ZAMBO.

Direct answer

Which parts were AI, which parts were you? Answer with the delivery record, not with your memory.

The question sounds simple and arrives at the worst possible time: weeks after delivery, in a thread about something else, when the client asks which parts of the work an AI did and which parts you did yourself. Your honest memory of step level authorship is already gone. You remember the job, the hard parts, the late fix, but not which of forty steps came from which hands, and any answer you compose now is a reconstruction that sounds more certain than it is. Guess generously and you overclaim your own work. Guess cautiously and you undersell it. Both guesses share the same flaw: they are essays written after the fact about a question that had a factual answer at the time. If the work ran on a delivery record, the answer is not an essay. Each step of the record carries the label of who or what produced it, recorded while the work ran, so the answer you send is the record itself, with the labelled steps readable one by one and the unlabelled ones honestly marked. Here is how that answer works, and how to handle the delivery that ran without one.

See the instrument first: the live demo runs one real call and hands you its verifiable receipt, no signup.

Updated November 7, 2026. First published November 7, 2026. The two delivery records quoted below are real and public, with their client decision states read back from the live service on November 7, 2026.

Step 1. Notice that the question is factual, and memory is the wrong instrument

Which parts were done by AI is not a matter of positioning, and treating it as one is what turns a routine question into a dispute. The client is usually asking for a practical reason: they need to know what they can reuse, what they must disclose onward, what a reviewer should check hardest, or simply what they paid for. A memory-based answer fails them even when it is honest, because it cannot be checked, cited, or passed to the person who actually needed it. It also fails you later, when your recollection and theirs drift apart and the essay you wrote becomes the disputed document. The standard to aim for is the one other delivery questions already meet on a good record: do not describe the work, point at it. Authorship at step level was a fact when each step ran. The only question is whether anything wrote it down at the time.

Step 2. Read what a labelled step record actually carries

On a handover record, every step carries a producer label alongside its saved result, its time, and its verifiable receipt. The label is producer-stated: it names the AI or the producer as recorded when the step ran, in the open, and steps that arrived without a label read Unspecified AI rather than being quietly assigned to someone. A real mixed record reads exactly that plainly. One public handover in this system, 3ba7cb75, carries a step labelled Claude, a step labelled Grok, and a step labelled Unspecified AI, each with its own receipt, and the job was accepted by its receiving agent with acceptance receipt 1600adb6. Nothing in that record is a claim about the job as a whole. It is three small facts, one per step, each checkable where it sits, and together they answer the client's question at the grain the question was asked: this part, that part, and that part nobody labelled, stated as unlabelled rather than smoothed over.

Step 3. Send the record, and walk the client through three piles

The answer you send is one link and one short message. The link is the delivery record itself. The message sorts its steps into the three piles the client actually asked about: steps labelled with a named AI, steps that are yours, and steps marked unspecified. Keep the message factual and short, name the counts, and stop there. Do not apologise for the AI steps, do not inflate the human ones, and do not editorialize about which steps mattered most. A client reading a labelled record is watching you be exact about your own work, which is the entire credibility of the exchange. Where a step's label matters to a decision the client faces, such as a part they intend to publish, offer the step's own receipt link so they can check that step directly rather than taking the summary on faith. The summary is yours; the record is checkable; sending both is what separates an answer from an assurance.

Step 4. Read the decision states that sit on top of the labels

Labels say who produced each step. The record also carries whether the client accepted those steps, and the two facts belong in the same conversation. The record format is AER-1, published as an IETF Internet-Draft (draft-zambo-aer1-26), and each step carries a verifiable receipt: a receipt proves the saved result was not changed since it was saved. The two public records below show the decision states a labelled delivery can sit in:

Accepted: the decision is on the record

Pending: checked, and still waiting on a decision

The accepted record, 4 steps, 4 checked, decided_by client at 2026-10-16T14:31:51.146Z, release_hint ok_to_release, shows the settled shape: every step checked, labels readable, and the client's acceptance recorded against that exact record, so a later authorship question is answered by reopening a document the client already decided on. The pending record, 5 steps, 5 checked, release_hint hold, shows the live shape: the labels are just as readable, and the record is honest that acceptance is still outstanding. If the authorship question arrives while a delivery sits pending, answer it from the record the same way, and let the answer carry its state with it: here is which parts were AI, here is which were mine, and here is the acceptance that is still yours to make. Facts first, decision second, both on the same page.

Step 5. If the delivery ran unlabelled, reconstruct honestly and mark the gaps

Many deliveries predating this question ran with no labels at all, in ordinary chats and files. For those, do not manufacture a labelled past. Reconstruct what can be grounded: files and drafts whose tool of origin is visible in their history, steps a client watched happen, work you can tie to a dated artefact. Present the answer in the same three piles, with the third pile, unknown, as large as it honestly is, and label the whole answer as a reconstruction from artefacts, dated today. Clients accept a bounded answer far better than a confident one that later cracks, and the bounded answer has a second use: it is the natural moment to change the setup, in one sentence, so the next delivery carries its labels as it runs. What you must never do is backfill labels from plausibility, assigning steps to an AI because it probably wrote them. A record with honest gaps is evidence. A record with invented labels is a liability wearing evidence's clothes.

Step 6. Set up the next delivery so the question answers itself

The durable fix is procedural, not rhetorical. Run the AI steps of a job through the recording connection so each one lands on the job record with its producer label and receipt as it happens. Log your own substantial steps into the same job record as your steps, so the human parts are present as steps rather than as the invisible gaps between AI steps. At delivery, send the handover as the delivery document: the client receives the work and its authorship map in the same link, can accept or flag step by step, and can generate the AI-use picture from the labels whenever their own disclosure needs ask. From then on the question, whenever it arrives, is answered by resending a link that already existed on delivery day. You will notice the side effect quickly: a record that can answer which parts were AI can also answer when each part ran, what it returned, and whether it was accepted, which retires a whole family of after-the-fact essays at once.

What a record does not settle

A producer label is producer-stated, not independently proven identity: the record shows the label that was recorded for the step, and the system's own notices say the same, in the form Logged by Zambo on behalf of the visitor. Not identity-verified. A verifiable receipt proves the saved result was not changed since it was saved; it does not prove who truly authored a step, and it does not judge whether AI use on any step was appropriate, permitted, or disclosed on time. Those remain matters of your agreement with the client and your professional judgement. What the record settles is narrower and, for this question, sufficient: what label each step carried when it ran, what the step saved, and what the client decided about it.

Frequently asked questions

My client asked which parts of the delivery were done by AI and which parts I did myself. How do I answer with a record instead of a guess?

Send the delivery record itself, with one short message that sorts its steps into three piles. On a handover record, every step carries a producer label recorded while the work ran, alongside its saved result, its time, and its verifiable receipt, and steps that arrived unlabelled read Unspecified AI in the open. So the answer is: steps labelled with a named AI, steps recorded as your own work, and steps marked unspecified, with counts for each and the link where the client can read every label in place. One public record shows the grain: handover 3ba7cb75 carries a step labelled Claude, a step labelled Grok, and a step labelled Unspecified AI, each with its own receipt. Answer at that grain, state plainly that labels are producer-stated, and do not reconstruct from memory what the record does not carry.

Are the AI labels on a step record proven, or just stated?

Stated, and the record says so itself. A producer label records what the step declared when it ran, and the system carries the notice that entries are logged on behalf of the visitor and are not identity-verified. That honesty is a feature of the answer, not a flaw in it: the client gets the exact status of the information instead of a confidence performance. What is checkable on each step is different and stronger: the saved result behind its verifiable receipt either recomputes as saved or it does not, and the acceptance state shows what the client decided about the step. So the authorship split rests on stated labels, the integrity of each step's content rests on a check anyone can rerun, and a careful answer keeps those two resting places separate and named.

What if some steps of my delivery say Unspecified AI?

Leave them unspecified, and count them openly in your answer. An unspecified step means the work arrived without a producer label, which happens with steps saved outside a named account or handed in without one. It does not mean the step was secretly yours or secretly an AI's, and assigning it either way after the fact would invent exactly the fact the client asked about. In the three pile answer, unspecified is its own pile with its own count, and on a mixed record it is normal: the public handover 3ba7cb75 carries one Unspecified AI step beside its Claude and Grok steps. If the client needs that pile smaller on the next job, the fix is procedural: run the steps through the connection under named producers from the start, so labels are recorded as the work runs instead of debated after it.

What do I actually send the client when they ask this question?

One link and a short factual message. The link is the delivery record, the same handover you delivered against, so the client reads the labels where they sit, step by step, with each step's receipt and the record's acceptance state. The message names the three piles with counts: how many steps carry a named AI label, how many are recorded as your own work, and how many are unspecified, and it states in one line that the labels are producer-stated as recorded at the time. Attach nothing else. Do not add a defence of the AI steps or a story about the human ones; the record's exactness is the whole reply, and it travels well: the client can forward the link to a reviewer, a colleague, or their own compliance contact, and every reader sees the same labels you sent, which no summary email can promise.

How do I make sure the next delivery can answer this question on its own?

Run the job so its steps are labelled as they happen, and deliver against that record. When the work runs through Zambo, each AI step lands on the job record with its producer label and a verifiable receipt under AER-1, published as IETF Internet-Draft draft-zambo-aer1-26, your own steps can be logged into the same job, and the handover you deliver carries the labels, the acceptance state, and the material for an AI-use statement generated from those labels. New users get 20 calls and 1 wrap-up free every day, and picking up work in another AI is always free. For $0.99, deliver the next job that way and let the authorship answer assemble itself while the work runs: get the Day Pass. A Day Pass is unlimited for 24 hours, no subscription. Ongoing use is the Zambo Pass at $29 a month or $290 a year, and $19 a month for $ZAMBO holders, which adds private handover links, your branding, and the approval record as PDF and JSON files.

Related