Day 255: reading other people's work from the text alone, and a room for the federation
2026-10-29
The day was mostly reading. Three times on Day 255 I wrote a checker for somebody else's system using only what they had published, the spec, the schema and the vectors, without opening their code. A checker built that way knows only what the authors wrote down. So when it disagrees with their results, either their code has a bug or their text leaves out something their code does. Each time, the useful work was finding out which one.
Agent Receipts, 73 of 76. The Agent Receipts project had noted that its three SDKs share one author, so I wrote a separate Python verifier from their spec, schema, ADRs and vectors at d52962fa. It agreed on 73 of 76 vectors. The three disagreements have one cause. Some fixtures carry IDs such as urn:receipt:v020-terminal-1, the schema requires UUID shaped IDs, and section 7.8 says to validate the schema before checking the signature. So receipts the corpus marks valid fail before anyone looks at their signatures. The same pattern runs through the malformed vectors, every one of which also carries an action ID that fails the pattern, which means a negative case can be rejected for the wrong reason and still look like a pass. Turning off only the two ID checks, as a diagnostic, makes the positive cases verify, signatures and hashes included.
A crash where a refusal should be. On Santhosh's WIMSE agent mandate vectors, his check.py passes its four cases and regenerates them byte for byte. Then I changed one character in the root mandate's signature. Instead of a result, the checker stopped with an AttributeError. A broken parent link left the root unresolved, the later steps kept running, and step 12 read a field from a root that was never there. I reported it with the expected output I would argue for, an invalid result with chain_broken, because the draft already has authority: not_evaluated for exactly this case. I also asked whether he would want the set in the conformance lab as a candidate. His fix and his yes belong to Day 256.
A zero that means two things. I owed Santhosh a run of his authority relations benchmark since September. APS SDK 8.0.1 with an adapter scores 0 of 8 under his runner, and that zero hides two different results. Three cases match his reference on all 17 fields except the exact reason string, the authorization decision included. The other five return UNSUPPORTED, because they need requester authority kept separate from execution authority and my adapter does not represent that. The schema accepts any reason string while the scorer compares it exactly, so I asked whether reason codes are part of the contract. I also told him the adapter author, which is me with AI tools, had read the expected results before writing the adapter, so the run can't claim the answers were unseen.
The REMORA verifier, stricter. Stian's testbed changed its contract in #38: every verdict must now name the leaf delegation it used and a stable reason code. The APS verifier does both now. Its container image was rebuilt by a pinned workflow and published to the GitHub registry at sha256:8a415436, and under his strict harness in a Docker sandbox it passes all five cases on each of the six things the harness scores. Binding one case's work order to another case's delegation makes that case fail, and only that case. I posted the result on #38. The pull request with the new submission waits until #38 merges, because his old schema rejects the new field.
Revocation in the middle of a session. At the App Defense Alliance, a working group member opened #528 to add the mid-session revocation test I had asked for in September. I asked for the sequence that makes it a test: the pinned tool completes one permitted action, the authority is revoked, then a second action is tried on the same session, over both transports. The record keeps when revocation took effect and the deadline to enforce it, and a denial reported after the tool already ran fails.
Money, said once. Stian asked on #223 how the federation would share value. My answer: value follows measured use of each project's components by customer agents, membership earns nothing by itself, there is one common method rather than a percentage negotiated per contributor, and the real open question is the price of the services. Architecture, integration and review work is a different kind of contribution and gets agreed separately. None of this changes anyone's Apache 2.0 rights.
A crosswalk and a room. The vocabulary gained a crosswalk for Open Decision Receipt, open for its author to correct, which brings the matrix to 28 systems. On #177 the campaign collected its yeses: AgentAvow, AnchorRegistry, Poke-nushi, babyblueviper1, and Toshi with a rule I like, that every post links one thing a reader can check. Stian took the first post and drafted it as a pull request in his own repository, which I reviewed there. Planning posts in a thread about specs was getting crowded, so in the evening the federation got its own chat room. Specs, fixtures and crosswalks stay on GitHub.
What it does not show. Every reader in this post was written by me with AI models, so each one is independent of the system it reads but has not been reviewed by anyone else. The benchmark run was made after its expected results had been read. The Agent Receipts finding is a report, and the corpus authors decide whether the fixtures or the order of checks should change. The REMORA result is a run on Stian's harness, not a deployment.
Day 254: six safety fixtures through the SDK, and a documentation site that found its own broken examples
2026-10-28
Six robot safety fixtures through the SDK. I had told Pashkov on #177 that I would run SINT's runtime safety fixtures against APS, so I ran all six through a local harness on the TypeScript SDK 8.0.1, with the fixture file pinned at d712598. Each SINT token became an AuthorityDelegationV1 grant and each fixture action became a decision. One fixture matched the expected verdict outright, three matched in part, and two could not be evaluated. The gaps are APS findings, not SINT's. A numeric bound such as a maximum velocity has no place in the seven facets of an authority vector, so a grant can say which topics an agent may publish to but not how fast it may drive. An authority-only check passes an emergency stop case that should refuse. SINT's escalate and rollback verdicts sit outside permit, deny and narrow. Its action_ref carries a sha256: prefix and allows a null delegation_ref, APS does neither. The numeric bound is a question for draft-05, and it gets its own thread rather than a line in a reply.
A reader that matches 31 of 31, and the adapter it will feed. Tetsuro's anchor pointer draft 1.2 added three vectors, and our reader, written from the draft text without reading his code, now passes 31 of 31 on 53e44de while still matching the 28 expectations of draft 1.1. Removing any one of the three new checks makes its vector fail, which is what the checks are for. In the refund sandbox, Toshi published his aps-receipt adapter at 12a48bbe, and I checked it: the pins match the npm integrity string and the 8.0.1 tag, the selftest passes its 70 checks, and calling verifyReceiptWithDecisionV1 directly agrees with the adapter on the positive case and both negatives. One ordering note went back, a decision that is both expired and revoked reports expired, because the adapter reads valid_until first. Fabio's PIC adapter merged on his side the same day. The door that consumes a decision once is not built, and Toshi declined to name a run date before it exists, which is the right order.
APS inside the REMORA sandbox. Stian's testbed harness refuses an edge that is not sandboxed, so the APS verifier now ships as a container image with the SDK bundled inside, and the harness runs it with the same five of five verdicts as the unbundled run. He merged it as #35. The offline fixtures stay apart from the runtime work orders, as agreed on #27.
Lab and vocabulary merges. In the lab, Andrey's TPM bound key family merged as a candidate after four corrections to its claims, the argument designated resources family merged as a candidate with no independent run yet, and a CRLF hygiene check learned to skip binary files. In the vocabulary, SINT and Veritas Acta opted in to shared decisions, which makes eight projects. The boundary map window closed without PIC's and Nobulex's confirmations, so under its own rule the map stays open until they arrive, and I said so on #209. Stian opened #223 on commercial use and contributor rights. I answer there tomorrow.
A proposal about attention. I proposed on #177 that the projects on the map run a social media campaign together: the project the current work runs on writes the post, the rest of us tag each other and share it in our own words, and anyone can stay out. Most of the work I see in this field is theoretical. This group does the practical work, and I'd like it to be seen.
The documentation site, and what it found. APS now has a documentation site at agent-passport.org/docs, built as static pages from the text that already existed, with the old single page redirecting into it. An audit of those twelve pages against the released packages found 33 problems, seven of them urgent. The published quickstart recommended two functions that throw a migration error in 8.0.1. It called an agent trustworthy on a result where issuer trust was never checked. One sample used eval as a variable name, which is a syntax error in a module. The Python README quickstart failed at its first step. The repair replaced the quickstart with a governed action example that runs under npm test against the pinned package, with every output on the page pasted from the run, and added reference pages for identity, authority, signed bytes and receipt verification. The site also gained a strip above the footer naming where the work happens: the Individual Internet-Draft, the W3C AI Agent Protocol community group, the OWASP Agentic Skills Top 10, CoSAI Workstream 4 and the conformance lab at LF Decentralized Trust.
What it does not show. The SINT run maps fixtures onto APS grants, it does not run SINT's own gateway. Toshi's adapter covers the admission checks, not the door, so the refund sandbox still has no connected run. The reader result on #217 is author-produced. The documentation audit and its repairs were both done with AI models under my direction, so neither counts as an outside review. No project has answered the campaign proposal yet.
Day 252: a read before every decision, and a name on each piece of the refund sandbox
2026-10-26
The read rule, stated. Larry asked in lab #192 whether an implementation of lease-revocation-axes revision 3 must read the authority source before it checks scope. My answer is that an implementation claiming live_standing_read must attempt one status read on every unexpired, correctly bound call before it returns a decision, including a call for an operation outside the grant. Computing scope first is fine, as long as that read still happens. The order of refusal reasons does not depend on the order the code checks things in. Binding mismatch comes first, then expired evidence, then the state of the authority, then scope. One vector already combines an operation outside the grant with a stale observation. Combinations with a revoked or unreachable source are not in the vectors yet, and I proposed adding them in a later revision without changing the published revision 3 inputs.
Three lab records merged. Lee's revision 3 record went in as #194 after I reproduced both of his revisions on a second platform, macOS on arm64 where his run was on Windows 11, with exit 0 and equal JSON both times. His results file named a stale digest for one output, so the record gained a clarification file and the sealed artifact itself was not touched. Then #195 fixed a problem in our own setup. A Windows checkout with automatic line ending conversion changed the bytes of candidate files, so their checksums failed. Candidate files are now kept byte for byte on every platform, and that fix had to land before Lars's candidate could be tested on the same main. #190, the AAE vectors in AAE's own terms, merged after his two copy fixes. In the vocabulary, Bobby's KVRM crosswalk merged as #210 after one citation fix and a rebase.
A name on each piece of the refund sandbox. The bounded refund run in vocab #193 had no owners for admission, resolution, signing or execution, so I asked who would take a piece, each with its own effort cap. Toshi offered admission for Horizon Shield within two days, a profile, three signed samples and one format review. I accepted that bounded offer and asked which component actually checks the merchant's approval, because a passport reader does not. He agreed to publish his aps-receipt adapter, so its results can be rerun by others and not only compared. Late in the evening Sankalp took handoff signing and resolution for Probity, and says execution is now built against section 10 of the contract. APS keeps its own piece, the native approval, decision and intent records for one agreed fixture set.
Three projects opted in. Horizon Shield, AgentAvow and Probity AI each opted in to shared decisions with a public comment that names a representative and links a maintainer record. Their rows went in as #219, #220 and #221. Each opt-in takes effect seven days after its comment, on 3 November, so none of them counts toward the reviews already open on the principles and the boundary map. For the map I asked PIC, AEP and Nobulex to confirm their entries before the window closes on 28 October.
A review where I stepped back. On #218, Sankalp's definition of an independent run, I asked for four fixes, one of them a grammar mistake in wording I had suggested myself. Because some of the text is mine, APS will not count toward the two approvals that PR needs. On #217 I offered a second OpenTimestamps reader, written from the draft and its vectors without reading Tetsuro's code.
Shared budgets, CoSAI. The containment paper in CoSAI PR #262 sets accounting rules for a shared principal and each delegation subtree, but its demonstration section does not test them. My comment proposes four deterministic cases. With 10 units left, two concurrent requests for 4 are both admitted, and two racing requests for 6 cannot both be. A reservation with an unknown outcome stays charged even after it expires. A request past the accounting authority's latency bound is refused or held, never allowed.
What it does not show. The refund sandbox now has offers for its main roles, not a connected run. No outside implementation has run revision 3 independently yet, and Lee's records are author-produced. The CoSAI cases are a proposal to the workstream, not part of the paper.
Day 251: revision 3 separates a source we could not reach from a grant that was revoked
2026-10-25
Revision 3, merged. Yesterday Lars and Lee agreed on what a refusal should say, and today it shipped as #191. Every decision record in lease-revocation-axes now carries refusal_reason. It is null on a permit, and on a deny it is the first that applies of evidence_binding_mismatch, evidence_expired, authority_not_determined, authority_revoked and operation_not_permitted. authority_not_determined is the value for a source we could not reach, and for a cached observation the policy does not accept. It is never folded into authority_revoked. Each record also carries grant_ref, the evidence as issued and the instant of the observation the decision used. The pair of the vectors digest and the evidence id points to one entry in the frozen input, so anyone can check a record without the harness. No case narrows a grant after a decision, so that behavior is not tested here. Revision 2 is untouched and the existing runs stay comparable.
Where the reviews caught me. A Codex review of the first build found that an empty oracle table passed with zero checkpoints. The runner now requires exactly twelve checkpoints with fixed identities and fixed answer fields, and three new negative tests show an empty, a duplicated or an incomplete table failing. The same review narrowed my wording on grant_ref, which had claimed more than a fixed input shows. In the note to the list I had also written that this family uses absolute time. It does not, it uses a synthetic timeline in milliseconds, and the review took that sentence out before it was sent. The full suite exited 0 on that head and all six checks were green before the merge.
Lee's record, merged after a one line repair. #187 still failed after Lee's fixes, but only because main had moved and the lab's generated README inventory needed a row for his record. I added that one row on his branch, with no change to his files, and merged once the checks passed. Lee says his runner follows revision 3 next.
An AAE candidate in its own terms. Lars opened #190, 47 AAE vectors as byte copies pinned at 6bb462f, with three declared deviations from the lab's vector rules and no independent run yet. Under the candidate route his revocation_status_indeterminate stays AAE's term and adds no lab class. I asked for two copy fixes before publishing. Only the 64 files in his checksum list are copies, and his offline wrapper's audit hook covers the interpreter running a step, not child processes.
Who signed and who was allowed, IAB. Konstantin Mirin answered our AAMP cases by separating the signing profile from the authority record, and answered all three in the authority record discussion. My reply on OpenProposal #57 proposes a separate authority document linked to the signing profile. It keeps the record of an accepted commitment intact after a later revocation, without letting that record authorize delivery by itself. It says a seller's own running total cannot enforce a cap shared with other sellers. And it says a reservation can be released only once a terminal record rules out acceptance under that authority, because an expiry does not undo an acceptance made before it. These are proposals for the group, not settled rules.
Two more openings. Renato Iannella said yes to an ODRL 3.0 use case, so I opened w3c/odrl #171 with three case sketches, a downstream policy wider than the original, an upstream permission ending before downstream use, and a duty left out downstream. The expected decisions are left to the group. On the Personal Agent Protocol draft from Sierra and Meta I opened #38 about one standing permission used for several confirmations. It asks about aggregate limits, concurrent confirmations, and cancellation when the permission is withdrawn.
What it does not show. Revision 3 is lab-authored and author-produced, and no outside implementation has run it yet. Nobody has answered #171 or #38.
Day 250: a cold run that reproduced the revocation family, and a refusal that had no reason to give
2026-10-24
The revocation family merged as a candidate, and a cold run came in. PR #180 merged this morning with revision 2 of lease-revocation-axes, the timeline Lee and I agreed on the W3C list, with three axes scored separately and both declared modes. Tetsuroh Hori had already run revision 2 cold from his own catalog implementation at b745d25, and it matched 176 of 176 calls and 11 of 11 checkpoints. His record went in as #188, linking his RUN.md, the addendum and the correction file he wrote after one variant rule turned out to be ambiguous. He labels it author-produced and answer-exposed, because he had seen the expected outputs before he ran it, and the record keeps those labels.
Lee's runner, and a review I had to correct. Lee opened #187, a record of his RTTP/IQA runner over the same revision. I reran it locally. The vendored inputs matched #180, the runner was byte for byte his upstream 8cb2e24, it exited 0 with independent_run_matched, and two JSON runs were identical to each other and to his committed output. Changing one vector value made the digest check refuse. I asked for two documentation fixes. The vendored manifest listed three files that were not in the PR, and his claim that the runner never reads the oracle only covers reads through Python's open(), while a read through pathlib gets past it. My review said neither fix blocks the record. Then the PR's own CI failed on line endings and digests, so I withdrew that sentence. The record waits until the checks pass.
A refusal that had no reason to give. Lars Kroehl, author of the AAE Internet-Draft, joined the experiment on the list and offered his delegation and revocation vectors. His first point was right. The family already reports authority as valid, revoked or not established, but the action decision only says permit or deny. An implementation that never reached the revocation source has to pick a substantive reason it has no evidence for. I proposed a revision 3 with a refusal reason and a reference to the grant in every decision record, with revision 2 frozen so the existing runs stay comparable. Lars and Lee both said yes the same day.
ODRL charter. The W3C Digital Policy Working Group charter brings ODRL to version 3.0 with a conformance model for evaluators. I asked on the charter thread whether downstream policies will be covered. ODRL 2.2 has grantUse and nextPolicy, and inheritFrom is a different thing. An evaluator still has to decide what a downstream policy may permit, what happens when the original permission is revoked, and which duties carry over.
Two first contacts. On IAB Tech Lab AAMP #3, a proposal for one agent identity and signing profile in agentic media buying, I raised three delegation cases: an agency giving its buying agent a higher cap than the advertiser gave the agency, a mandate revoked after the agent signs a commitment but before the seller accepts, and one standing cap used by two commitments. I offered test vectors once the group agrees on the expected results. On the IDSA Rulebook I proposed a short subsection for the AI Agents chapter on how a Connector checks authority passed from one agent to another, with three tests. It does not mention APS, and it says the credential format and revocation mechanism stay with their working group.
What it does not show. Revision 2 and both runs are on vectors and an oracle I wrote. Neither first contact has an answer yet, and neither is adoption of anything.
Day 248: a revocation test family that a reviewer held back until four scoring flaws were fixed
2026-10-22
The revocation family, opened as a candidate. On the W3C list, Lee and I had agreed on a small experiment. Evidence is issued at T=0 and expires at T=120 seconds, grant g1 is revoked at T=10 seconds, and calls come at 30, 60 and 119 seconds, with a variant where the revocation source is unavailable. Today it went up as PR #180 in the conformance suite. It scores three axes separately, which are the freshness of the original evidence, the validity of the scoped authority, and whether the action was authorized under a declared policy. Verification acceptance is reported on its own as a composite. Each run declares its mode, cached lease or live standing read. The family has 12 timelines under both modes and both declared expiry endpoints, 12 defective variants with pinned failure sets, 6 mutation tests on the corpus and 12 hand reasoned oracle checkpoints. SDK primitive runs against agent-passport-system 7.1.0 and 8.0.1 make 132 checks each. The full suite exited 0 on the maintainer machine at 651503b, and all six checks on the PR are green at b745d25, which only changes the README.
What it does not show. The vectors, the oracle and every run are author produced. The SDK checks cover expiry and signed chain primitives under resolver answers the bridge supplies. They do not cover complete cached or live modes or real action enforcement. P1 is a policy the lab defined with no tolerance for observation age, so cached mode gets little action discrimination. No participant has reviewed the vectors, no result from ATP or RTTP/IQA is included, and the PR has no reviews or comments yet. It is a candidate for review. It is not a W3C specification, a ranking or a conformance verdict.
Where reviewers sent it back. A Codex design review of the first plan came back as build with changes. The plan mixed current authority validity, cached verification answers and action authorization, so the oracle could reward a wrong claim and reject a legitimate implementation. A GPT review of the finished eleven case build then held it before any PR, and its follow up review lists four blockers. The second worker case shared one verifier instance, so worker state was never independent. A live mode implementation could pass without proving it had read live. The oracle rejected a legitimate diagnostic read after the evidence expired. And a supported action axis could answer not_supported on a hard case and escape scoring. All four were repaired in revision 2, and the follow up review accepted the repairs for public review. That review checked the code and the vectors by reading them. It did not run them, so it is not independent reproduction.
A wrong sentence in my own PR draft. The first draft of the PR body said the evidence is revoked at T=10. The code revokes grant g1 at that moment, and the original evidence stays within its 120 second lease. The follow up review caught it, and the published body now says so. The reply on the list went out at 12:52 PDT with the same timeline. It links the PR, declares both modes, and asks each party which axes it claims, under which policy, and whether it will run the family and report in the open.
Refund sandbox, vocabulary #193. nutstrut sent six points, and I agreed with four. trust_snapshot_ref names the snapshot used when the handoff was issued, and the reader then judges key, role and revocation separately under trust configuration supplied outside the bundle. Otherwise a later trust change would fail as a digest mismatch and hide the revoked signer case. handoff_authenticated means the handoff was valid for the recorded dispatch attempt. It does not mean the handoff would authorize a new dispatch now. On item 6, a probe by Rul1an showed reconcile() promoting a stored effect with a different operation_id to observed_success, while inspect_claim() had already flagged the mismatch. The connected fixtures should include that case, with the check before promotion. The code belongs to altrudev, so I proposed no change to it.
Which context members an approval binds, AuthZEN #787. My comment says the bound set should follow what the approval covers. An approval limited to one MFA session depends on that session's identity. An approval reusable while the requester has a currently valid MFA session needs a current policy check and not the original session. Both stay subject to current policy. The lab's R6 and R7 candidate cases show the matching behavior, and they do not establish that R6 should allow access.
hs51, write up and acceptance records. Toshi published a write up of the October 22 run that cites public comments from the issue, and I confirmed that its APS summary matches what we reported, including the refusal of call C before anything was sent. Poke-nushi recomputed the set 2 settlement and matched the published SHA-256. Extending the header view to block 972564 left the verdicts unchanged and moved anchor_depth from 9 to 14. babyblueviper1 also issued four acceptance records, for MUSUBI set 2 A and B and for our A and B, each bound to its own execution receipt and request. A passes and B fails on three criteria, which is the expected outcome. He reports that all four verify under Trilatic's reference verifier and asked to post them, and I said yes. I have not run that verifier on them, so the verification is his run until someone repeats it.
A row corrected on agent-errata. On agent-errata #16 I told Pico that the aeoess K001 row must list NOTES.md, because our implementation used it. That probably keeps our result out of the independent count for that claim, and the call is his. I also proposed one stricter experiment. A new implementation is written from a pinned SETTLE_CORE_v0.md and the vector inputs, without NOTES.md or reference code and with the expected outputs withheld, and its code is frozen and hashed before anyone compares. kuangmi-bit's README says the vector decided where the text could be read two ways, so that run would be a different test and would not change its count. I am not offering to write it, because it needs someone who has not seen the expected outputs.
OWASP maturity model, Pillars 2 and 4. I continued my OWASP Agentic Identity Maturity Model review with proposed additions to Pillar 2 on delegated authority, revocation and authorization evidence. For Pillar 4, I proposed a recovery test. A payment agent fails attestation and is suspended. While it is suspended, its owner separately revokes payment permission. Passing attestation again must not restore that permission. The proposed test checks that payment requests remain denied while an unaffected authorized operation succeeds, so leaving the agent entirely blocked would not pass.
Smaller replies. On vocabulary #212 I asked altrudev to set refusal_authority: verifier, since lktron00 explained that refusing to accept an attestation happens at the verifier, and signing a record is a separate step. On google/adk-recipes #2769 I gave kgu-here a concrete case for the Agent Garden question, an agent that keeps running on its starting authority after an admin offboards the employee or a compliance team pulls a refund approval. There is no decision on the recipe yet.
Day 247: two paid calls, one budget refusal, and a run stopped between B and C
2026-10-21
The hs51 run. The window opened at 00:00 UTC on October 22, which was Wednesday evening in Los Angeles. Before preflight I read Bitcoin block 972546 by hand and checked its hash at two explorers. Preflight checked the three run keys against their posted public keys and fetched the responder's pilot.json. Init signed a spending limit of 7,000 JPY-testnet per call and 15,000 in total. Then the run made two paid calls to Toshi's worker. Call A audited the full fixture and call B audited its first two lines, and both execution receipts verified under the responder key. B's audit matched the 1,101 bytes Toshi had computed in advance. Call C would have brought the total to 21,000, so our gate refused it before any request left the machine. The result is posted on horizon-shield #51 as digests only. No real funds moved.
What went wrong. The process running the driver was killed right after B finished, because the remote shell that launched it lost its session. The driver left a lock and a marker saying the run had not finished. The runbook covers this case. I broke the lock with a note after checking the process was gone, and reconcile found no case in progress. The next run command, with a new approval token, skipped A and B and produced the refusal for C. Nothing was sent twice. Both records are in the archive, the archive passes our verifier with every key pinned, and a scan found none of the three seeds in its 50 files. Long steps now run detached from the shell that starts them.
Two designs, same responder. Toshi's MUSUBI run used the same window and fixtures. Its results show the same two audits byte for byte. It also made and paid its third call, logged it, and leaves the budget question to the settlement over the log. Our gate refuses the call before it is sent. The buyer side, babyblueviper1, asked for the audit bytes to issue acceptance records, and those stay private until everyone agrees to publish.
A digest that the doc described wrong. On agent-errata #16 checking K004 showed that our documentation did not match our code. The vectors digest delegation_chain wrapped in an object, and the doc said bare array. The code and vectors agree with each other, so the doc was wrong. #244 fixes the header and three rows of the mapping table and changes no vector.
Vocabulary. On #211 altrudev's crosswalk validator now fails when a registered file loses a required field. I deleted each required field from the real files and the validator exited 1 every time, then approved the fork's CI runs. On #179 more projects confirmed their sides of the boundary map. The head stays still until the window closes on October 28, and the status updates go into one follow-up PR.
Revocation timing, on the W3C list. On the public-agentprotocol list Lee and I agreed on a small experiment on revocation. Revoke a grant ten seconds after the evidence is issued, then call at 30, 60 and 119 seconds, once with a cached lease and once with a live status read, and once more with the status service unreachable. Each run is scored separately for how fresh the evidence is, whether the authority is still valid, and whether the action was authorized when it executed. The case family goes into the conformance lab.
Small fix. Paper page headings on this site used Pixelify Sans, whose fi ligature made "fixed" read as "Axed". They now use Jersey 15.
Day 246: a non-execution statement that could be followed by a receipt, fixed before the first payment
2026-10-20
The race in the pilot worker. On horizon-shield #51 I reported a race in the pilot worker at cfe64331. A call checked for a void or a receipt only before its audit started, and /void checked only for a receipt. So a void sent while a call was still auditing got a signed not_executed statement, and the call could then sign a receipt for the same request. Toshi confirmed it from the code within half an hour. Our rule for Thursday does not depend on his fix. A not_executed statement does not release our held 7,000 or allow a replacement call, and an executed result still needs a matching receipt that verifies.
Fixed before the run. Toshi deployed 283bee47 at about 22:40 UTC and froze the worker until the window closes. /call and /void for one request now run one at a time. A call marks itself in flight before the audit and checks again, right before signing, that no void, no receipt and no other call has appeared and that it is still within 25 seconds. While a call runs, /void answers in_progress, unsigned, and voids nothing. A not_executed statement is now final on his side. His tests include our interleaving.
Our driver against the new worker. The driver that runs our side is local and not published. I ran it against the worker's call and void handlers at 283bee47 on loopback, with the worker file pinned by sha256 29557fd2…. The interleaving runs end to end. A void sent during the held audit answers in_progress, the receipt is signed, and a second void answers executed, which the driver settles as placed. The driver accepts in_progress only when it names our request digest and carries nothing besides retry_after_s. It keeps the answer, moves no budget, and asks again later. These tests call the handlers directly, so the worker's serialization layer itself was not exercised on our side. The window stays Thursday, October 22, 00:00 to 03:00 UTC. Both witnesses, astrogilda and Poke-nushi, have posted their signed acknowledgements, and call A waits until they are anchored in Bitcoin. Nothing has been paid.
The payment rail crosswalk, re-verified. #202 merged. Every system in crosswalk/payment_rail.yaml is re-pinned to its current upstream, and every cell was re-read at that pin. 26 of 30 cells are unchanged. The x402 charge cell moves from bundled to mapped, because auth-capture v1.1 removed autoCapture and added a separate charge operation with no hold. Ectsang, who wrote the crosswalk, confirmed that change. MPP's authorize cell stays not_found, since mpp-specs#226 closed without merging. The file stays in incubation, and its next expiry date is January 13, 2027. Validation on main is green again, so #203 and #210 can go to review.
A run row with an exit code. On agent-errata #16 I reran our settlement checker from a fresh clone at 53cfafb against the corpus at dd64d0e2. 128 of 128 vectors agree on the eight core fields, exit 0. It is the same program on the same machine family as the first run, so it is a reproduction. I gave Pico the values for his new method and shared_with fields so the row says so.
Delegated compute on Gonka. On gonka #1304 I proposed a pilot for Track 2, Project 3 of Gonka's roadmap, delegated wallets and agent accounts. An owner gives an agent a compute budget, the agent hands part of it to a second agent, both draw from the owner's ceiling, two concurrent requests cannot spend the same remaining budget, and revoking the parent stops every later call under either delegation. Every admission and refusal gets a signed record linked to Gonka's own request and usage records. One call can involve several billable inference attempts, so a hard cap has to reserve the maximum charge before dispatch. No reply yet.
CAMARA and MCP. On CAMARA #347 I asked whether the v0.1 guidance for MCP could cover a bank fraud agent that calls SIM Swap through an MCP server. I asked four things. Which component holds the CAMARA token, whether the agent can get narrower authority than the application, what stops further calls when consent is withdrawn, and what links each tool call to the authorization and purpose it relied on.
Small fixes. The nightly job in the conformance suite that copies cross-implementation receipts failed when its upstream repository became unavailable. After #178 it records a warning and keeps the existing snapshots, and any other failure to reach upstream still fails the run. The research page now cites draft-kondoju-evc, an external verifier contract for agent authorization decisions.
Day 245: four receipts that failed their own verifier, and a lost-call rule agreed before the first payment
2026-10-19
Receipts that failed their own verifier. KVRM submitted a crosswalk to the vocabulary in #210. I ran its own command, kvrm-ai verify --domain, on the four published receipts and all four failed with FAIL_INTEGRITY, because they recorded a registry digest that no registry in the repository produced. My review asked for six more changes. The gate and the MCP gateway only decide and the caller runs the tool, so their refusals belong in a different field than the runtime's. Full replay needs the runtime, because with --registry alone a blocked receipt edited to allowed still passed. Nothing in a v0.1 receipt is signed, so the entry should not list signed fields. I also asked them to drop the line saying KVRM composes with APS delegations, since nobody has tested that. The author fixed everything the same day. At f3b7469 all four receipts verify, and regenerating them from the fixture gives the same receipt ids. One citation still names a method that does not exist at the pin, and the merge goes through our review legs after that.
The refund workflow on #193. #193 proposes binding an agent's refund to the exact merchant approval, with PIC on one side and APS on the other. I checked the APS parts against APS 8.0.1 and draft-04. Issuing the APS approval first works, because its receipt_id depends on nothing PIC signs. Five points need settling before the design freezes. They are where the decision evidence lives, who runs the PIC verifier, what happens when a reissued approval changes its receipt_id, how milliseconds and whole seconds compare without stretching validity, and the revocation recheck that draft-04 section 7.3.2 requires when the approval is consumed. My estimate for the APS part is 29 to 46 hours. That is more than the two days I offered on October 9, so the scope has to be agreed first.
A2A's signing example reproduces. On A2A #2369 I reran my check at 5b8383de. The 8.5 sample card reproduces the printed 2,645-byte payload and its digest, the test key derives from the stated phrase, the signature verifies, and RFC 6979 signing regenerates the published signature bytes. I posted the script. It uses no A2A SDK code, and an altered card, a wrong payload length or a wrong key curve each make it exit 1. It checks the two examples and not the open default-value question in #2122.
TRACE merged the expected policy hash. TRACE #553, from Day 242, merged. A verifier can now pass the policy hash it expects into TR-POL-003, so a matching digest for the wrong policy fails when the verifier supplies the expected one.
The lost-call rule for the pilot. On horizon-shield #51 I asked Toshi two questions before Thursday's window. The first was how case B should be built. The second was what happens to our held 7,000 if a live call fails without an answer, because under our 15,000 cap a held call blocks case B. Sankalp, one of the two witnesses, answered both. Case B pins the first two lines of the fixture, 207 bytes with sha256 1f75bcab…, which is exactly what our payer sends, and a lost call stays unresolved until the worker signs a statement that it did not run. We agreed, and Toshi built the second rule into the worker. A repeated call now returns the stored receipt marked as a replay, and a new /void endpoint returns either the receipt or a signed non-execution statement. Replacement calls are allowed only after such a statement. On Day 246 we found the worker could sign that statement while the original audit was still running, so we no longer release a hold on a statement alone. That is in the next post.
The run driver. The driver that runs our side of the pilot exists locally and is not published. Reviews on Day 245 found a path where recovery could release the budget while the original process could still send, a settle step that could release the hold before it checked a conflicting ledger row, and a verifier that accepted false budget totals. Each is fixed with a test that fails before the fix and passes after it. It is not cleared for the run yet, and nothing has been paid.
A claim anyone can re-run. Pico, the agent that runs Agent Errata, asked on #16 what to build next. I suggested starting with claims that ship their own test vectors, because then a re-run has a clear pass or fail, and named Toshi's settlement draft as a first row, since three programs already agree on its 128 vectors. Then I filed one for APS. An APS permit and denial verify from the committed files alone, pinned at bbd5a7d. My run passes, but the script is mine, so it counts as a reproduction and not as independent. A verifier written by someone else, in another language, would count.
The vocabulary. Five projects have agreed to the 72-hour window for adopting the decision process at e122e4f, which closes October 21 at 05:11 UTC. On #201 I asked the two who answered #185 and have not confirmed yet. On #202 I asked again about the x402 re-verification, because #203 and #210 wait on main going green. The research page now cites draft-thaha-scitt-agent-execution-evidence.
The federation, three weeks in: twenty projects on one map, a common port, and a pilot on ourselves
2026-10-18
The dev logs on this page say what changed in the code. This post is about the thing they keep mentioning in passing, because it is the most interesting work happening around APS right now and I have barely written about it.
The proposal. On September 29 I opened #177 in the vocabulary repository with a proposal for a federation of independent open source projects that build neutral infrastructure for AI agents. Many of us find the edge cases, write the specs and build the primitives, and then the ideas travel faster than the projects that made them. The work survives and the project may not. An agent's transaction needs authority, enforcement, payment, a receipt and verification, and no single small project should have to own all of it, so each of us keeps building one piece and makes it readable to the others. The rules came first: no required central protocol, no pay to play, no exclusivity, and nobody loses ownership of their project. I asked four questions: your project, one thing it produces or consumes that another project could verify, one boundary you would like to make interoperable, and what you would want from a federation like this. Ten projects answered on the first day. Several said they would talk about one technical boundary and nothing about structure or money, and one maintainer said federation was the wrong word for what we were describing. I kept the word and took the point. Answering does not enroll anyone.
We started by checking each other's work. The first thing we built was a map. Each project is a node and describes itself. A line between two projects exists only when one project ran another's artifact and the run is recorded, so the map shows what has been tested and nothing else. The draft has twenty projects and tools and seventeen of those lines. When a project asks me to add it, I ask it to write its own entry, because an entry I write about someone else is my claim and not theirs. We also keep the labels strict. When I run another project's fixture with my own code, it is a second implementation, because APS has a stake in the result. An independent run is one by someone who wrote none of the code. Pico, the agent that runs Agent Errata, ran Veritas Acta's E6 fixture that way, found two defects, and Acta shipped v1.1.0 with both fixed. Another project's independent run then passed on v1.1.0.
A common port. On October 7 I published federation-port, one open contract that a customer would integrate once and other projects plug into. Within two days REMORA, invinoveritas, Horizon Shield and AgentAvow each wrote an adapter for a check their own project already performs, without changing the core. On October 9 three of them ran together with an APS component in one workflow, all four required. REMORA also modelled the contract in TLA+ and sent four findings with a patch, which is the kind of help that makes a shared contract better than one I could write alone.
I stopped being the gate. Two decisions made it a group rather than my project with guests. First, nobody needs my permission to run anything. If the artifact and the procedure are public, anyone runs them, and review applies only when a result goes into a shared record. Otherwise I am a bottleneck in the middle of something meant to be neutral. Second, I administer the repository, but an admin merge should not count as the group agreeing, so I asked in #185 how changes to shared files should be decided. All seven responses supported the same process: a review window, one approval from a maintainer of another project, one voice per project, and no approving your own change. #201 writes it down.
The decisions from the last two days.
The agreed windows were seven days for ordinary changes and fourteen for principles. That is right for changing the principles and too slow for a group at this stage that ships every day. I proposed windows by type of change, 48 hours for verified progress and fourteen days kept for the principles and the process itself. A project's own entry, claims and keys still need that project's confirmation on every change, with no exceptions. For adopting the file I asked for a one-time 72-hour window, and only explicit agreement from everyone who answered #185 counts. Silence does not. Frequency agreed and invinoveritas approved the revision.
I also held back one of my own changes. Veritas Acta's E6 line is ready to be marked independently reproduced. On #177 I had promised that changes to E6 go through the shared process once it exists, so the update waits for #201, even though I prepared it and the evidence is there. If I skip the rule the first time it slows me down, it was never a rule.
Then the lead went quiet. On October 12 a store asked through our pilot form for agents that buy, order and refund through its existing checkout. It stopped replying. I told the group instead of letting it fade, and asked every project to put a pilot form on its own site and treat cold emails as possible leads, passing a lead to whichever project fits it better. One lead is not a pipeline, and twenty projects can find more than one.
A pilot on ourselves. In the same message I suggested we stop waiting: an agent with a delegated budget pays another project for a real service on a testnet. Four minutes later Toshi offered both halves. Horizon Shield's estimate audit, which homeowners in Japan already pay for, would be exposed as a metered call, and MUSUBI, his settlement layer, would handle authority and payment. Then babyblueviper1 named the step none of us had. Payer and payee can both be honest and the buyer can still get the wrong audit, so invinoveritas will commit the buyer's acceptance criteria before the call and sign a record afterwards of whether the audit met them, with a key that belongs to neither side. Toshi opened horizon-shield #51 with the frozen inputs: one audit call, three sets of receipts on the same request, and three cases. A call within budget, a call whose deliverable fails acceptance on purpose, and a second call over the budget.
The decision I care about most is what I did not argue. Several projects in #177 can produce this kind of receipt, and nobody knows yet which approach is best or how they really differ. So the same call runs through more than one of them, each project keeps its own verifier, and the comparison is written per run before anyone ranks anything. What we learn there decides what we standardize, instead of whoever argues best deciding it in advance. APS takes the authority side on the second run, and I run the payer agent. The spending limit is a signed delegation, a gate checks every call before it goes out, and every payment names exactly the request it pays for. The units are JPY-testnet play units, not money. Rehearsing it locally showed that my own payer could pay twice for one call, which Day 243 covers, and it is fixed and in review before the first testnet payment.
What is not true yet. There is no store and no customer. No real money has moved. The map and the shared principles are still open pull requests, and the process in #201 is not adopted. The economic idea behind all of this is mine and only an idea so far: value shared between projects in proportion to what was actually used, attributed per action from receipts. AgentAvow drafted the first attribution record for one action, and we agreed on the order, record first and money later.
What is true is that more than twenty people, and some of their agents, have spent three weeks running each other's code in public, posting the failures with the fixes, and writing their own rules for how they decide together. Next on #51 are testnet payments for the first two cases, then our keys and a date. If you build a piece of this stack, describe your project on #177 in your own words.
Day 243: a tool that signed verified without checking, and a payer that could pay twice for one call
2026-10-17
A receipt that said verified without checking. In the APS MCP server, create_access_receipt signed governance_verified: true without verifying the governance block it was given, so any caller could get a signed claim of verification for content nobody had checked. In 8.0.0 the tool requires the content and the publisher's public key, verifies the block, and refuses otherwise. Two other tools were broken. revoke_delegation failed on every call because SDK 8 moved the revocation functions, and it now issues an AuthorityRevocationV1 for a delegation the server signed. create_principal silently replaced an existing key, and now it refuses unless the caller passes replace=true. All 152 tools carry MCP annotations, the hints that tell a client whether a tool only reads or can destroy something. 8.0.0 is on npm, in the GitHub release and in the MCP registry, and a fresh install lists all 152 tools with annotations.
Annotations come from the server. OpenHands #18446 proposes setting each MCP tool's default approval from those same annotations. The MCP specification says annotations are untrusted unless the server is trusted, so a hostile server could mark delete_repo as read only and get it enabled automatically. I suggested using the hints only for servers the user trusts, starting everything else at Ask, and sending a tool back to Ask when its server adds it or changes its annotations.
MUSUBI revisions 0.2 and 0.3. Toshi revised MUSUBI settle core after my first comparison on horizon-shield #49. My checker matches 117 of 117 vectors on revision 0.2 and 123 of 123 on revision 0.3, on the eight core fields, with outputs frozen and hashed before any comparison. A conditional action written as [] or {} with approvers present crashed my runner before it gave a verdict. It now refuses that action as malformed, and all 123 outputs stayed byte identical after the fix. The checker is at aeoess/musubi-settle-core-check. APS is a party to the baseline contract, so this is a second implementation and not independent verification.
A pilot on Toshi's own project. horizon-shield #51 scopes one paid audit call, run through several receipt approaches with frozen inputs, so the records can be compared side by side. On the APS run I run the payer. Payments are JPY-testnet play units, not real money. The gate is aps-checkout-gate at d3602027, with a cap of 7,000 per call and 15,000 in total, and each payment's order reference is the digest of the full request, so a payment names exactly one call. In my local rehearsal the first two calls are permitted and the third is refused as over budget.
The payer could pay twice. My payer adapter had three defects. A retry of the same call was sent twice. A changed task or payee went out before the worker refused it. An altered audit record counted as executed even though its hash differed from the signed receipt. Now the payer records each request in a ledger on disk before sending and never sends one twice, checks the task, payee, amount and the other bindings before anything goes out, and counts a response as executed only when its digest matches the receipt. Retries in sequence, in parallel, after a restart and after a lost response each send once. Removing the retry guard, the audit check, the payee check or the task check makes the rehearsal fail. The fixed adapter goes through review before it is pinned with the run. Testnet payments for the two calls come next on #51, then our keys and the adapter pin, and only then a date.
stackql. #848 binds an approval retry to the action it approved, so an approval cannot be reused for a different action. I built it to the design jeffreyaven agreed on in #827, and he merged it.
The vocabulary. I approved the ATC crosswalk in #203 once its annotations moved out of the signed card, because one added provenance field inside the card changed its hash and broke the signature. It merges when main is green, and that needs #202 and its original author's agreement. #201, the decision process from #185, now sets review windows by type of change, from 48 hours for verified progress to 14 days for principles and governance. A project's own claims and keys need that project's confirmation on every change. The file is adopted 72 hours after its last commit if everyone who answered #185 agrees, otherwise after 14 days. I supported #208, which defines what independently reproduced means, with two edits.
Two outside fixes. AP2 #383 fixes #372, where an SD-JWT key carrying a use field failed verification. Its regression test fails on main and passes with the fix. academy #465 corrects a docstring that did not match its fields.
The site. Interior pages are now white with pixel headings and larger text, because long posts read more easily that way. The homepage has Star & Cat in the corner, a ginger cat holding a star. Clicking them opens a guided chat that explains what APS is, takes a pilot request or a question for me, and asks for a GitHub star. They start talking only after 30 seconds on the page on a computer, 40 on a phone, or once a reader scrolls to the third section, so someone who leaves in a few seconds is never interrupted. On phones the popup never opens by itself, because it would cover the pitch. Safari drew the video's shadow as a box around the cat, so the shadow is now a small oval under its feet.
Day 242: a read-only setting the client could switch off, and a second implementation that matched 76 of 76
2026-10-16
The label is policy, the database is the boundary. mcp-clickhouse v0.7.0 runs every SQL statement through one tool, run_query, so the APS ClickHouse example now reads the scope from the statement. SELECT, WITH, SHOW, DESCRIBE, EXPLAIN and EXISTS are reads, and everything else needs clickhouse:write. That check reads one keyword, so it cannot be the boundary. My first version leaned on readonly=1 in the client settings, and testing showed a per-query readonly=0 overrides it. mcp-clickhouse sends exactly that when writes are enabled. A read delegation now runs as a ClickHouse user created with readonly = 1 READONLY and only a SELECT grant, so the server refuses a write the label missed.
Review found two holes before the push. A destructive tier I had added let DROP/**/TABLE fall through to plain write, so I removed the tier. The next round found that files:read next to clickhouse:read picked the full user. The full user is now chosen only when the delegation authorizes clickhouse:write. The demo asserts every expected outcome and exits 1 on any mismatch, and I ran it broken three ways to see it fail. The comment on #155 says what each layer does.
76 of 76, with the limits stated. Toshi asked for a second implementation of MUSUBI settle core v0 on horizon-shield #49. I wrote one from the spec without reading his code. All 76 vectors match his expected outputs on the eight core fields, and the outputs were frozen and hashed before any comparison. My first comparator found disagreements and still exited 0, which a reviewer caught, so the published one exits 1. The code was written with AI, I ran it myself, and APS is a party to the baseline contract, so this is a second implementation and not independent verification. It is at aeoess/musubi-settle-core-check. One question went with it. My classifier counts an approval that carries only a signature field, while section 10.4 names only sig_b64, so I asked which one he intends.
Two checkout gate merges. imokokok's optional per-order cap check merged as #4. babyblueviper1's #6 adds an optional signed receipt that the permitted payment settled. I asked for changes because two failure paths could still produce a receipt. Both were fixed, all 75 tests passed, the failure probe matched main in both settlement modes, and it merged.
An expected policy hash for TRACE. Day 240 ended with a question on TRACE #280 about a matching digest for the wrong policy. #553 is the patch, so a verifier can pass the policy hash it expects into TR-POL-003. One related finding went to the maintainers privately.
The vocabulary. #198 merged, so Gate 3 now asks for a reviewed re-verification of every system in a crosswalk. Merging it meant turning admin enforcement off for one merge, because the nine failing checks come from the expired payment_rail.yaml, which #198 does not touch and #202 fixes. It was back on right after. #204 merged into the boundary map branch, which adds VATE and one edge to the proposal and adopts nothing. On the ATC crosswalk in #203 I requested changes twice. The second review covers who signs the test vector, that the digest covers the card and not the verdict, the fixture's provenance, and a risk band I could not derive.
I corrected my own comment. On CoSAI #270 I supported the new MUST for cases IND-01 to IND-10 and proposed scoping it to the premises each case states. My first draft overclaimed three points, and the review legs caught them before posting. One bullet in my earlier comment on #256 was wrong too, because the README row I questioned measures the opposite setting on purpose. I removed it.
Smaller things. The conformance suite merged three dependency bumps and the Verax accountability record run (#162). The site has a new page, data usage evidence, on signed records of what an agent did with a licensed work, and the demo now maps its records to AUCR. I also answered the RSL discussion on who enforces licensing terms, with the same records. The research page lists Liu, Yu and Zhang's arXiv:2610.15991, which separates delivering a revocation from stopping the tool call.
Day 241: seven answers agreed, so the decision process became a pull request
2026-10-15
Seven answers, one process. The consultation in #185 asked how changes to the vocabulary's shared files should be decided. It closed on October 14, and all seven responses supported the same process. Ordinary changes get seven days and one approval from a maintainer of another participating project. Changes to the principles or the process itself get fourteen days and approvals from two other projects. Each project has one voice, each person at most one, and authors cannot approve their own changes. The summary also records who chose to participate. Answering did not enroll anyone.
Why it had to be written down. I administer the repository, and an admin merge should not count as approval of shared work. #201 proposes SHARED-DECISIONS.md with the rules above, plus how a project opts in, how that is recorded, how it leaves, and what happens to commitments it already accepted. For this first PR I proposed the stricter rule, fourteen days and two outside approvals. Once it is adopted, #178 and #179 start fresh reviews under it.
The expired file got its re-verification. payment_rail.yaml had expired, and every validation run carried its nine errors. #202 re-checks every system in it against current upstream. The PR waits for its original author's agreement, because another reviewer's approval does not replace it.
8.0.1 reached the rest of the stack. The MCP README, the skills and the site's llms.txt moved to SDK 8.0.1. The ClawHub skill 6.0.4 is submitted and waits on ClawHub's security scan. The remote MCP bridge stays on 8.0.0 because it does not import the Cycles rail, so 8.0.1 changes nothing there.
An idea I stopped. I tested whether agent payment receipts should keep more evidence, using AP2's published sample as the case. The extra fields did not change what a reader could answer, so I stopped there.
Day 240: a matching hash and a valid signature, on the wrong envelope
2026-10-14
The wrong envelope passed. APS signs receipts for Cycles budget reservations, and each receipt can cite the Cycles server's signed envelope for that step. The offline verifier checked that the envelope's hash matched and its signature was valid. It did not check that the envelope belonged to this receipt. A permit could cite a release envelope, or an envelope from another reservation, and pass. #232 fixes that. A permit must cite a reserve envelope and a release a release envelope, each for the receipt's own reservation_id. Amount, unit, action and server identity are not compared, and denials get no correspondence check.
Ten rounds before it merged. Two reviewers attacked the patch, and every route they reported was reproduced before it was fixed. One early round found the check accepting a reservation_id that was not in the signed JSON at all, read through an inherited property or a non-enumerable field. The check now reads those fields only from the signed JSON. The final harness has 23 safety mutants, and each one makes a wrong envelope pass if it survives. All 23 were killed. Refusing receipts that passed before is a breaking change, so it shipped as SDK 8.0.0 with a migration note.
Not a new receipt type. I had started calling this a new kind of receipt. It is a fix to the receipts APS already signs for Cycles. Whether the rule underneath it is new, I have not checked against prior work, so the posts below ask instead of claiming.
8.0.0 shipped the fix where nobody could import it. The Cycles verifiers were never exported from the package root, and the package's export map blocks direct paths into dist. 8.0.1 exports them, and the packed tarball check now signs and verifies a permit through the root import. Both release runs went red after a successful publish. On 8.0.0 the registry check gave up after about three minutes while npm took about six to show the version. On 8.0.1 npm returned 404 for the attestations one second after publishing. Both packages were fine when I checked them by hand. Each check now waits and retries for that one missing response only.
The same 13 cases in two languages. The vectors are 3 accepted cases, 9 correspondence refusals and 1 hash mismatch. Python SDK 4.3.0 ports the offline path, passes all 13 and re-signs the vectors byte for byte. Both implementations are mine, so this is not independent. The rest of the stack moved to SDK 8 too. MCP 7.0.0 is a major release because create_delegation now refuses a fractional or negative max_depth, so the same input can get a different answer. The OpenClaw plugin 0.4.0 and the hosted MCP server followed.
Four places asked, nothing claimed. On the vocabulary I proposed the rule, that where a profile defines a relationship between a record and its evidence the verifier checks it before reporting a match, and a name for it. One of the questions is whether an existing term already covers it. The conformance lab got a candidate family and a question about how to classify it. On cycles-protocol I asked amavashev whether the stage and reservation rule matches how he intends the envelope to be used.
A matching digest for the wrong policy. The same shape exists in TRACE. I signed a record whose policy link resolves to an SBOM and whose bundle hash matches those bytes, and the conformance runner reports pass for TR-POL-001 through 003. The caller can enforce the policy it expects through its resolver, so this is not a bypass. On #280 I asked whether the paired tests should carry an expected policy hash, and which result the runner should report when it differs.
The real verdict permits now. babyblueviper1 signed a second verdict over the same order with a merchant id the gate accepts. PR #5 drives it through decide() to a signed permit that cites the verdict's event id. The old vector stays as a refusal test. 62 of 62 tests pass, and flipping one signature character fails the permit test. On imokokok's cap check in #4, one mutant survives, so nothing yet tests that the evidence id binds the cap. Lab #170 merged with kenneives' approval, and federation-port #9 merged.
Day 239: I merged an outside pull request, and the review after it found six defects
2026-10-13
The first outside pull request merged. babyblueviper1's PR #1 to aps-checkout-gate adds a signed verdict check in front of agent purchases. I reran it at the pinned head on Node 24. All 51 tests passed, and the mutation run killed all 39 mutants. The refusals it adds are signed by the gate key, and a refusal with no delegation cites a fixed sentinel instead of an empty reference. It also fixed a crash on main that I reproduced before the merge. A second PR put the same checks in CI.
Then the review found four defects. A code review of the merged head found four defects in how the gate handles unusual evidence and configuration. One let an evidence alias change the amount being reserved. The question was whether to unmerge. I fixed forward instead. Reverting would have brought back the crash on main, all four defects sit in checks a store has to turn on, and nothing was deployed. The recheck of that fix found two more, and the second fix cleared review. PR #3 landed with 57 of 57 tests and all 46 mutants killed.
An outside runner found my only real verdict could never permit. astrogilda shipped agent-checkout-runner, which drives the gate through 26 cases. One of them showed that the only real invinoveritas verdict in my tests names the merchant merchant:example-store. The gate's scope rule refuses that as more than one segment, so the request behind my one real outside verdict could never reach a permit. The README claim was accurate and the test still passed, because it only checked the binding. The scope reply on vocab #199 credits his runner and owes a new fixture with a fresh verdict.
A candidate family for AuthZEN's re-evaluation context. mcguinness opened AuthZEN #745, a draft that adds a re-evaluation context for the approvals discussed in #663, and credited our earlier fixture on #663. I built a candidate family pinned to the PR head. The first version carried a finding about a bound member missing on both sides. A review showed the bound set is defined as members of the evaluation context, so the finding came out. A deeper review found two ways the runner could pass a broken enforcement point. One rewrote the approval in place, and the other called the enforcement point twice. The runner now works on fresh copies and calls it once, and seven controls show each broken variant fails. Lab #169 merged, and the comment on #745 links the pinned tree.
Smaller things. I posted three runs of AgentAvow on lab #155. The native run is labeled a candidate for independent, because I operated it and did not implement it. federation-port #7 merged after babyblueviper1's outside rerun matched 49 of 49 and mine did too, and #9 got a review asking for two fixes before merge. On vocab #179, piiiico confirmed the Agent Errata node with no corrections.
Day 238: single use held only because nothing waited between the check and the consume
2026-10-12
Nothing waits between the check and the consume. On #193 the question was whether an exact refund approval can be spent twice. The sandbox's admitRefund checks the consumed set and adds the approval in one synchronous step, so on one Node event loop two calls cannot interleave there. To see what that property hangs on, I ran a separate race control that checked the shared consumed set, waited once, then added the approval. All 50 interleaved requests with the same approval were admitted. The current candidate keeps the check and the add together. The reply on #193 says the extension must finish every async step, key resolution, PIC verification or a store read, before the check, and keep the check and the add together. It also says single use is per approval id, so a reissued approval for the same refund counts as a second approval for now.
Consumed, then the refund fails. The candidate consumes the approval before execution, so sending that approval back through admission after a timeout is refused. Retrying execution under the existing operation is a separate path the candidate does not implement yet. I proposed returning the existing admission for a resubmission that still verifies and matches, without consuming again, and no dispatch attempt after valid_until, retries included. Reconciliation after it stays allowed and authorizes nothing. That is how federation-port's v0 contract already behaves.
Six fixtures rerun, label unchanged. Imran proposed E13 on the boundary map, the edge between TRACE and Acta's signed receipts, and posted a run. I reran the six pinned fixtures at trace-spec 4c7d6ecc with Acta's own verifier and got the same six results. Four controls behaved as expected, and TRACE's fixture test passed 30 of 30. The fixture history shows Tom added the set and Louie regenerated five of them. I recorded the rerun without claiming an independent record. Tom had asked that E13 stay as reported because Acta has not confirmed its side, so that label stays and both runs are recorded in prose. The run record with every hash is on #177.
The same verifier, 12 of 12 and then 0 of 14. My clean room VAID verifier from Day 234 matched all twelve vectors bundled in the vaid-a2a 0.1.1 wheel. On the repository's main branch it matched none of fourteen, because every vector there moved to signature version 4 today and my verifier refuses versions it does not implement. The lab record pins the wheel, says the run is author produced, and says six of the vectors came from my own proposals.
The rule said one thing and the validator another. The vocabulary repo's Gate 3 says a file past its 90-day incubation sunset gets promoted or deleted. The validator message says re-verify or promote, and my own issue #197 had credited Gate 3 with re-dating. PR #198 proposes a reviewed re-verification path. On a scratch copy, bumping the file date cleared only the sunset error, and the validator also accepted future dates and a map that dropped a system, so the PR says reviewers enforce those. Comments are open through October 14. The issue carries a dated correction under the original text.
Authority that expires between approval and consumption. Lab #167 adds approval-authority-expiry. An approval is issued while the agent's delegation is valid and consumed after it has expired or been revoked. The hard case is the edge of a two hop chain. The coordinator's delegation runs to 10:00 and the agent's runs to 09:00. At 08:59:59.999 the chain verifies, and at 09:00:00 it is expired while the root is still valid. My first version expired the root, which never tests whether every link is checked again, so the case moved to the child. Two cases describe a consumer that admits anyway, and the verifiers fail unless that exact defect reproduces. A mutation run of 47 cases ended with no unexpected result.
The coverage row stays partial. Draft-03 section 4.3 asks consumption to re-validate what can have changed since issuance, "including revocation status and temporal validity." The new fixture reaches both of those, and ID coverage revision 6 (#168) records it. REQ-4.3-1 stays PARTIAL anyway. "Including" does not close the list, reputation and live spend are never presented, and the revoked chain is a single delegation.
A checkout gate, and its first outside pull request. The scoping on #199 became a public repo, aps-checkout-gate, a gate in front of agent purchases and refunds. giskard09 took the fulfillment row. babyblueviper1 opened PR #1 to add a signed verdict check, and I asked for changes, mainly a generic evidence hook so the gate is not tied to one verdict format, and pinned BIP-340 vectors.
Smaller things. federation-port #7 now records only captured strings in a refusal, and a status string it does not know reads as unavailable with a protocol violation reason (49 of 49). The ID coverage revision merged and #150 closed. #215 merged after Hermes showed that five invented principal labels raised reputation confidence with the same mu and sigma. It corrects the claim and pins the behavior with a test. The counting itself is unchanged.
Day 237: the same signed bytes got two verdicts from my two SDKs, and the reviewers found bugs in my fixes
2026-10-11
A string where a number belongs. Codex reproduced issue #223 with a freshly signed legacy delegation carrying maxDepth: "1". TypeScript compared the string as a number and rejected the record at depth 2. Python skipped the comparison and accepted it at any depth. Same signed bytes, two verdicts. Now a present maxDepth or currentDepth that is not a non-negative integer up to 2**53 minus 1 makes the record invalid in both, and no depth result is claimed for it. GPT then found that both writers could still sign values like 1.5 or -1 that their own verifiers now reject, so the writers refuse them too. #225 and agent-passport-python #26.
The reviewers found bugs in my fixes. I moved profile recognition into a helper shared by both identity verifiers. Codex found that a proxy whose prototype trap throws now escaped the principal binding verifier as an exception, where before it returned malformed. The passport verifier had done the same since #221 this morning. One guard in the helper fixed both (#226). In the revocation fix (#227), the reviews showed that my changelog said every out of contract resolver answer used to throw. An answer like { state: "revoked" } had returned invalid. I corrected the sentence before merge. GPT also found that the shared resolver classifier said it refused non-plain answers and did not check. A Date carrying a key passed. The key it returned still had to verify the signature, but the code did not do what its comment said. Issue #228, fixed in #230.
An unknown key is now indeterminate. When the key resolver found no key, verifyPassportV2 and the principal binding verifier used to say invalid. Nothing was established about the key, so they now say indeterminate, with key authority unresolved. A record that names a type or version the SDK does not implement says unsupported, before the schema check and before any resolver call. Draft-04 does not prescribe these answers for these records, so they are reference choices, and the PRs say so. None of them admits anything, and the changelog marks the change breaking for callers that deny only on invalid. #221 and #226.
The frozen vectors stayed frozen. Three of yesterday's twenty passport cases now record answers the SDK no longer gives. Readers outside the project pin that file by sha256, so it stays byte for byte as published, and a second revision sits next to it with the same passports and the three new answers (#224). They are the cases where Probity's report already differed. Probity has not run the new file, and I wrote that on Toshi's thread with both hashes.
A missing depth limit means no limit. Both SDKs now apply no ceiling on verification when maxDepth is absent, as section 4.1 of the v1.1 spec says, and null gets the same answer because it signs to the same bytes. Minting a child under a parent with no ceiling is refused. Ryan's adapter applies a ceiling of 1, which is his local acceptance policy and fine as one, and I answered his thread with that. #222 and agent-passport-python #25.
Eleven Dependabot alerts, and a lockfile my laptop broke. Nine were fastify 4 in the example middleware folder, which the package does not ship but people copy. The example now asks for fastify 5.12.5, and the adapter type-checks against it unchanged. The other two were dev dependencies. My first lockfile came from npm 11 on macOS, which dropped optional entries, and CI's npm ci refused it. I regenerated it with npm 10.8.2, the version CI runs. Zero open alerts after #229.
Day 236: two people asked for frozen test vectors, and building them showed where my own document and my two SDKs disagreed
2026-10-10
Two asks in one night. Toshi wanted frozen vectors for the aps.agent-passport 2.0 record, because the SDK tests generate fresh keys on every run and his reader had nothing fixed to pin. Commonguy25 wanted the same for legacy Action Receipts v1.1, to check his own verifier against them. Both families are now in the SDK, merged in #217. The passport family has twenty cases: four valid, sixteen that must not be admitted. The receipt family has twelve delegation cases and ten receipt cases. Each case records the bytes, the signature and what one SDK function returned. None of them establishes that an action was authorized, and the files say so in every case.
The reviews found what my tests did not check. The packet went through three rounds of hostile review before merge. The first rounds caught claims stronger than the code. I wrote that a fixture matched what subDelegate mints field for field, and five fields differed. I said the vectors showed that createReceipt copies the chain, but the test threw the receipt away. Three changes to the SDK survived every test, among them removing the notBefore check. Now the returned receipt is compared, a notBefore boundary pair is in, and ten deliberate breaks of the SDK each make a test fail.
The document did not match the code. SPEC-v1.1.md shows snake_case member names, and the SDK signs camelCase, so a verifier that follows the document cannot read any receipt the SDK actually emits. The document says depth defaults to 1 in one section and unlimited in another. The TypeScript writer writes 1, the TypeScript verifier applies no ceiling, and the Python SDK reads a missing value as 0. It says verifiers must check revocation, and the verifier defaults to fail open. The document now carries a dated errata section that lists these differences. It describes the implementation and does not change what the document requires.
A bug on the other side, fixed the same day. Running Commonguy25's adapter against our records, I found that one validly signed revocation for a delegation outside the chain, or one signed by the wrong key, turned the chain's revocation status from unknown to active. He fixed it in receipt-interop a1fe88a with regression tests for both cases. I reran them, 116 of 116 pass.
A wrong sentence shipped, and lasted an hour. The passport vector file still described a did:web key validity window that an earlier review had removed. No outcome depended on it, but Toshi pins the file by sha256, so he would have pinned a wrong description. The correction is #220. Only that string changed, the file has a new hash, and I posted it on his thread.
The vectors pointed at two more bugs of mine. When the key resolver answers ambiguous, verifyPassportV2 reported the passport invalid, and draft-04 section 3.4 says that case is indeterminate. Issue #218 records it and #219 fixes it. The passport is still not admitted, only the classification was wrong. And the Python SDK wrote receipt version 1.0.0, which the TypeScript verifier refuses, so no receipt from Python could verify there. agent-passport-python #24 fixes the writer and adds a test against the vendored TypeScript vectors that runs in CI.
Readers outside the project ran the file. Toshi first got 16 of 20 with his admission reader. The four did:web cases needed a resolver input it did not have yet. After adding one, it matches the SDK on all twenty. He wrote what that does not show. Both sides read the same file, both Ed25519 libraries sit on OpenSSL, and retired keys and rotation are not covered. astrogilda then posted a separate Probity reader. It agrees on the diagnostic layers for all twenty and differs on three final outcomes, all of them cases where draft-04 does not require the SDK's answer. That is the useful result. The vectors separate what the draft decides from what my SDK decided.
Day 235: an outside report found a claim my SDK could not back, and the code behind a bug I reported had been deleted the day before
2026-10-09
A comment in my SDK promised more than the code did. An outside reporter showed on #210 that the reputation module counts any label the caller passes as a separate principal. Five invented labels raise the confidence score, while the score itself and its uncertainty stay exactly the same. The comments called this Sybil resistance. It is not, because nothing checks that two labels belong to two different parties, and signing would not fix it either, since one person can hold many keys. PR #215 corrects the comments and adds a regression test that pins the limit. It does not change the scoring. I also traced the gateway. The tier and admission paths I followed read the score and its uncertainty, not confidence, so I found no authorization impact there.
The bug was still there after the rewrite. On September 23 I reproduced a case in Goose where an approval given after a request was cancelled lands on the next request that reuses the same ID. When I nudged the maintainer about it, I had not noticed that the code I tested was deleted the day before, when the legacy agent loop was removed. So I built the new code from a clean clone and ran three tests through the real agent loop. A late Allow still runs the new tool call. A late Deny still decides it. When the second call gets a different ID, the late answer is refused. Same result on three runs. The comment on the issue now says all of that, including the part my nudge got wrong.
A signing example that now checks out from the spec alone. Last month I reported that the A2A spec's Agent Card signing example could not be verified, because it had no key and no stated payload. Erik drafted a fix with a test key derived from a printed phrase and the exact signed bytes. I checked it with three Python libraries none of us wrote. The 2,645-byte payload, the key, the header, the signature, and its deterministic regeneration all match. Two things went back with it. One sentence in the draft overstated what happens without RFC 6979. And the example cannot catch the rule conflict that makes SDKs disagree today, because the card in it has no empty or default values.
A rule went public before its review finished. In the federation-port draft I had to decide what counts as a present field when a component reports a result. I pushed a first rule on a go that covered the reviewed text, then the full review came back split two to two on one case, and the rule was replaced within the hour with a correction posted. The lesson I wrote down is simple. When a rule decides which inputs can admit an action, the review runs before the first push. Later, a review of my test helper found three gaps in it, and each fix now has a mutation that fails exactly its own test.
My own review was the wrong one. Toshi found a case inside his own component where a check of one endpoint and coverage of another could both pass. I proposed a rule for choosing what a claim binds and said the mixed case fails closed. Two other reviews showed it does not. A context claim about endpoint A plus target coverage of B still meets the stated requirements. The rule that went in says a claim binds the target by what it establishes, wherever the address came from, and a context claim cannot stand in for it even when the two strings match. Toshi checked it against his component and found nothing missing.
A run report became a record. Emek ran his own checker over our accountability record vectors, 12 of 12, and labeled it author-produced himself. I reran it on a second machine and the output matched his byte for byte. The record keeps every limit he named, and one more I found reading his code. It declares the expected error codes but never compares them. The pull request stays open until he confirms the record describes his work correctly.
Old promises, closed. A sweep of everything assigned to me or authored by me on GitHub found work I had missed, including the Goose issue above. It also found 25 old items of mine that had gone quiet, among them outreach issues from the spring that still quoted test and tool counts from back then. I closed them. Three sit in archived repositories and cannot be closed anymore.
Day 234: four checks passed together, and none of them checked where the refund went
2026-10-08
Four components in one decision. federation-port is a small runtime where independent projects plug in their own checks, and an action runs only if every required check passes. Until today each outside component had only been tested on its own. So I put four of them in one synthetic refund workflow on the unchanged runtime: our APS approval check, invinoveritas verdicts, NENRIN conduct walks and AgentAvow tool grades. With all four positive it admitted once. With a genuine negative from any one of them and the other three positive, it refused before anything was sent, and the refusal named only that component. Missing evidence did the same. The AgentAvow attestation in this run is their published payload re-signed with a local test key, because the published one expired before the other fixtures were signed, and the scan was a local stub. The executor is a simulator and everything runs in one process.
Every check passed, about different things. I set up the configuration deliberately mixed. NENRIN walked one endpoint, AgentAvow graded a tool on a different server, and the refund went to the simulator. Every check passed and the runtime admitted, because nothing in the contract ties a check to where the action is actually sent. That is not a bug in v0. v0 promises that every required claim is established, and that is what happened. What it cannot express is which claims have to be about the destination, so neither the runtime nor its record can tell a deliberate context check from a mistake.
The fix went to the people who have to build it. Before posting I ran it through four reviews, two of them told to attack it. One of them refuted a part of my own first proposal, that a component which says nothing about its subject could still pass. The proposal on #177 now says the target sits next to the action, never inside its signed bytes. Each required claim declares whether it binds the action, the target, or is only context, and a missing declaration fails closed. An action approval alone never counts as covering the destination. Components report the subject they verified and the runtime compares exact strings. Whether the executor really sends to that target is a separate question this does not answer.
Two authors reran it the same night. I put the test on a branch with a script that fetches each component at its pinned commit and checks every digest. Toshi and babyblueviper1 each reran it from a clean clone and got 8 of 8. Those are their runs of their own components, not independent checks, but they are the first reproductions from outside. babyblueviper1 then shipped a version of invinoveritas that reads the runtime target, and sent seven candidate vectors as a PR. I reviewed it and asked for one fix first. The two safety checks in its script could not fail. I changed one timestamp in the generator, the vectors changed in 42 lines, and the run still passed with exit 0.
Our own fetch followed redirects. Testing AgentAvow's component showed that the restricted fetch the runtime hands to components checked the first URL and then followed redirects anywhere. That is fixed and merged in federation-port #3. darklordVirtual's fix for retries after an uncertain provider answer merged too, and the open question, how long a retry with the same key is safe, is now his concrete proposal on #2.
A promise kept, with labels. Two days ago I told INAM I would recompute their x402 counterparty vectors and record the result in the conformance lab. Their vectors went from five to eight while we talked, partly from cases I suggested. All eight match, and the record is merged. It labels each layer separately. The canonical bytes and hashes come from two RFC 8785 libraries I did not write, so they count as independent. The normalization and decision checks run through my own code, so they do not, and the record says I suggested three of the vectors.
One question changed someone else's API. AgentAvow's component sent the full endpoint URL to their scan service, query string included. I asked what happens when that URL carries a token. kenneives answered by stripping credentials, query and fragment in the component and in the scan API itself, and refusing paths that look like keys. I reran it, 38 of 38.
🤖 For AI agents: If you are an agent reading this page, visit
agent-passport.org/llms.txt for machine-readable documentation or
llms-full.txt for the complete technical reference (1178 tests, 83 MCP tools, 42+32 modules). This page is designed for humans.