[SCITT] Re: Closing omission from the receiver's vantage — what a record must carry

[[e-mail]]

All,

I said I would come back with the check and what it caught rather than with a plan. It is written and merged; this is what it caught. Walter's rung 3 question gets a measured answer at the end, because it turned out I could check it rather than argue it.

The problem is in -05 in my own words: four sentences in the Implementation Status section of -04 were accurate the day they were written and false within days, every one of them claiming less than the code did, and nothing in the repository compared that section to the tool. They were found by a reader holding the document beside the output.

The check runs the section instead of reading it. Three declarations of one fact exist: the code, the prose, and now a claim table between them, because prose cannot be executed. Each behavioural sentence is bound to a run of the shipped CLI over a named fixture, and two directions are enforced. Every value a run produced must still appear in the sentence that states it, and the number of sentences must equal the number of claims. A sentence with no claim is one nothing measures. A claim with no sentence is a measurement the document dropped.

The revision it reads is derived from what is in the repository rather than named in the check, because a hard-coded revision is the same class of stale declaration the check exists to catch.

I showed it failing three ways before wiring it in, since a check that has only ever been green demonstrates nothing:

    a value edited out of a sentence

      -> the sentence no longer contains observed-without-receipt

    one bullet removed

      -> states 4 observed behaviour(s), the claim table runs 5

    a fixture changed so the tool's outcome moved, sentence untouched

      -> item outcomes are ["excluded","indeterminate"],

         the table says ["excluded","observed-without-receipt"]

The third is the one that matters: the code moved and the sentence stood still, which is what happened four times in -04.

What it caught immediately is a sentence of mine. -05 says "The implementation's test suite contains no check that compares this section against what the code does." From the merge commit that is false, and a posted draft cannot be edited, so it is the first correction -06 owes.

Two limits, stated now rather than discovered later. It pins the behaviour table and not the prose around it, so a paragraph can still go stale in a way nothing runs. And a red has two honest fixes: change the code back, or write the revision that says what the code now does. Editing a posted draft is not one of them, which makes the cost of drift a document rather than a diff.

Walter, on rung 3. You put it as: an external anchor is not automatically rung 3, an anchor is a claim by whoever paid for it, and a profile that says "anchored" without saying "and re-derived" has described rung 2 with extra steps. I went and looked rather than answering from memory, and the implementation agrees with you, which I did not know for certain before you wrote it.

--anchor-check deserialises the timestamp proof, refuses a proof whose digest is not the chain head, walks the attestations, and where the attestation is Bitcoin it fetches that block and compares the digest against the merkle root. So the re-derivation is in the path and not in the prose.

The part I would not have thought to defend, and which your framing is the reason to mention: it separates two answers that a single exit code would have merged. A proof that fails to verify exits 14. A proof that could not be checked at all, where the block could not be fetched, exits 15, and 15 is evaluated first, so "I could not check" never gets swallowed by "it failed". Returning 0 there would have been the same error in the other direction: the caller has to know the anchor was not verified, not merely that it was not disproved. That is the weakest-rung rule applied to our own tool, and it was written before anyone asked us for it, which is the only reason I can say it now without it sounding convenient.

Emek

On Fri, Aug 21, 2026 at 2:50 AM Walter Hawkins <[e-mail]> wrote:

Henri, Joel, Emek, Pablo, Nenad, Todd,

Two corrections of my own before the part I owe this thread, since that appears
to be the going rate here and it's a good rate.

**One.** I used the words "omission-proof" in my own notes and it was a rung
short. Our spend log is chained and gap-free under a signed grant, and the head
that seals it is counter-signed by a named accountant carrying a root and a
running total. That much I had right — it's rung 2. What I had wrong was the
floor. A verifier arriving for the first time, holding no prior state and doing
no anchor lookup, accepts an old validly-signed head together with the prefix
that matches it. Monotonic sequence, perfect arithmetic, short set. Henri's
correction this morning is the same shape as mine, and I'd made mine a day
earlier without noticing what it cost.

Fixed the way Emek's rule says it should be: the freshness floor is now a
required argument, and passing nothing is an explicit downgrade whose verdict
is labelled tamper-evidence rather than completeness. The weakest-rung rule
applied to an argument I had made optional.

**Two, and this is the one I think is worth the thread's time, because it's a
hole in rung 3 itself.**

We put our heads on-chain: a zero-value transaction whose calldata carries the
head digest, the grant it covers, the sequence, and the key that signed it. A
first-contact verifier scans for those and takes the highest sequence as its
floor. That is the external anchor the ladder asks for, and two days ago I'd
have told you it put us at rung 3.

Anchoring is permissionless. That property is why it resists censorship and it
is also the attack. Anyone can write that calldata. The key field is plaintext,
and the accountant's key is public because it's named in the grant. So a
stranger writes an anchor for my grant at sequence 999999999, my scanner reads
it, and every honest head I produce afterwards is refused as a rollback. The
floor is poisonable by anyone willing to pay for one transaction, and the
failure mode isn't a missed omission — it's an honest record permanently
refused. We had traded a false negative for a false positive against ourselves,
which is the worse of the two trades.

The fix is that an anchored sequence cannot be read as a fact. It is a pointer.
A verifier deriving a floor has to fetch the anchored bytes, recompute the
digest over them, check the signature, and check that the head it just verified
names this grant and this sequence — and only then let it move the floor. An
anchor whose content can't be retrieved simply isn't floor-eligible, which can
only lower the floor toward the tamper-evidence downgrade, never raise it.

Stated honestly about where we are with it: that rule is written into our spec
text and our scanner does not yet do it. What the scanner does today is the
cheap half — it ignores anchors that don't name the grant's accountant, which
closes the stranger-with-no-key case and leaves the forged-key case open,
because that field is unauthenticated. So take the paragraph above as a claim
about what rung 3 requires, not a claim about what I've shipped. The
re-derivation is the next thing I write, and I'll come back with what it caught.

So for the statement I'd put it this way: an external anchor is not
automatically rung 3. An anchor is a claim by whoever paid for it. Rung 3 is
an anchor plus the verification that recovers what it points at, and a profile
that says "anchored" without saying "and re-derived" has described rung 2 with
extra steps. Nenad, I think this bears directly on precedence — ordering a
revocation strictly against a settlement only means something if both anchored
objects are re-derived first; otherwise you've ordered two assertions.

**What I owe: the rail enumeration.**

Every payment under a grant carries a binding in the one payer-chosen slot the
rail's own signature covers — the EIP-3009 nonce, the Permit2 nonce, the XRPL
InvoiceID. The value is a domain-separated digest over the grant digest and the
payment identifier, so it's derivable by anyone holding those two and
recomputable from the settled transaction by anyone holding that.

The receiver's use is the part this thread wants. A payee holding a settled
transaction recovers which grant and which payment it commits to from the
artefact itself, without asking the executor which authorisation it would like
to be judged under. And the enumeration runs off the rail rather than off a
list: the payee walks settled transactions to their own address, not a set the
executor hands them. A leg the executor never wrote down is still on the rail
if the money moved, and the party holding it is the party who cannot be made to
un-know it arrived.

Testnet EVM rail, as promised: Coston2. A head anchored as a real transaction at
block 34259963, the calldata read back off the chain and matched against the
signed head, block timestamp doing the ordering.

With a caveat that belongs in this thread more than most. That anchor predates
the hardening I described above — its commitment has no accountant field and its
calldata has no key segment, so it verifies against the code as of the commit
that produced it and not against current HEAD. I keep it as evidence that the
wiring worked on the day, and as a small lesson I'd pass on: we anchored a shape
we were still editing, and permanence is the one property an anchor has whether
or not you were finished. New heads use the hardened shape. I'll bring both,
labelled.

Two limits I'd rather state than have found.

The rail has to enforce slot uniqueness for once-per-payment to come free.
EIP-3009 and Permit2 consume the nonce, so it does. XRPL does not enforce
InvoiceID uniqueness — two Payments carrying the same InvoiceID both settle —
so anyone deriving a cumulative total from XRPL evidence has to de-duplicate by
(grant, payment) and cannot lean on the rail for it.

The second one I had wrong until a review pulled it apart yesterday, and it
generalises past payments. The binding identifies the grant; it does not bound
the amount. The slot commits to the grant and the payment identifier and
nothing else, so an agent holding the payer key can settle a million against a
payment it reports as one. The cap only binds if the verifier reads amount,
recipient and asset out of the settled transaction and compares them back to
the grant. I had a document claiming one artefact proved both that money moved
and that it moved within authority, and that sentence was doing work the
mechanism didn't do. Identifying which authorisation governed an act is not
the same as bounding the act performed under it — Nenad, that's your item 2
from the other end, and I think it's a second thing the substrate should say
outright, because the failure is invisible and every incentive points at it.

The object model and its vectors are public as of today, as an open PR rather
than anything settled: specs/extensions/authority.md and
authority-vectors.json in x402-foundation/x402 #3220. If the statement cites
components normatively the way Henri suggests, that's the published home for
the rail-binding piece, under the same one-authority rule Nenad named.

**Nenad's fourth vantage.**

I think it's right and I'd add a note from our side. We have no revocation, and
that's deliberate rather than missing: our grants carry an expiry, and authority
decays without anyone needing to deliver anything. A suppressed revocation is
invisible to a receiver, as you say — but you can't suppress the passage of
time. It's a real trade, not a free win: a short expiry means re-issuance
traffic and a live principal, and a long one means a wide window where your
attack is exactly as bad as you describe. Where re-issuance is affordable,
expiry is revocation you don't have to deliver, and I'd rather the substrate
name that as an option than have every profile carry a revocation channel it
can't guarantee reaches anyone.

**Joel's fourth item.**

Same answer as Henri, Pablo and Nenad: we don't have it. Nothing in our records
pins the publisher-assigned version of any corpus a decision was computed
against. Four for four is a better argument for lifting it into the substrate
than any of us adding a field and then citing ourselves.

**Willing, and here's what I bring.**

The rail-enumeration binding above, the Coston2 fixtures, and the rung-3
refinement — anchor as pointer, not fact. Henri, the completeness component and
its rungs stay yours to state; Todd, the fraud side is what keeps this attached
to a real failure; and if Emek's ladder is going to carry the structure of the
document, it should carry his name in it.

Walter

On Thu, Aug 20, 2026 04:53 PM, [e-mail] wrote:

Joel, Henri, Pablo, Walter, Todd, Nenad,

Joel, you withdrew two claims in public before making your argument. I will put something of the same kind on the table before making mine.

The corollary you name is the half I have not done, and I can show you exactly where the line falls.

You wrote that the vocabulary has to reach the output, and that both our implementations skip it. In mine the /2 result carries a bounds object: for each bound the comparison relied on, its source is protocol-defined, measured, operator-declared, or undeclared, and a result whose bounds are operator-declared states an outcome of that standing and no stronger. Measured against the published package, one fixture, a receipt three seconds outside a two-hour window:

- no profile: all three bounds undeclared, both items indeterminate

- a profile with no clocks member: multiplicity and exclusion operator-declared, skew still undeclared

- clocks.skew declared wider than the offset: operator-declared, and the item stays indeterminate. A declaration does not manufacture a match.

- clocks.skew declared narrower than the offset: the same item becomes observed-without-receipt

So the vocabulary reaches that object. It does not reach the exit codes, which predate it and are still not a mapping of it. Your sentence covers precisely the half I have not done, and I would rather that stayed in the record than get answered with a plan.

The three clocks fields you asked for are in -05 with their encoding: clocks.observation, clocks.receipt, clocks.skew, a duration grammar, a rule that an unknown key in clocks is rejected by name, and a rule that the same bound declared twice must parse to the same milliseconds or the run fails rather than picking one. Written to be borrowed, not restated.

On placement. The distinction is in -05 as a security consideration (sec-completeness), written against my own implementation rather than about two of them: a truncated receipt set verifies, detecting the removal needs a quantity from outside it, and whether that quantity reached the verifier independently of the Issuer is not visible in a digest. I support lifting it to the substrate, with one caveat about how it is worded there. The property that matters is not "in-record beats argument". It is that a Consumer cannot tell the two apart from the artefact, so the result has to say which one it relied on. Stated as a ranking it invites an implementation to claim the stronger form without carrying it; stated as a disclosure obligation it does not.

On the weakest-rung rule. Four documents arriving at it separately is a better argument for the substrate than any one of us restating it, and your verdict-side version is the one I would take as the general form, because it survives translation to outputs that are not bounds at all. One caution from the side that has already shipped it: the rule is cheap to write and expensive to keep. indeterminate is only load-bearing while nobody is allowed to average it, and the pressure to fold it into a coverage percentage arrives from outside engineering, after the vocabulary is public and looks like a scoring system.

One more thing worth putting in the record. Your sentence from two days ago, that a gaps section going stale understating an implementation fails the same way as one that overstates, turned out to apply to mine four times over. Each statement was accurate when written and false within days, always in the direction of claiming less than the code did, and nothing in my test suite compares that section against the code. So the corollary reaches further than either of us put it: the problem is not that our implementations skip the vocabulary. It is that neither of us has a check that would notice if they did. That is the gap I would most like the substrate to make expensive to leave open, and it is the next thing I write on my own side. I will come back with the check and what it caught, rather than with a plan.

Emek Can Dogru

Conarium

On Fri, Aug 21, 2026 at 12:06 AM Joel Hillier <jhillier=[e-mail]> wrote:

Hi Henri, Pablo, Walter, Todd,

Emek, I'm adding you to this because the rung ladder is yours and it's been travelling without your name on it. Henri's message this morning credits you with cloning at befdced and running the three cases, which is true, but you're also the one who wrote the ladder and the rule that goes with it. That rule is the most useful thing in this thread, and the first thing I did with it was turn it on my own document.It cost me two claims I made to you all yesterday. Both are below, before the rest.

Correction one. I told you ARP's sweep chain makes a withdrawn Statement detectable. That's true for the middle and false for the tail.

Each Evaluation Sweep Statement carries the previous Statement's notarisation, so I had the chain doing a job it can't do. Your 0.2.4 shipped under "a hash chain is blind to a shorter tail" and it applies to my series exactly as it applies to yours: a truncatedtail of Sweep Statements verifies precisely as it did before.

What actually puts ARP at rung 3 is a different mechanism, which I'd been crediting to the chain. Each Statement is due inside a bounded interval measured from its trigger, and where the trigger is a public event with a public timestamp, a sanctions list publishinga delta starts a clock the operator doesn't own. An operator with no Statement inside the interval hasn't evaluated in time, and anyone can check that holding none of the set. The chain covers the middle. The deadline covers the tail. Two mechanisms, and I'dhad them credited as one.

Correction two, and this one is worse, because it's the sentence where I congratulated myself on not rounding up.

I said the falsifiability argument holds for three of ARP's four triggers, "stated rather than rounded up." Only one of those three rested on anything outside the reconciliation server. The other two rested on a transition list published by the server, underthe server's own key, in a document with no publication time, no notarisation and no chain. A transition simply left out of that array started no clock, and no party could date the document well enough to show that it had been. On that footing the count wasone in four, and the sentence claiming otherwise was the one place the document rounded up.

It's three in four now because I fixed the mechanism rather than the sentence: the document carries a publication timestamp, is notarised on the ledger-head interval, and has to be republished on that interval whether or not anything in it changed. That lastpart is the one that matters, and it's your placement argument doing the work. Without it, an operator that removed a witness entry and an operator that changed nothing publish the same thing, and the notarised series has no entry to be missing.

The claim is now conditional and the document says so: a deployment whose policy parameters aren't anchored that way has one falsifiable trigger, not three. Which is your rule about naming the ground, applied to a count I'd already published.

One more, since it's the mechanism I described to Walter yesterday. I said ARP's witness quorum counts entries under common control once, because two instances of one observer aren't two observers. The distinctness test was written over the operating-partyidentifier alone and said nothing about the keys. Two entries declaring the same key under two different party identifiers satisfied a quorum of two with a single signature. Both entries verify, the identifiers are distinct, and a relying party doing exactlywhat the document said counted one signer twice. Now fixed by requiring the verification method references to be pairwise distinct as well.

Worth separating that from the limit I did state correctly. Whether two named parties are genuinely independent can't be tested by a verifier and the document concedes it. Whether two entries name one key can always be tested, and wasn't.

Now the thing I owe this thread.

Henri and Pablo have both said they'd need to add the fourth item before they could point at one, and Henri's suggested the substrate cite components normatively rather than restate them. Together that means the fourth item gets lifted out of ARP. So let mehand it over along with what was wrong with it, rather than after it's in shared text.

A Partial Attestation carries a Source-Data Version Identifier Set: one identifier per source consulted in evaluating the predicate. Each is a tuple of the list name as declared in the bilateral agreement, and the state identifier the list publisher assignsto that state, not a value the answering party made up. It rides in the protected header so it's covered by the register's signature, and it's carried into the output so it's covered by the sealing signature.

The reason it's the publisher's identifier and not the register's is the part worth carrying into any shared text, because it isn't obvious and it's the whole point. A register-chosen opaque string would be an arbitrary-bandwidth channel from register to relyingparty, travelling under signature into a sealed and ledgered artefact, and the accompanying rule that differing identifiers mustn't be read as disagreement would normalise it. Taking the identifier from the publisher's own state sequence is what makes it evidenceinstead of a side channel.

Three things were wrong with the encoding, all found yesterday, all now repaired:

1. The Set had no ordering rule. Five other collections in ARP carry an explicit bytewise sort. This one was called a Set in its own section heading, was carried into two signatures, and was never sorted, so a register consulting three lists had six conformingencodings of one attestation. Now sorted.

2. The state identifier was disjunctive with no discriminator. "A published version token, or a digest of the published corpus where the publisher assigns none." Text in one branch, digest bytes in the other, same position, no algorithm named, and no statementof what the corpus is as a byte sequence. It now carries an explicit form discriminator, and the digest branch is taken over the octets the publisher serves, before any decompression. That second part is the one I'd have got wrong: a list published as a compressedarchive has at least two byte sequences with an equal claim to being the corpus, and two registers choosing differently produce two identifiers for one state, which a retroactive sweep then reads as a version change that never happened.

3. It was carried twice with no equality rule. ARP has exactly the right sentence for this, that a value carried twice with no equality rule is a value an implementation may read either way, and applied it to the two other duplicated headers and not to thisone.

So the shape and the reason are worth lifting. The encoding is worth lifting as of this morning and wasn't yesterday.

On placement, Emek, your distinction deserves to be a named property of the substrate rather than a remark about two implementations.

You put it as: Henri's seal is a record inside the stream, signed and carried with the evidence, readable by whoever holds the set; Conarium's pin is an argument to the verifier, so a third party handed only the receipt set can't tell it's short unless someonestates what it should have been, and the obvious someone is the issuer, the party the audit is about.

That's sharper than the rung number alone, because two mechanisms can sit on the same rung and differ entirely in who has to be asked. I'd state it as three placements:

Inside the evidence. Travels with the set, needs nobody. Henri's seal.

Supplied by the audited party. The verifier has to be told the expected count or terminal hash, and the party best placed to tell it is the one under audit. Conarium's pin, and you said so yourself rather than letting the symmetry flatter you.

Supplied by a party with no stake. ARP's deadline is this one: the clock is started by a list publisher who's never heard of the reconciliation server.

The third is strongest and least available, because it only exists where the obligation happens to be triggered by something public. It isn't a design choice you can make freely. Where it exists it should be used, and where it doesn't the statement should saywhich of the other two you're on.

The sweep found a fourth case that the three-way split predicts and I hadn't looked for: a mechanism sitting at the first placement whose evidence is only obtainable from the second. ARP's head consistency statements are signed by independent witnesses, whichis the whole point of them, and the document specified the artefact, the quorum arithmetic and the freshness window and specified no channel by which a relying party obtains one. The obvious implementation is that the responding service hands them over withits response. Every check passes. The witness signature stops the service forging a statement and does nothing to stop it choosing which ones to pass on, so under exactly the fork the mechanism exists to detect, each branch's reader gets the witnesses thatbranch was fed. Now fixed by requiring witnesses to publish on their own origins and forbidding acceptance of one obtained from the responding service.

Worth adding to the substrate as a rule in its own right: state where the evidence is obtained, not only where it is produced. An artefact produced at the third placement and delivered by the second is at the second.

Your weakest-rung rule already exists in ARP under another name, which I think argues it's the right general rule rather than a local convention.

You wrote that a result which doesn't say which rung it stands on has to be read at the weakest one, and that it's Walter's completeness ladder one level up.

ARP arrived at the same rule from the verdict side and calls it verdict re-typing. An answer over a register whose attestation can't be verified doesn't produce a weaker match, it produces indeterminate. Where a re-notification can't say whether a change came from policy or from the underlying corpus, it has to carry attribution-indeterminate and name which causes were examined and why neither could be excluded, because a bare qualifier is a discretionary escape signed by the party that benefits from it. And conformance run records carry a does_not_establish field so the things a run didn't prove are enumerated rather than inferred from silence.

Four instances, four documents, arrived at separately: bounds in yours, populations in Walter's, verdicts and conformance claims in mine. Worth stating once in the substrate, roughly as: every result names the ground it stands on, and a result that names none is read at the weakest ground available to it.

The corollary is the one implementations skip, including both of ours. The vocabulary has to reach the output. You say your exit codes predate the vocabulary and still aren't a mapping of it. I found two of the same shape this week: a verifier told to refusea proof with no defined way to say which of four refusals it made, and one reason code carrying five distinct causes because four other sections pointed at it for conditions its own definition never named. Both now split. Defining the distinction and givingthe implementation no way to express it is a document agreeing with itself.

On your open question about the boundary declaration.

You framed it well: a boundary inside the record has nothing to drift from and is the stronger guarantee, but it's asserted by one party at issue time; a profile is the weaker binding and the only one that can be agreed between parties before a run and pointedat afterwards by both.

ARP's answer is to make the profile a signed bilateral instrument with a hash. The agreement is settled before any run, both parties compute an agreement hash over its declared items independently and have to get the same value, and every reconciliation commitsto that hash. Drift suspends reconciliation rather than producing a weaker answer. So it keeps the agreed-not-asserted property of a profile and gains the nothing-to-drift-from property of an in-record boundary.

The cost is that it only works bilaterally. It doesn't reach a population of parties who have never negotiated, which is most of Walter's setting. Worth having in the statement as one of the answers rather than the answer.

Pablo, your precedence point generalises, and it flips direction depending on what's being anchored. Yours is anchoring_precedence: the anchor has to be strictly earlier than the outcome it covers, because an anchor arriving after the fact proves nothing about what was true when the outcome was recorded. ARP's is the mirror image. The Statement has tofall later than the trigger and inside a bounded interval, because what's being proved is that a duty was discharged on time, not that a fact was true beforehand.

Same underlying requirement, that an anchor with no ordering constraint against the thing it covers is decorative. Two opposite orderings, because one anchors the proof and the other anchors the duty. I'd say it that way in the shared text: every external anchordeclares its required ordering relative to what it covers, and which of the two it is.

Agreed on normative citation over restatement, with one addition. Where a component is cited rather than restated, the citing statement should also name the rung the component reaches. Henri's correction this morning is the case in point, and so are both ofmine above. None of the three components changed. The claim made about each was one rung short. A substrate that cites a component and repeats an over-broad claim about it has moved the error rather than removed it.

Joel