The evidence annex: wiring every artifact so the Module 13 audit finds a paper trail, not a scramble
The short answer
The side with the paper trail wins
Li v. Liu was decided not by talent or honesty but by provenance: the party who could document who made the AI-generated image, through prompts, parameters, iterations, and an identifying watermark, prevailed, and the party who could not show where their copy came from lost. Every serious challenge to an AI decision, from a court to a viva, resolves the same way, and the evidence annex is your standing paper trail prepared before the challenge arrives.
What you will be able to do
- Assemble an evidence annex that wires every governance artifact you have built across the program into one indexed register, so any piece of evidence can be produced on demand rather than hunted for.
- Structure each annex entry so it carries the six things an auditor needs: what the artifact is, what it proves, who made it, when, where it authoritatively lives, and which system and obligation it maps to.
- Stamp provenance on every entry, recording who made what and when in a way that can be verified as unaltered, so the annex answers the Li v. Liu question ("who made this, and how do you know") for your whole AI estate.
- Map evidence in two directions, tying each artifact to the system it concerns and to the obligation it discharges, so a request framed either way ("show me everything about this system" or "prove you meet this duty") lands on the right artifact.
- Read your own gaps as findings, using the annex to expose which obligations have no evidence and which systems have no trail, and treating each blank cell as a task, not an embarrassment to hide.
- Keep the annex live, maintaining it as artifacts change rather than assembling it in a panic at the end, so the register is always current when the audit arrives unannounced.
- Defend the annex against the charge that it is a cosmetic index over a hollow file: prove that every entry points to a real, dated, provenance-stamped artifact an auditor can open and trust.
The ideas, one by one
The annex is an index with provenance, not a warehouse or an essay
It points to every artifact where it authoritatively lives and stamps each pointer with who made it, when, and proof it is unaltered. Copying artifacts into one folder creates a museum of stale duplicates; writing a fresh narrative creates an essay, which is a claim without a trail. The value is entirely in the wiring.
Six fields or no row
Every entry carries what it is, what it proves, who made it, when, where it authoritatively lives, and what it maps to. Each field answers a question the auditor will otherwise ask you live under pressure. A row missing any of them is a false promise of evidence, and an artifact whose "what it proves" is blank may be evidence of nothing.
Provenance turns a document into evidence
Anyone can produce a document that says a model was tested; evidence is a report you can show a named person produced on a real date and nobody has altered since. Version control gives you most of this for free; where artifacts live outside it, a dated signature or a stored file hash approximates it. Evidence that can be tampered with is decoration, not proof.
Map in two directions or fail half the questions
Auditors ask by system ("show me everything about this model") and by obligation ("prove you meet this duty"). Build both lookups over the same rows, using the AI systems inventory for the system axis and the binding frameworks for the obligation axis, and the annex answers either shape with equal ease.
Gaps are findings, and finding them yourself is a gift
The most valuable output of building the annex is the blank cells it exposes: obligations with no evidence, systems with no assessment, stale cards, unfindable reports. A gap you mark as a dated finding with an owner is a task and a signal that you govern yourself; a gap you hide is a landmine the auditor will step on, and then wonder what else you concealed.
Keep it live, and build it to be inherited
An annex assembled the week before the audit is a scramble with better handwriting, and one that lives in a single person's head leaves when they do. Wire its maintenance to the events that change evidence, give it a named owning role, and name each artifact's accountable maker, so any audit, ninety days out or two years out under a new lead, finds a paper trail.
The burden of demonstration is on you
Under the EU AI Act, demonstrating compliance means producing the evidence on request, not merely having done the work. An organization that did everything right but cannot produce and authenticate its trail has not demonstrated compliance, and the annex is precisely the mechanism that makes the whole program demonstrable, which is why it is the last thing built before the audit reads it.
One annex serves every regime that asks you to show your work
The same register, with an obligation axis that can group rows by the EU AI Act, ISO/IEC 42001, the NIST AI RMF, or a state law, discharges the demonstration expectation of all of them at once. Wiring evidence well once, with provenance and two-way mapping, is leverage across every framework and regulator, which makes the annex an investment rather than a compliance cost.
Match provenance strength to the stakes
A low-risk tool's card can live under version control; the evidence behind a decision that harmed a person, or a log that a forensic reconstruction will examine, deserves the strongest integrity you can give it. Record per entry how each artifact's integrity is shown, so a weak-provenance row on a high-stakes artifact stands out as its own kind of gap before an adversary points at it.
Listen to it
The same lesson, talked through by two hosts as episode 81 of the podcast.
Real cases
These are documented cases used to illustrate the annex, not to predict your organization. Each is cited and used for a specific point.
Example 1: Li v. Liu and the trail that won the case (this topic's anchor). On 27 November 2023 the Beijing Internet Court ruled that an AI-generated image was copyrightable and belonged to the person who prompted it, because he could document his intellectual inputs: more than one hundred and fifty prompts, their arrangement, specific parameters, and iterative adjustments, plus a watermark carrying his identity (Kluwer Copyright Blog, 2023; China IP Law Update translation, 2024). The value for the annex is exact and load-bearing. The plaintiff prevailed because he had provenance he could produce; the defendant, who had removed the watermark and could not say where she obtained the image, had none. The case is the evidence annex in miniature: the party with the documented trail of who made what wins, and the party who scrambles for provenance loses. Your annex is the standing version of the plaintiff's trail for your whole AI estate.
Example 2: When the records, not the algorithm, decided the outcome. In the Dutch childcare-benefits scandal, the collapse came not only from a discriminatory risk system but from the state's inability to produce clean records of who was flagged, why, and on what basis, which turned a policy failure into a governance catastrophe (see Topic 10.1). Owned by Topic 10.1, it is referenced here for one point: the scandal shows that when evidence is scattered and unreconstructable, an organization cannot defend even the parts of its conduct that were defensible. The annex exists so that the records are never the reason you lose.
Example 3: Provenance had to be extracted because it was not offered. In the Rotterdam welfare-fraud case, investigators could only prove the scoring discriminated after obtaining the source code, model, and training data, evidence the organization had not organized for inspection (see Topic 10.2). Owned by Topic 10.2, the point for the annex is the inverse lesson: an organization that had wired its own evidence into a register could have produced that trail itself, on its own terms, rather than having it pried loose. The annex is the difference between disclosing your evidence and having it extracted from you.
Example 4: The EU AI Act puts the burden of demonstration on you. The AI Act's documentation and record-keeping obligations for high-risk systems, deferred by the 2026 Digital Omnibus to 2 December 2027 for standalone (Annex III) systems and 2 August 2028 for high-risk AI embedded in regulated products (Annex I), require the organization to be able to demonstrate compliance once they apply, not merely to have complied (Regulation (EU) 2024/1689; established). The example matters because it converts the annex from good practice into the mechanism of a legal duty: "demonstrate" means "produce the evidence on request," and an organization that did the work but cannot produce it has not demonstrated anything. Building the annex before the deadline binds, rather than after, is what turns the deferral into an advantage instead of a false sense of extra time. The annex is how the burden of demonstration is met.
Example 5: Provenance written into the content itself. China's 2025 labeling measures require AI-generated content to carry implicit provenance metadata (provider and content identifier) inside the file, so origin travels with the artifact (Cyberspace Administration of China, in force 1 September 2025, with GB 45438-2025; established). The example shows regulators independently reaching the annex's core instinct: provenance should be attached and verifiable, not asserted from memory. An organization operating in China must map that duty in its annex's obligation axis, and the deeper lesson generalizes everywhere.
Example 6: A management standard's evidence expectation. ISO/IEC 42001:2023, the first certifiable AI management system standard, expects an organization to maintain documented information and to make evidence of its controls available for audit (ISO/IEC 42001:2023; established as a voluntary, certifiable standard). For an organization pursuing certification, the evidence annex is close to the exact artifact an auditor of that standard will ask to walk, which means the same annex that serves the Module 13 audit also serves a real external certification. The point for the annex: build it once, and it discharges the demonstration duty across several regimes at once.
Example 7: The framework that assumes you can produce your trail. The NIST AI Risk Management Framework organizes governance around Govern, Map, Measure, and Manage, and every one of those functions produces artifacts a mature program is expected to be able to show (NIST AI RMF 1.0; established as voluntary guidance). The example illustrates the obligation axis: the annex can group its rows by RMF function to show, at a glance, that each function of the framework is backed by real evidence rather than by a claim of adherence. Frameworks assume the trail exists; the annex is where the assumption is made true.
Example 8: The generic scramble that precedes most failed audits. Across regulated industries long before AI, the recurring pattern in failed inspections is not that the work was undone but that the organization could not produce the records fast enough or prove they were genuine, so the inspector concluded the controls were not real (the pattern is visible in audit findings across financial, safety, and data-protection domains). The point for the annex is structural rather than tied to one named case: the scramble itself is the finding. An inspector who watches an organization hunt for its own evidence has already learned the most important thing about how that organization actually operates, regardless of what the eventually-produced document says.
Example 9: Records that could not be reconstructed after the fact. In several documented public-sector algorithm failures, the harm was compounded because the deciding organization could not later reconstruct which version of a system made which decision, on what data, under whose sign-off, so affected people could not even be told why they were flagged (a pattern the Dutch and Rotterdam cases both show, owned by Topics 10.1 and 10.2). The point for the annex is about timing: evidence captured as the work happens, with provenance attached, is producible; evidence reconstructed under legal pressure years later is weak, contested, and sometimes impossible to build honestly. The annex exists precisely so that the trail is captured while it is cheap and true, not excavated when it is expensive and doubtful.
Example 10: One annex serving several regimes at once. An organization certified to ISO/IEC 42001, subject to EU AI Act documentation duties, and voluntarily aligned to the NIST AI RMF does not need three separate evidence systems; the same well-built annex, with its obligation axis able to group rows by any of the three, discharges the demonstration expectation of all of them (ISO/IEC 42001:2023; Regulation (EU) 2024/1689; NIST AI RMF 1.0; all established). The point for the annex is leverage: the effort of wiring evidence once, with provenance and two-way mapping, pays off across every framework and regulator that asks the organization to show its work, which is why the annex is an investment rather than a cost.
Where people go wrong
- Copying artifacts into one folder instead of pointing to them. The most common structural error. A folder of copies creates duplicates that drift from the originals the moment anyone updates a file, and within weeks the annex is a museum of stale versions that misleads the auditor. Keep every artifact where it authoritatively lives under version control and make the annex point to it. The annex is a map, not a warehouse.
- Writing the annex as a fresh narrative essay. If your instinct on hearing "annex" is to author a new prose account of everything you did, stop; that is an essay, and an essay is a claim without a trail, which this program refuses. The annex is pointers and provenance, almost no original prose. Its value is in the wiring, not in new writing.
- Leaving out the provenance and just listing documents. A list of document names is not evidence; it is a table of contents for claims. Without who made each artifact, when, and a way to show it is unaltered, every entry is a "trust us." Stamp verifiable provenance on every row, because provenance is what turns a document into evidence, and its absence is exactly what lost Li v. Liu for the defendant.
- Building only one lookup. An annex indexed only by system fails "prove you meet this obligation," and one indexed only by obligation fails "show me everything about this system." The auditor asks both ways. Build both axes over the same rows, using the AI systems inventory for the system axis and the frameworks that bind you for the obligation axis.
- Hiding the gaps to look seamless. The strongest temptation and the worst move. A gap you conceal is a landmine the auditor will step on; a gap you mark is a task that shows you govern yourself. Write every blank cell into the annex as a dated finding with an owner and a plan. The annex that shows its own holes is trusted; the one that pretends to be seamless and is caught is not.
- Treating a stale artifact as current because it exists. A model card written before a retrain, an assessment done before a system changed, is stale evidence wearing a current face, which is a trap you set for yourself. The date and version fields exist so you and the auditor can judge freshness; tie refreshes to the changes that age artifacts, as the module's card and log logic requires (see Topic 10.3).
- Letting a pointer dead-end. An artifact you cannot locate is, to an auditor, an artifact you do not have. A "where it lives" field that resolves to a moved file or a dead link is a scramble waiting to happen mid-audit. Every pointer must resolve to the authoritative copy, and checking that they do is part of keeping the annex live.
- Assembling the annex the week before the audit. An annex built in a panic is a scramble with better handwriting, and real audits, regulator letters, and incidents do not wait for you to tidy up. Maintain the annex continuously, wired to the events that change evidence, so it is always ready rather than being made ready.
- Giving the annex no owner, so it rots when a person leaves. An annex that lives in one person's diligence dies when that person moves on, taking the trail with them. Give the annex a named owning role and name each artifact's accountable maker, so the register is inheritable, which is the whole point of a trail that must outlast individuals (see Topic 12.5).
- Mapping every system to your home jurisdiction only. A multinational annex that ties all systems to European duties and ignores that a system in China must carry content-provenance metadata, or that a United States hiring system sits under state bias-audit rules, is confident and wrong outside its home country. Map each system to the provenance duties of the jurisdiction where it actually runs.
- Confusing the conformity file with the annex. The conformity file is a bundle of evidence for one high-risk system (see Topic 5.6); the annex is the register that points to every artifact across every system, including that conformity file as one entry. Do not mistake having assembled one system's file for having wired the whole estate. The annex is the layer above the files.
- Believing that having done the work is the same as being able to prove it. Under the EU AI Act the burden of demonstration sits on you: "demonstrate compliance" means "produce the evidence on request." An organization that did everything right but cannot produce and authenticate the trail has not demonstrated compliance, and in an enforcement posture that is close to not having complied. The annex is how doing becomes demonstrable.
- Giving every artifact the same weak provenance regardless of stakes. A model card for a low-risk internal tool may be fine under version control, but the log behind a decision that harmed a person, or evidence that will be examined in a forensic reconstruction, deserves the strongest integrity you can give it (a stored hash, a signed immutable log). Match provenance strength to the stakes, and note per entry how integrity is shown, so a weak-provenance row on a high-stakes artifact is visible as its own kind of gap.
- Adding a row without confirming the pointer opens. It is easy to record a plausible-looking location and move on, but a pointer you have not actually followed is a pointer you cannot trust; the file may have moved, the repository may have been archived, the link may resolve to an old copy. Open every pointer as you add it, and re-check them in the backstop review, because a dead pointer discovered mid-audit is indistinguishable from a missing artifact.
Questions people ask
- What is evidence annex?
- A single indexed register that points to every governance artifact an organization has produced and, for each, records what it is, what it proves, who made it, when, where it authoritatively lives, and which system and obligation it maps to. It is a map to the evidence, not a warehouse of copies or a narrative essay, and it is the register the Module 13 audit reads from.
- What is paper trail (versus scramble)?
- The state of being able to follow a register straight to authenticated evidence on demand, as opposed to hunting through folders and hoping to find a document you cannot confirm is genuine. The paper trail is provenance you can follow; the scramble is provenance you are reconstructing while someone watches, and the scramble itself is a finding.
- What is provenance?
- The account of an artifact's origin and life: who made it, when, what version this is, what changed between versions, and a way to show it has not been altered since. Provenance is what turns a document into evidence, because it lets you answer "who made this, and how do you know it is genuine" without saying "trust us." More on Provenance
- What is chain of custody?
- The unbroken, accountable record of who created and handled an artifact and of its integrity over time, so that its authenticity can be shown rather than asserted. Version control provides most of it for free; outside version control it is approximated with dated signatures or stored file hashes. More on Chain of custody
- What is file hash?
- A short cryptographic fingerprint of a file that changes if a single character of the file changes, letting anyone check that a stored artifact is the same one that was recorded, without trusting anyone's word. One practical way to make an artifact's integrity verifiable.
Keep going
This lesson builds Evidence collection and audit-ready documentation, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
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 →