot 7, state plainly that no fully harmonized standard yet applies, name the international standards and internal methods you used instead, and explain why they meet the requirement. Waiting for a harmonized standard to appear before assembling is how the deadline arrives with an empty file.
  • "A filled slot is a passing slot." Not the same thing. A slot can contain a document that fails its own test: a performance slot with only a headline average, a data slot pointing to an untraceable source, a risk slot describing function but no necessity case. Completeness (something in every slot) is easy to fake; a passing file has the right, traceable, dated evidence in each slot, tested against the one question a reviewer asks of it. The fix: grade each slot against the reviewer's question in section 3H, not against whether it is empty, and treat a filled-but-failing slot as a gap.

Questions people ask

What is conformity file?
The single assembled evidence pack for a high-risk AI system, structured on the EU AI Act's Annex IV technical documentation, that lets a reviewer open it and follow the system from intended purpose through data, testing, risk decisions, and human oversight to the signed declaration. More on Conformity file
What is provider?
Under Article 3(3) of the EU AI Act, the party that develops an AI system (or has it developed) and places it on the market or puts it into service under its own name. The provider owns the Annex IV documentation, the declaration of conformity, the CE marking, and registration. More on Provider
What is deployer?
Under Article 3(4), the party that uses an AI system under its own authority in the course of its activity. The deployer owns the Article 26 obligations: use per instructions, human oversight, monitoring, log retention, incident reporting, and (in defined cases) a fundamental rights impact assessment. More on Deployer
What is Annex IV?
The list in the EU AI Act of the nine elements of technical documentation a provider of a high-risk system must draw up and keep current, referenced by Article 11. It functions as the table of contents for the conformity file. More on Annex IV
What is article 9 risk management system?
The continuous, iterative process of identifying, evaluating, and mitigating the risks a high-risk AI system poses across its life. This is where the necessity and proportionality reasoning for building the system lives.

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →