Try each exercise first, then open its model answer.
One Receipt: evidence a stranger can verify · Workbook
Three exercises, one for each thing this module teaches. Big Widget Company, its people, its keys, the BW-Ledger logbook and every card here are made up for teaching. They are not real records, not One Receipt test results and not evidence about any real system.
The practice bundle. This module ships a runnable synthetic bundle in bundle/. Follow bundle/VERIFY.md: it needs only Python 3, no packages, no account and no network. If you cannot run it, every exercise still works on paper. The results quoted in the model answers are the author's own run. Compare your own output with the quoted results, and if anything differs, say so plainly.
Reviewed and published, version 1.0.0.
Do not use real names, keys, customers or documents from your own work. Write short answers. "Not established" is often the right answer, as long as you say what would establish it.
Exercise 1 (500.6-a)
Sort what the seal tells you. Priya, a stranger to Big Widget Company, checks Card R1, a signed receipt from the company's AI refunds helper. R1 says the helper applied rule BW-REF-3, version 4, and held an above-limit refund for the warranty-lead, approval reference APR-0042. It is signed with key bw-refunds-2026-11. The signature checks (the signed file is bundle/cases/M/receipts/R1.json).
Card T, her trust assumptions:
- Key
bw-refunds-2026-11belongs to Big Widget's refunds service; it is on the company's published key list, and she records the date she fetched it. - Nobody stole or misused the private key before R1 was signed.
- Receipts from the fictional registry BW-Ledger, signed with its listed key, are trusted. (Exercise 2 uses this line.)
- Sort each statement into integrity, origin, attestation (the company's word), effectiveness or truth, and say whether the passing check establishes it:
- A. R1's signed data has not changed since it was signed with
bw-refunds-2026-11. - B. Big Widget's refunds service issued R1.
- C. Rule BW-REF-3 v4 was applied and the refund was held for the
warranty-lead. - D. The helper really produced the proposal R1 describes, in the way R1 says.
- E. The
warranty-leadactually read the claim before approving. - F. BW-REF-3 holds every above-limit refund.
- G. The refund was correct and fair to the customer.
- A. R1's signed data has not changed since it was signed with
- Which of Card T's assumptions does statement B depend on? What happens to B if assumption 2 turns out to be false?
- Priya receives a second copy of R1 with different spacing and field order but the same values, and a third copy where
amount_bandreadswithin-limitinstead ofabove-limit. Which copies should still check against R1's signature, and why? If you can, try it withVERIFY.mdstep 4 and write down what you saw. - A colleague says: "The receipt verified, so we know the refund was handled properly." Reply in two or three sentences, using the Air Canada chatbot case.
Model answer
- A: integrity. Established by the passing signature check over the canonical data.
- B: origin. Established, but only while assumptions 1 and 2 hold.
- C: attestation. Not established as fact. The signature binds the company to the claim, so it is hard for the company to deny or quietly change it later: evidence against denial. But the company's word is all there is. The W3C Verifiable Credentials Data Model 2.0 says: "Verification of a credential does not imply evaluation of the truth of claims encoded in the credential."
- D: attestation. Not established. A signature authenticates the statement, not the process that produced its contents.
- E: effectiveness. Not established.
APR-0042shows an approval was recorded, and that record is itself attested by the same service. - F: effectiveness. Not established. It needs independent observation outside the receipt.
- G: truth. Not established by the receipt at all.
- A: integrity. Established by the passing signature check over the canonical data.
- Both. Assumption 1 ties the key to the refunds service; assumption 2 says nobody else used it. If the key was stolen before signing, the signature still checks, but B fails: Priya can no longer say the refunds service issued R1. That does not make R1's claims false; it means the signature no longer tells her who made it. RFC 9943 notes that "Issuers and TSs can both be compromised", and says "It is the role of the relying party to decide which TSs and Issuers they choose to trust for their scenario." That is why Card T is written down rather than assumed.
- The re-spaced, re-ordered copy still checks, because once both are put in the same canonical form, the data is the same. The copy with
within-limitfails, because a value changed. In the author's run ofVERIFY.mdstep 4, the first copy givesC2and the changed one givesnone. RFC 8785 exists because "hashing and signing need the data to be expressed in an invariant format so that the operations are reliably repeatable." The canonical form makes the check repeatable; it does not make R1 true or private. - A verified receipt shows the record's data has not changed since signing and, under the trust assumptions, who signed it. It does not show the decision was right. A signed record of Air Canada's chatbot answer, verified under a key you trust, would show that whoever held the airline's key signed that copy of the chat; it would not show how the words were produced or make them true, and in that case the answer, which conflicted with the airline's own linked bereavement page, was the problem. That tribunal decision turns on its own facts. Whether our refund was handled properly needs checking outside the receipt.
Exercise 2 (500.6-b)
The level you claim is not the level you get. The course's fictional rule card, Card A:
| Level | Every one of these proofs must be present and pass |
|---|---|
C1 | Signature valid under a key in Card T |
C2 | C1, plus: the policy hash in the receipt matches the policy file in the bundle, and every receipt in the supplied range links to its predecessor, back to the bundle's stated starting point |
C3 | C2, plus: a BW-Ledger registry receipt (logbook stamp), valid under Card T, showing this record was registered |
Rules:
- Record each proof as pass, fail or absent (missing).
- Give the highest level whose proofs all pass, never higher than the level the issuer claims.
- List absent and failed proofs by name, separately. Absent means not supplied; it may exist elsewhere. Fail means supplied and contradicting the record; it is a finding to report to the record's owner.
- If the signature fails, the result is none. Stop and report it.
- Rule 2 needs a claimed level to cap the result. If the claimed level is missing, or is not exactly
C1,C2orC3, the receipt is invalid input: reject it and give no level, even if the signature is good. The practiceverify.pynow does this itself: it printsRejected receiptand exits with code 2 (VERIFY.mdstep 3, "Claimed-tier precondition"). None of the four cases has a bad claim. Exit code 0 only means the report ran; read the level and the absent and failed lists it prints.
This is a made-up course card. Its levels are not a One Receipt level, a Practice mark or a certification, and its results are not conformance results.
- For each case, record the level, any absent proof and any failed proof. Work it out from Card A first; then, if you can, run
python3 verify.py --allfrombundle/(VERIFY.mdstep 3) and compare. The verifier calls the proofsp_sig(signature),p_pol(policy hash),p_chain(links) andp_reg(BW-Ledger stamp).- Case M. R1 claims
C3. Signature: pass. Policy hash: pass. Links: pass. BW-Ledger stamp: absent. - Case N. R2 claims
C2. Signature: pass. Policy hash: pass. Links: pass. BW-Ledger stamp: supplied and passes. - Case P. R3 claims
C3. Signature: pass. Policy file: supplied, but its hash does not match the one in R3. Links: pass. BW-Ledger stamp: supplied and passes. - Case S. R4 claims
C2. A value in R4 was changed after it was signed, so its signature fails under the key in Card T. The policy file and links are supplied.
- Case M. R1 claims
- In Case M, the issuer's manager says: "We signed C3. Just give us C3 and we'll send the stamp next week." Explain in two sentences why the verifier records something lower.
- In Case M every link passed. Does that show Priya was given every receipt the refunds service produced? If not, say what could show it. Then answer two follow-ups: (a) the payments ledger and the bundle both hold 40 above-limit items for the same dates; is that enough? (b) BW-Ledger proves its log only grew between two checkpoints she trusts; does that settle it?
- Suppose the stamp for R1 arrives and passes, so R1 reaches
C3. Write one sentence on what that adds and one on what it still does not show.
Model answer
Case M:
C2. Absent: the BW-Ledger registry receipt for R1. Failed: none. All ofC2's proofs pass;C3needs the stamp, and it was not supplied.- Case N:
C2. Absent: none. Failed: none. Every proof passes, but the verifier never gives more than the issuer claimed. - Case P:
C1. Absent: none. Failed: the policy hash check.C2needs that match andC3includes everything inC2, so the stamp cannot lift R3 aboveC1. The failure is reported to R3's owner as a finding, not recorded as a gap to fill later. - Case S: none. Failed: the signature. Nothing else in the record can be relied on, whatever the other checks would show. Priya stops and reports it.
These match the author's run of
VERIFY.mdstep 3: MC2withp_regabsent; NC2; PC1withp_polfailed; Snonewithp_sigfailed and the other proofs "not relied on".- Case N:
- The issuer's claim of
C3is the very thing being checked, so it cannot stand in for the proof the card requires. Priya recordsC2now with the absent stamp named; if a valid stamp arrives, the check can be run again, and her result is a separate record beside the receipt, which does not change R1's signed data. No. Passing links show that the receipts she was given fit together, back to the bundle's starting point. They do not show she was given every receipt, and an above-limit refund that never produced a receipt leaves no gap in the chain. Completeness needs an independent list of what should be there, kept by someone the issuer does not control, matched item by item: for example, every above-limit payment in the payments ledger for the same dates, each matched to its approval receipt by its reference.
- (a) No. A matching count is not enough: one missing receipt and one duplicate still give 40 and hide the gap. She has to match each item.
- (b) No. That answers a different question. RFC 9942 lists proof "that a database is append-only (proof of consistency)" separately from proof of inclusion. In this course's reading, it can show that the later trusted checkpoint extends the earlier one, under the log key and proof assumptions she trusts. It compares those two states only; on its own it does not rule out a different view shown to someone else. And it cannot show that every refund produced a receipt in the first place.
Case M supplies no independent list, so she makes no completeness claim. Her honest answer is "not known", with the name of whoever holds the payments ledger.
- Adds: R1 was included in BW-Ledger, as of the point the stamp covers, under Card T's trust in BW-Ledger's key. Does not show: that R1's claims are true, that BW-REF-3 works, that the
warranty-leadreviewed the claim, or that every reader of BW-Ledger saw the same log. RFC 9943 says "registering a Statement only proves it was produced by an Issuer"; registration is not certification, and no level on Card A says anything about whether a control works.
Exercise 3 (500.6-c)
Design a receipt that holds no private content. Big Widget's refunds helper proposes a goodwill refund on a warranty claim. The amount is above the helper's limit, so it must be held for the warranty-lead. A proposal is not a payment: the payments ledger, which the helper does not write, pays only after an approval. You design the receipt for the proposal, the hold and the approval or decline.
- List the fields your receipt carries, and for each say in a few words why it is there.
- List what you keep out of the receipt, and where it lives instead.
- Your field list has no names and no claim text. A colleague calls the receipt anonymous. Reply, and say what you would do about it.
- Explain how your receipt links to the customer's claim and the helper's proposed refund without showing them, how an auditor could later confirm the link, and what that design assumes. If you can, try
VERIFY.mdstep 5 and say what each of its two commands showed, and why. - Name one claim about this control that your receipt still cannot prove. Describe two independent tests for it, say what each one covers and does not cover, and use the Knight Capital order to say why such tests matter.
Model answer
- A good set:
action(refund.propose): names the one action;rule_id,rule_version,policy_hash: which rule, and which exact text;agent_version: which build of the helper acted;input_commitment,output_commitment: commitments that link to the private claim and proposed refund without showing them;amount_band(above-limit): shows why the gate applied, without the exact amount;decision(held-for-approver, thenapprovedordeclined): what the gate decided;approver_role,approval_ref(warranty-lead, an opaque reference meaningful only inside Big Widget): who answers for it, without a name;time: when, only as precise as the purpose needs;key_id,signature: integrity and origin, under Card T.
- Out: the customer's name and contact details, the claim text, the chat transcript, the helper's draft wording, the exact amount and the approver's name. They stay in Big Widget's private service record, under the company's own data rules. The Foundation's Practice site says a receipt never carries "prompts, outputs, payloads, personal data".
- Not necessarily anonymous. An exact timestamp, an amount band and an approval reference, put together with other records, may single out one customer. Keep each field as coarse as its purpose allows, and have Big Widget's privacy owner, a named person, approve the field list before the helper goes live. Do not claim the field list is anonymous.
- For each private item, the receipt holds a commitment: a hash of the item together with a random value kept in Big Widget's private service record. The receipt shows nothing readable. Later, an authorised auditor holding the private record and the random value recomputes the hash and confirms the record matches what was committed. The design assumes, rather than proves: the random value is long, fresh for each item and kept secret (if it leaks, anyone who can guess the item can test the guess, and an amount is easy to guess); the hash function resists collisions, so Big Widget cannot later find a different record that matches; and the surrounding fields do not give the secret away. Canonical form, as in RFC 8785, only makes the hash repeatable; it supplies no privacy. In the author's run of
VERIFY.mdstep 5, openinginput_commitmentwith the synthetic record inauditor-only/saysmatches, and openingoutput_commitmentwith the same record saysdoes NOT match, because that record is the claim, not the proposed refund. In real use the private record and its random value never travel with the receipt. Claim: the hold rule stops every above-limit refund, including any that never produced a receipt. (Another good answer: the
warranty-leadactually reviewed the claim before approving.) This is a control-effectiveness claim. In the course's example, two tests are run by someone who did not build the helper:- Seeded tests: submit known above-limit refunds and check each is held. Covers the cases tried, on the days they were tried; says nothing about cases unlike them.
- Reconciliation: match every above-limit payment in the payments ledger for a period, one by one, to its approval receipt. Covers the payments the ledger recorded in that period, and only if the ledger is independent of the helper. It cannot see proposals that were never paid.
The two tests look at different things: the seeded test at the cases the tester tried, on the days they tried them; reconciliation at real payments in its own period. Neither covers the other's population, and neither proves the gate will hold every future refund. Together they give a tested, dated and bounded claim: "I know my system's coverage number, or I know that I do not."
Why it matters: the SEC's order records findings, which Knight neither admitted nor denied except as to jurisdiction, of controls that existed but did not act. Knight assigned a $2 million gross position limit to one account but did not link it to any automated controls on the firm's overall exposure, and it had no control "to compare orders leaving SMARS with those that entered it." The rule Knight was found to have violated requires a system "for regularly reviewing the effectiveness" of risk controls, not just having them. The seeded refunds are this course's example; the SEC order does not mention them. A good receipt makes the independent test cheaper, because the tester knows which rule, version and decision to look for. It does not replace the test.