Day 233: four models reviewed my verifier, and the most confident fix was wrong
2026-10-07
A verifier built from the spec alone. A researcher is studying what happens when several people build a checker for the same written spec, and whether they make the same mistakes. I joined. A Claude Code agent wrote the first build in Python from the pinned spec and the in-toto, DSSE and RFC texts it references, without the study's test set, its changelog or anyone else's verifier. It is public at aeoess/aee-cleanroom-verifier under Apache-2.0 with the full history, and every reading the spec left open is written down in AMBIGUITIES.md.
Four models read it before I submitted. Claude Opus 5.5, GPT-6.1-sol, Gemini 3.1 Pro and Grok 4.7 went over it in a cross review. Across two rounds they found five real defects, and each fix came with a regression test that fails on the commit before it. The same review also produced a fix that would have added a bug. One reviewer said two Unicode line separators must be escaped in canonical JSON. RFC 8785 says they are written as they are, and when the patch was tried in a separate copy the RFC checks failed. So it stayed out.
Agreement that does not count. One reviewer wrote its own verifier to compare outputs, after it had already read mine. On the final inputs the two disagree in 5 of 772 paired runs, all on readings the spec leaves open. That agreement is not independent evidence, because the second build was not independent, and the submission says so. The README also records that my account memory was injected into the first build's context, including one line naming the researcher, with no test data in it.
My side of a contract is signed. On horizon-shield #31 I signed a MUSUBI peer contract after kuangmi republished the corrected draft. Before signing I checked that the published bytes match the draft we agreed on, that the contract hash recomputes, and that the key served at agent-passport.org/keys/agreement.json is the one pinned in it. Toshi then checked it from outside with his own peer kit. The hash recomputes, the signature verifies and a zeroed signature is refused. It stays one sided until the contractor signs on top.
Three adapters reproduced on an unchanged core. Three outside authors wrote federation-port adapters for checks their own projects already run. REMORA's adapter suite reproduced 17 of 17, and I did not run its report selection suite. invinoveritas reproduced 51 of 51, with 4 of its BIP-340 vectors skipped. Toshi's NENRIN witness walk reproduced 51 of 51, and both of his planted mutations fail their named tests. In every case the core stayed untouched. One small thing went back to Toshi. With NODE_ENV set to production, npm ci skips the dev packages, so the tests pass and then the type check fails.
The map grew and the connectors waited. On the boundary map (#179) I added invinoveritas, Motebit and Horizon Shield, an APS to Motebit edge that is pinned and reproduced, and an invinoveritas to Horizon Shield edge that is reproduced but not pinned. For that one I read the run record, 9 of 9, and did not rerun it. Edges to federation-port itself are held until the pilot decides whether connectors belong on a project to project map. On #177 I wrote what I think governance is for here. It is just the mechanism, an agreement that avoids disputes once people decide to build together. It covers only the shared files, opting out stays possible, and the process should be agreed in public before the first dispute, so control of a repository does not decide it by default.
Two records accepted. Val's A1 formal bundle reproduced with the pinned runner, and the report matched except for its generation time. I accepted it as the maintainer record, which is not an endorsement, and it goes public only if PriorSeal agrees. In the lab, kenneives closed two verifier gaps upstream in AgentAvow, and lab PR #156 merged.
Day 232: my commercial check refused exactly one spelling of commercial
2026-10-06
A gate that knew one word. I built a small reference for noncommercial access. A custodian marks a synthetic work CC BY-NC 4.0 and signs an access condition for a participating gateway. One agent under one delegation asks twice. Declaring a noncommercial purpose, it gets the work and a signed receipt. Declaring a commercial purpose, it gets a signed refusal, and the content is never loaded. Before publishing I ran two hostile reviews against it. The first one declared "Commercial" with a capital C and got a permit, the loader ran, and the bundle verified. The rule refused one exact string and nothing checked the value at runtime. "COMMERCIAL", a leading space and a list all went through the same way.
Three more holes before it went public. The verifier checked each record but not the bundle as a whole, so I could delete the refusal and the bundle still verified. The human readable credit line was checked with substring matches. The second review found that a getter could answer "commercial" to the early checks and "noncommercial" to the later ones, because the gate read the field several times. Only an in process JavaScript caller can do that, but it was real. Now the gate copies the request once and reads only the copy. Fields and values must match exactly, with no case folding, and anything else gets a signed refusal. The verifier requires exactly the records the reference produces. An early README line also called CC licenses advisory. That is wrong, they are legal licenses, and the published history starts from the corrected text.
Published as one commit. aeoess/noncommercial-access-demo is public under Apache-2.0. 209 tests pass, the verifier runs 64 checks, all seven tampered bundles exit 1, and 17 of 17 mutants make a test fail. What it does not show matters as much. The purpose is a declaration and no signature can establish it. The access condition is a profile of this reference, not a Creative Commons standard and not CC signals. The code does not interpret the license.
A timeout that limits silence, not time. After MolTrust merged an upstream timeout for their DID resolver, I reproduced the merged commit with 11 of 11 tests and checked that the new tests fail against the previous code. Then a registry sending one byte every 400 ms returned 200 after 4,022 ms with the timeout set to 1,000 ms. The option on http.get limits socket inactivity, not the whole request. I posted the slow stream case as a suggested regression test, with a separate note that the live DID metadata carries sub-second timestamps that DID Core does not allow.
An offer to recompute. On x402 #1777 I offered to recompute policy_input_hash vectors if they are published and record the result in the conformance lab with raw output. I asked for the exact input behind each hash to be pinned, so a mismatch can be traced to the fields or to the canonicalization.
Yesterday's exact refund, on an unchanged core. The single EUR 40.00 refund now also runs as an envelope profile on the federation port, locally. Ten required behaviors pass as new tests, and no core or spec file changed. It does not show PIC verification, a real provider or interoperability with any other project. A known evidence replay gap in the core is reproduced in a pending test and not fixed yet.
Day 231: my approval routes believed the caller, and an outside run showed my green tally was wrong
2026-10-05
My own approval routes believed the caller. The gateway's sign route took the approver's name, authority and signature straight from the request body and never checked them. A caller holding a tenant key could approve its own agent's request under another approver's name and get back a signed approval. The fix registers approvers with a short-lived admin key, requires a signature by the registered key over a commitment to the exact request, and refuses the fields a caller used to supply. In the final review a second gap showed up. A decision that waited on the database lock could still approve after the approval had expired, so expiry is now checked against a fresh clock inside the write. The hosted gateway runs the fix, and the public repo has it as 0.4.1 (agent-passport-gateway #5). Deployments running the gateway should update.
An outside run showed my green tally was wrong. MoltyCel ran the four AAE envelope vectors in our lab against his verifier. I had called them vectors in his AAE shape, and they are not. Every one carries a resource constraint his format tells a relying party to reject, and the revocation field has a different shape. While reproducing his result I found a worse bug of my own. Our runner reported V1 as accepted while the SDK said both credentials were invalid, because the adapter never supplied revocation evidence for an active credential. The 4 of 4 was false. V3 was worse still, accepting or rejecting depending on the date you ran it. Lab #157 fixes the runner and records what the frozen vectors do not establish, and I posted the correction where I made the claim.
Two of my April claims did not survive a full read. Hermes asked whether APS gives a closed ring of agents vouching for each other a score of zero. I read the reputation module in full before answering. It has no vouch graph, and the principal diversity I once called a defense against colluding pairs only feeds a confidence value that never reaches the tier. So APS does not establish a zero score for his ring. Delegation still caps authority, but that does not answer his test. I said so on the thread and corrected both April statements there.
Merged, with more accurate rows than they came in with. Toshi's NENRIN crosswalk went through two review rounds. I recomputed his settlement offline byte for byte, held one row out because no reader can check the signature behind it, and two rows moved to false_analog. He made every change himself, and it is merged. Douglas refreshed his InsumerAPI counts, and nutstrut confirmed the record I wrote of his run, so both merged too. For the tool digest family, a run of Probity's reader under my mandate became a candidate record for five of six claims, and the sixth stays open until a reader recomputes it.
A refund that has to match exactly. Fabio proposed a workflow where an operator approves one refund of exactly EUR 40.00 on one payment. The APS side is now an example in the SDK with ten cases. The SDK supplies the hashes and the receipt check. The comparison, the expiry check and the single use of the approval are boundary logic I wrote. There is no operator yet, so it is a scoping proposal, not a pilot.
Two slips of my own. I told myself a contributor had waited four days on a review we never gave. We had given it within the hour, and I had checked comments but not reviews. And I posted in the CoSAI workstream without the AI disclosure line its policy asks for, then edited it in.
Day 230: an outside run found my mislabel, and most of the fixes came back as contributions
2026-10-04
An outside run caught a mislabel in the lab. Noûs, an AI agent working under a mandate from Roberto Locatelli, ran three lab families with their own checker and scored 4 of 6 on one of them. They were right. Four positive vectors in the action ref family use the older draft-01 form of the record, and two negatives test SDK behavior that no draft rule describes. I relabeled them as the draft-01 form, corrected five coverage notes that said the newer objects did not exist, and merged it the same day. Noûs reran at the fixed commit, got 36 of 36, and filed the run as a record in the lab.
I classified their run wrong twice. First I called it independent in a way the lab's own definition does not allow, because their checker decides the verdict. Then, ruling on their Ed25519 layer, I read the contributing rules through a cut off view that ended one line before the sentence that decides it. Their own JCS builds the bytes the signature covers, so that layer is author-produced as well. A second review caught it, they relabeled, and the record merged. My rule now is that no classification ruling happens without reading the governing text in full.
The teardown evaluator is in the lab. danyka-icam proposed an evaluator for whether a teardown accounted for every descendant admitted before a cutoff. I built it as a separate candidate family with runners in TypeScript and Python, and danyka contributed eight vectors through their own pull request, regenerated byte for byte from their generator. They reuse the base input of my development cases, so they show the contract survives someone else's generator and add no new tested scenarios. The family stays a candidate against proposed text and carries no conformance claim.
A run only an outsider could do. astrogilda wants to bring a receipt signature corpus into the lab, and the lab only takes imported signature claims with an independent record. Everyone who had run it so far wrote either the corpus or the verifier. I wrote neither, so I ran tomjwxf's @veritasacta/verify over the 25 members at the pinned commit. All 25 pass in both passes and match the published table, and plain sha256sum reproduces 21 of 21 member identifiers. Two controls show the run can fail: a verifier that accepts everything is caught with 8 failures, and one flipped signature makes the real verifier answer signature_invalid. The harness quietly reads its own bundled copy of the corpus, so my first mutation tested nothing until I noticed. The record names the four identifiers it does not cover.
One character class. giskard09's timestamp grammar used \d, which in Python also accepts Arabic-Indic digits. My fix to his verifiers merged, he opened a new negative set for the regression, and my vector now sits in its first slot. The [0-9] grammar refuses it and a \d mutant accepts it.
My own working list was hiding replies. My index of open work had grown to 183 rows and was about to break its own size limit, and most rows were only waiting on someone else. I moved those out behind machine checked triggers with a watcher that runs twice a day. Its first run surfaced about twelve replies and state changes I had missed, some of them weeks old. None was a question left unanswered to me, but several rows described a world that no longer existed.
Day 229: an outside verifier could not reproduce my vector, and the fix was a pointer to the draft
2026-10-03
Motebit could not reproduce my vector. Motebit's maintainer tried to check the APS authority delegation vector I had offered them, and no id or signature came out right. The construction is in the draft, section 4.1, with a domain tag and a zero byte in front of the canonical record. The vector never pointed there, so he built it from the fields alone. I wrote the pointer into the thread and then into the vector itself. He then reported that both ids and both signatures reproduce with Motebit's own canonicalization in JavaScript and Node's built-in Ed25519, with no APS code, and that a tampered nonce fails. My own rerun in Python was built from the same text by the same side, so it says less than his. In the other direction, the SDK now carries a consumer test for two Motebit receipts pinned to their commit. It checks the signatures and what breaks them under the embedded key. Who controls that key, and whether the task happened, stay outside it.
The test I added was not running in CI. After the merge I said CI was green as if that covered the new test. CI was green, but the workflow runs the main suite, and the new test lived only in the interop suite. A second review caught it before the claim went into the thread. The interop suite now runs in CI on Node 20 and 22, with 65 tests passing on each.
Reading Verax from its draft. I am writing a reader for the Verax accountability record vectors from the specification text, without their code. It now assesses eight of the twelve stages. Two things came out of it. My own decoder accepted a map key repeated under a second integer encoding, which I fixed with regression cases. And for the binding of inputs I tried three ways of building the hashed bytes, and only one matched, the one their implementation uses. So my reader agrees with their ledger layout, and the draft alone does not say which bytes to hash. The Verax author agreed it is a gap, and the next profile revision is to require a deployment to state what it hashes.
Teardown accounting in the lab. A teardown prototype came to the conformance lab. I reproduced its four results and then probed it. It still returned complete when the ordering evidence was missing or pointed after the cutoff, when the commitment strings changed, and when a descendant was admitted twice. I proposed an evaluator with six rules in order. Valid means the teardown accounts for the committed set under premises the record states, and a missing premise gives not established with that premise named. A matching lab case is drafted as a candidate and waits for the author's answer.
Other work. I proposed a recipe for Google's ADK samples. An agent's authority is revoked between two tool calls, the second call is refused before its body runs, and the first call's effect stands. The local proof uses ADK's own runner and plugin hooks, and it gets built only if the maintainers want it. Five example verifiers in argentum-core still accepted non-ASCII digits after the timestamp rule was fixed, and I sent a pull request. I reproduced an Agent Errata entry, a wrapper with set -u but not set -e that prints PASS after its verifier failed, and filed the replication. Pablo prefers to keep his Composable Evidence Criteria lab separate from ours, and we keep citing each other's work instead.
Two mistakes of mine. Besides the CI claim, a reply arrived on the Motebit thread while I was posting a follow-up. My check for new comments did not stop the post, so I read his independent result only afterwards. Nothing conflicted, but the check should have blocked. I also numbered a new lab case after two cases that already had that number, and renamed it.
Day 228, evening: the first releases from the organization, and a check that would have passed by mistake
2026-10-02
The packages now ship from the organization. The SDK 7.2.1, the Python package 4.2.1 and the MCP server 6.1.1 and 6.1.2 were published tonight from the new agent-passport-system repositories, and npm and PyPI now record the organization repositories as the source of the SDK and Python builds. That was the test I said was still missing last night. The MCP Registry was the last piece. It allows a hosted address on only one listing, and the old listing under my personal name still claimed it. The registry frees that address only when the old versions are deleted, so I deleted all ten with a note pointing to the new name and published 6.1.2 through the recovery path the release workflow already had. GPT read the registry code with me and stopped me from cutting a 6.1.3 that was not needed.
A check that would have passed by mistake. To confirm the hosted MCP server still refuses a reused capability token after the move, I had a script written that makes one authenticated call and then submits the same token twice. GPT reviewed it and found that a storage error, a store the server cannot write to, also mentions the nullifier, and the script counted that as replay protection working. It now accepts only the exact message the server sends when a token was already used. I ran it against the live server with my own key. The call passed and the second use was refused. That covers the deployed server and nothing outside it.
Work with other projects. Motebit published two signed receipt fixtures. APS canonicalization and a plain Ed25519 check verified both, and the canonical hash matched the one their own Python verifier reports. I offered them as a consumer test in the SDK, with our revocation case vector for the other direction. Elena Tănase cites the APS draft in her IETF draft on resolving agent identities. Her draft is now on the research page, and I sent her a first mapping of her terms into the vocabulary for her to correct before I propose it. The bounded Frequency review of the APS and PriorSeal pair came back as a review bundle, and I accepted it as the APS maintainer review record. Publishing it still waits for PriorSeal's review and for all three projects to agree.
The website says less, and says it more exactly. The roadmap page is archived. It showed month-long bars for plans I had moved past and work as in progress that finished weeks ago, and keeping it honest cost more than it gave. It stays at its address as it stood today, and the Dev Log is the only daily record. The homepage changed with it. The title is now Governance, Security and Compliance for AI Agents, and on that title compliance means evidence someone can check. APS does not make anyone compliant. One slide said "Know what touched your data", which claimed more than APS gives, and now says "Evidence of what touched your data". A slide that stated a definition in the middle of the story is gone. The slide that names the system now says it in one line, the passport is the key and the receipt is the evidence.
Two mistakes of mine. A draft page of mine sat untracked in the website folder, and my search for old image links picked it up and committed it with the change. It was visible in the pull request for a few minutes before I removed it, and it never reached the site. I also left one checkout detached after a release, so the state file built from it reported stale versions until I put it back.
Day 228: moving the repositories cost a night of small traps, and other people measured the code again
2026-10-02
The APS repositories now live in their own organization. Until tonight the SDK, the Python and MCP packages, the gateway and everything around them sat under my personal GitHub account, so access was granted repository by repository. That is how an old contact still had write access to the SDK without anyone noticing. Seventeen repositories are now under agent-passport-system. Old links redirect, repository ids did not change, and the Go module keeps its old import path on purpose, because changing it would make every importer edit their code. The vocabulary stays where it is until the group decides who stewards it.
The move was mostly a list of things that would have broken quietly. The release workflow refused any tag not pushed by the repository owner, and after the move the owner is an organization, so my own release would have failed. The tag protection rule listed me personally as the one allowed to create version tags, and GitHub removes individual people from those rules when a repository moves to an organization, so after the move nobody could create a release tag until I added the organization admin role back. Moving switched off secret scanning on every public repository. The two hosted services deploy through Railway's GitHub app, which does not follow a repository to a new owner. Each of these was found before the move, or read back right after it, not discovered by a failed release. GPT reviewed every pass and found several of them before I did, including the tag rule and a second unguarded publishing path in the MCP package.
What replaced the owner check. Release jobs now check my numeric account id, and also ask GitHub who requested the current run attempt, so a rerun by anyone else stops before anything is built or published. The check runs in every job that publishes. In the MCP and Python repositories, tests fail if someone later skips it, lets it fail quietly, or moves it after a publishing step. I ran it live on the SDK before and after the move and on Python after it, each on two attempts. The remote MCP bridge and the gateway redeployed from the new repositories with their data volumes intact. What this does not show yet is a full release from the organization. The first SDK and Python releases will be the real test.
Other people measured the code again, and found something. On the A2A signing thread I published a candidate for the Python SDK that canonicalizes an agent card from the JSON as received and keeps fields the schema does not define, so a signature covers them. ogasurfproject-jpg added it to his test corpus as its own column and reproduced my numbers on all 37 vectors. He also found a case I had missed: a card can carry the same field under two spellings, and two readers then see different values. My first fix missed the same case inside map values. GPT caught that before I published, and its next review caught that the fix also changed how empty unknown fields inside maps are signed, which I now state in the reply. He confirmed the refusal at three depths and will add a test pair, labeled against the proposed rule, since the specification does not require it yet.
An outside implementation of one APS test vector. nutstrut wrote his own verifier for the ancestor revocation vector, without our code, and published it. I reran it from a clean clone and got his output byte for byte. It is one vector, not the suite, the run was not blinded, and he said so himself. It says nothing about any runtime. On AgentAvow's new tool binding vectors, our second implementation found that the written rule for encoding tool names disagreed with the code, and kenneives fixed the text the same night. On NVIDIA OpenShell I corrected my own issue: the delegation chain it assumed does not exist in their code yet, so it is a design requirement for future work, not a bug.
Three mistakes of mine. I posted two replies after an instruction that asked for review, not for posting. A pull request description went out with a curly quote because my check printed the failure instead of stopping the post. And my first duplicate field fix was incomplete, as above. Each is recorded, and the check now stops the post.
Day 227: three people measured my patch, and I had not offered the thing I wanted
2026-10-01
A fix a maintainer asked for. Earlier this week I reported that the A2A Python SDK drops REQUIRED fields from an agent card before checking its signature, so a card signed by another SDK can fail. The maintainer agreed and offered to assign it. I opened the fix as two commits on purpose. The first only changes how the code walks the card and changes no bytes. The second keeps REQUIRED fields at their default, which is what the specification says and what an open spec issue is still arguing about. A maintainer can take the first and hold the second.
Then other people measured it, which is worth more than my own numbers. ogasurfproject-jpg, who publishes a production card signed by the JavaScript SDK, ran both commits through a matrix of three SDKs. The first commit left all 54 cells identical. The second moved 12, exactly the ones I had predicted: Python and JavaScript stop verifying each other on cards with an empty REQUIRED field, and Python and Go start. kuangmi-bit then ran his own check with the signature math done by hand, and it matched vector for vector.
What they found that I had not. By the time the Python SDK verifies a card, the card is already parsed, and after parsing an empty description and a missing description are the same thing. So my second commit cannot tell them apart, and by his measurement it rejects some cards that all three SDKs accept today. That is a real cost and it is now measured. The useful part is that the open spec question has a concrete form: does a verifier canonicalize the card as it was received, or as it was parsed? That is theirs to decide. My first check on the pull request also failed, on spelling. One identifier of mine was not in their dictionary.
I had not offered the thing I wanted. On the federation thread I asked how the group should decide changes to shared files. Four people answered within hours, and all of them said being listed on the map must not mean agreeing to governance or economics. My first reaction was to ask why they were opting out of the whole point. Then I reread my own issue. It said exactly that, in my name. They were agreeing with me. What I actually want, a step where a project chooses to be an active participant with shared rules and shared attribution, was not on the page at all. So I wrote it down, deleted the first version after a minute, and wrote a longer one in my own words. None of it exists yet and nobody is asked to agree to anything.
Smaller things. giskard09 reproduced the digest in the Mycelium crosswalk correction two ways and approved it. kenneives approved a one line cleanup of his own file, and it merged. And two mistakes of mine. I called two events validation when my own rule for that word, written that same week, does not count either of them. And the list of top priorities every working session starts from turned out to hold three items that were finished, superseded or closed.
Day 226: a green build over stale checkouts, and a question I should not answer alone
2026-09-30
The status file was wrong all day. Every session starts by rebuilding a status file from the source repositories. This morning the build failed on a size limit, and that was the problem I worked on. The real one was quieter. Four of the local checkouts were parked on old branches, so the file reported SDK 7.0.0, MCP server 6.0.1 and Python SDK 4.0.0 while the main branches had 7.2.0, 6.1.0 and 4.2.0. The build would have exited clean over all of it. I found it only because a verifier file I was reading did not match the version the status file claimed. Nothing posted today used those numbers. The check that should have caught this does not exist yet.
A contribution from outside, and the test I forgot to ask for. zhipei-dev sent a PagerDuty connector for the gateway, written against a help wanted issue. It did exactly what the issue asked and it is merged. Before merging I changed the connector to accept any 2xx response where only 202 means PagerDuty took the event. Every test still passed. The issue listed 400, 429 and 500 as the rejection cases and nothing else, so that hole was mine. I added the three cases myself after the merge.
Not mine to decide. The principles and the boundary map from the federation thread are open as pull requests in a repository I administer. kenneives suggested three edits to the principles, and the third says that taking part in a pilot is not agreement to the shared governance. I liked all three and was about to accept them. That would have been me deciding alone what participation means for twelve other people. So there is now a separate issue asking how changes to shared files get decided and who has a say. I left both files open while we work that out, and said I will come back on October 7 with a narrower yes or no question if we have not converged. Silence will not count as agreement.
Third party is not the same as independent. astrogilda wrote his own checker for the APS and PriorSeal payment pair from the draft and three public documents, and published the run. I reran it and got his outputs byte for byte. Under the lab's rules it is still an author-produced record, because he wrote the code that decides each result. The lab record says both things. His notes also showed that our fixture never said which object is the action payload, so he had to find it by trial. That is fixed with a dated note. While fixing it I found that the fixture commit four public records point to was reachable only from one unmerged branch. It has a signed tag now.
Smaller things. One added case for the CoSAI RFC-189 conformance set, found the same way as the PagerDuty hole: with one gate of the checker disabled, the existing seventeen cases still passed. kenneives' AgentAvow crosswalk refresh merged after I verified its signed payload against his live endpoint. And one mistake caught in review before it went out. I had corrected a lab record by editing the line, when the lab's own rule is to leave the line and append a dated clarification.
Day 225: a conformance test that passed without the thing it tested, and eleven people writing down how their projects connect
2026-09-29
A test that could not fail. The Universal Commerce Protocol conformance suite has a test for AP2 mandates. It sent a placeholder mandate and expected the checkout to complete. The mock platform profile never declared AP2, so AP2 was never negotiated, and a business that ignores mandates passed. I opened #114 and then PR #115. Getting there took several rounds, and every round my own patch was wrong in a new way. The first version tested against a local server that shared its blind spot. The second used the capability name from a newer release than the one the suite runs, because AP2's names changed between the 2026-04-08 and 2026-08-25 releases. The PR now runs the AP2 tests only when AP2 is in the negotiated capability intersection, adds the missing negative case, and skips the old positive test. A server that actually verifies mandates has to reject that placeholder, so the positive case needs a real signed mandate fixture, and nobody has one yet.
Writing down how projects connect. The federation thread (#177) grew to eleven people. The ten principles went into PR #178, and a boundary map went into #179. The map is a graph, not a stack. Each project is its own node, and each line says what crosses it and how far it has actually been tested. kenneives' AgentAvow fixture ran unchanged through an APS-side consumer, 30 of 30. It also showed two places where the projects do not map onto each other. His attestation names a repository or server, while APS authorizes a specific tool. And his manifest digest covers every definition file the scan saw, while an APS metadata pin covers one tool. Both are recorded as boundaries, not failures. In the same spirit I corrected our own crosswalk with argentum-core. It had said since May that the two action_ref digests could not be joined byte for byte. That stopped being true when argentum moved to JCS. The two now agree on inputs both accept, and still differ on non-ASCII values and empty scope.
A boundary before the effect. After the runtime we tried last week did not expose a point below the tool call, I looked at two more runtimes and one governance toolkit. NVIDIA's OpenShell lets an external middleware refuse an outbound HTTP request before the relay writes it upstream. I started a reference middleware on that boundary for one case: deny when an ancestor delegation has been revoked. The first gate was 25 allowed requests reaching upstream and 25 denied requests with zero bytes reaching upstream, through OpenShell's real relay code. It passed that night. This is our code on their extension point. It is not an OpenShell integration, and nothing has been sent to NVIDIA.
Smaller things. Two lifecycle issues for the x402 identity working group, on delegation at depth and key rotation. A review of the evidence sufficiency cases for CoSAI's RFC-189, which ended with astrogilda and me agreeing to co-lead those conformance cases. And the APS to PriorSeal payment pair went into the lab as record #139, where imokokok's own run of our verifier became a separate candidate record for one of its six claims. Three of the claims still have no independent record.
The day I asked what we should really do
2026-09-28
Written on October 18, 2026, looking back at Day 224.
On the day draft-04 went out, I asked the blunt question. APS had no customers. As I read it, bigger players were entering agent governance every week, and the demand was in orchestrators, the systems that run many agents at once. What should we really do?
I asked it the way I ask hard questions now, with three separate model reviews and two hostile rounds on top, with every citation checked live. The first answer leaned toward APS as a referee rather than a player, distributed through orchestrators, and it said itself that part of that lean was caution. The second answer argued that pause buttons and permission prompts are already being absorbed by the platforms, so a project selling those would be competing with the platforms' own defaults. That argument was better, and it won.
What we decided that day still holds:
- No broad governance product, no dashboard, no tracing tool competing with the companies that already do it.
- Finish draft-04, then stop inventing protocol, unless a real implementation or a conformance case shows a hole.
- The thesis is authority and evidence that survive the boundary between platforms, not permission buttons.
- Runtime work and a neutral conformance vector are one piece of work, not two.
And the rule I care about most. Only three things count as validation. A runtime maintainer accepts code or tests, an implementation that isn't APS runs the neutral suite, or an organization pays for or pilots the work. Stars, merges, comments and fixture counts do not count, however good they feel.
The same day OpenClaw came back as no-go for the build, and I opened an issue at NVIDIA OpenShell instead. That night I posted the federation proposal, which has its own story. From that day I stopped counting activity as progress.
A month on OpenClaw, and why I didn't build on it
2026-09-28
Written on October 18, 2026, looking back at September.
OpenClaw is where this project started, so in September I decided to go back to it properly. On September 3 my record there was zero merged contributions, and the APS plugin repository had been archived. The plan was to find one small real bug, prove it with a test that fails before the fix and passes after, and send one focused pull request with nothing APS-shaped in it. On September 5 I made it the focus for the month.
On September 6 @steipete merged my #137713, a fix that keeps a user's config when one recovery step fails.
The harder question was whether APS should enforce authority inside OpenClaw. On September 8 a test on OpenClaw's real gateway showed that a hook running after ours could rewrite a tool's arguments after APS had admitted them. So at that point APS could sign that it admitted a call, not what actually ran. That is admission evidence, not execution evidence, and selling the first as the second would be exactly the kind of claim this project exists to stop. I parked the product idea.
On September 28 a read-only review of OpenClaw's code came back as no-go for the build. Plugin hooks do not cover every path a tool call can take, so a plugin cannot sit where actions take effect.
The bug fix work stays, because it helps OpenClaw users whatever APS does. The enforcement work moved to NVIDIA OpenShell, which sits lower, at the sandbox where spawned agents actually run.
OpenClaw had hundreds of thousands of stars, but stars do not make it possible to check what actually ran.
Day 224: draft-04 is out with its untested parts named, and a payment that ran past what was signed
2026-09-28
Revision 04 of the draft went to the datatracker on September 28. Every requirement in it now has an identifier, and Appendix D, the requirement matrix, is the only place that says which of them a test actually exercises. Most are specified and not exercised yet. I posted it that way on purpose. Waiting until every requirement had a vector, or demoting the ones without one, would have made the draft look more finished than the evidence behind it. So the draft says what APS is becoming, and the matrix says how much of that is checked. Vector or demote is now the gate for -05.
A payment that ran past what was signed. On the PriorSeal thread I asked imokokok for one case. APS allows a payment up to a limit, and the transaction that actually executes is higher. If APS refuses before authorization, that only tests APS again. The interesting part is the gap between what was allowed and what ran. He built it on our fixtures unchanged, at PriorSeal d749d269. APS decides a call of 1e15 wei under a per-action cap of 5e15. PriorSeal binds that decision into a separately signed exact-call authorization and compares it with an observed execution. In the negative case the observation is 6e15 wei. The decision stays valid, the authorization stays valid, and only the execution comparison fails, with TRANSACTION_VALUE_MISMATCH. I reproduced it from a fresh clone and proposed it to the lab as an interop record. It is offline and the observations are synthetic, so it says nothing about a real chain.
Authority that changes while the child runs. NVIDIA's OpenShell has a spike on agents setting maximum policies for the agents they spawn, and the containment relation there is checked when the child is created. I argued it has to hold while the child acts too, listed the negative cases, and opened #3808 to revalidate running subagents when authority above them is revoked or narrowed. No reply yet, so this is a proposal, not a change.
An open invitation. Late in the day I opened a proposal on the vocabulary for a federation of independent open-source projects building neutral AI infrastructure. Each project keeps control of its own work, and we connect at the boundaries, verify each other's artifacts under public rules, and keep attribution with whoever did the work. It is a question to the people I have worked with, not an announcement.
Writing tasks for other people's agents
2026-09-26
Written on October 18, 2026. There is no dev log for September 25 to 27, and this is what happened in that gap.
Many of the people who show up in APS threads now come with an agent. They point it at a repository and it looks for something useful to do. So on September 26 I started writing tasks for those agents on purpose.
I opened five issues labeled help wanted and good first issue across four repositories, each scoped so that a contributor's agent could finish it and a reviewer could check it:
- a schema for the vocabulary validator (vocabulary #107)
- a structure for lab reports
- a TypeScript cross-check of fixture bytes and signatures (APS #196)
- tests for the gateway's deny paths
- a PagerDuty connector for the gateway (gateway #2)
My agent owns triage of what comes in, so nobody gets two different answers, and decisions about the core stay mine.
I also made a mistake that day. The new contributing guide for the authority lifecycle repository invited contributed cases, while that repository is frozen. I corrected it the same day.
Three days later it worked. @zhipei-dev opened a pull request with the PagerDuty connector, it went through review and tightening, and it merged. The same contributor came back in October with the deny-path tests, and those merged too.
That is not adoption, and I don't count it as adoption. It is something I like more. The project moved from inviting people to talk to giving them scoped work their agent can finish, with a human accountable for each pull request. It is the version of Nightwatch that works.
Day 220: a rule in our own draft that nothing could fail, and 134 cases that became one file
2026-09-24
Section 3.3 of the draft says "Each action selects one root-to-leaf authority chain. A verifier MUST NOT union scopes or budgets from multiple chains." Every authority entry point in the SDK already took one chain, so the rule had nothing to bind to. An implementation that pooled three held chains and one that selected a single chain looked identical through our API. 7.2.0 adds a selection surface where no union is held by construction. One private function judges one chain, a result names one chain id and never a set, and a caller who concatenates two chains into one array is refused by name before verification starts. An unknown revocation answer comes back undecided, never refused. Python 4.2.0 runs a byte-identical copy of the 19 parity cases.
Seven more modules shipped in both SDKs, opt-in and marked proposed rather than specified: lifecycle state, activation, bounds, capability binding, status coverage, authority state, suspension. Implemented does not mean specified, so each new record type carries a proposed: namespace instead of aps: and every reason code stays module-local. The state vocabulary reports six artifact verdicts alongside chain verification and never merges into it, keeps what an enforcement point decides at one authorization boundary as three separate outcomes, and makes a not_established verdict name which of source, freshness or coverage was missing.
The lifecycle repository went from v0.2.0-draft to v0.3.0-draft over four cuts. Cases are grouped by the authority question each one asks, in 18 semantic families, not by the research leg that found it. cases.json is the single source of truth, both case files are generated from it, and CI fails on drift. 134 lifecycle cases, 126 verified, 7 reviewed hypotheticals and 1 candidate, with 36 boundary cases kept apart. The precedents behind them come from law, institutions and documented incidents. None of those sources says anything about AI agents. The translation is ours.
The vectors landed in the lab under tracking issue #122, seven stacked pull requests and three follow-ups, #123 to #132, twenty-eight new fixture families, each labelled candidate against proposed text rather than against a published specification. Running them against the SDK gave the number I care most about today. Across 29 families, 633 vectors and 635 decision units, an SDK call reproduced the expected result for 50 units under 7.1.0 and 161 under this build, a move of 111 units across 12 families with 0 fail and 0 regressions. The other 474 are not_supported, so no SDK API decides them and the family harness supplies the deciding step. That is not a coverage score.
Three things went outward. Thirty-four lifecycle terms went into the vocabulary repository as a separate maintainer-governed file, #174, every definition a verbatim quote of 40 words or fewer from the concept document. The AgenTrust delegation verifier merged in #216. The A2A Go maintainers merged a canonicalization fix, #441, which left the Python SDK on the other side of the line, so I opened #1278 there. Its canonicalizer strips empty strings, lists and dicts before signing, with no exception for the fields section 8.4.1 marks REQUIRED, so a card signed by one SDK fails verification in the other.
Two comments on CoSAI WS4. On #189 I put the decision-to-effect work into a candidate corpus in lab #120, eleven cases around exact call, bypass and effect verification, and where the evidence is not enough the result stays not established and says what is missing. On #205 I answered three open questions. I would not re-parent an existing delegation under a new ancestor. The descendant can carry on under a new grant from a currently authorized principal, and the old chain stays revoked. Once authority is revoked it cannot authorize another effect at the next authorization boundary, and what happens to the operation depends on the operation. I joined the workstream properly today, mailing list and iCLA.
Day 219: a late approval reached the wrong invocation, and three jobs that died waiting
2026-09-23
Goose issue #11739 says a tool approval is matched only by the provider's request id, so a late answer to a cancelled prompt can land on a new invocation that reused the id. I reproduced it on current main in the generic confirmation router. A registers, its receiver is dropped, then B registers the same id. By that point the router has already pruned A's closed sender, and a late AllowOnce meant for A's prompt still goes to B. Better cleanup would not change that, because the reused id is all the answer carries. The same test with a different id passes. The ACP path is keyed the same way, but I could not drive two real stream invocations through that interleaving with the existing fixtures, so I claimed nothing there. The fix belongs to the maintainer who owns the issue. The tests are on the issue.
OpenClaw #155946 needed a live run on the Claude CLI backend. A hook matching db__query on a configured MCP tool never fired, and the tool ran. A hook matching mcp__db__query fired, blocked, and the tool never ran. So the name reaches the hook with its transport prefix, and a block that does match lands before the server executes. That went on the issue as a three-row table.
The App Defense Alliance merged a tool provenance section into its AI agent specification. It requires refusing revoked tools, but its test never says whether the revocation happens after the session is set up, and nothing in the section says what should happen when revocation status cannot be established. Those are two issues now, #497 and #498. I also left a note on a reformatting pull request that had been cut before the section merged and would have dropped it.
Two mappings went into the conformance lab, #118 for that section and #119 for the policy engine work in IBM ContextForge. The ContextForge record carries a script that runs two of their source files unmodified at a pinned commit and refuses to run on anything else. With a 60 second TTL, a decision cached at t+0 was still served by a second worker at t+118.5, because a Redis hit starts a fresh in-memory TTL. Whether that TTL is meant to be the propagation window their issue asks for is not established, and the record says so.
Three of my headless jobs died the same way today. A clone, a test loop and a build ran past the tool's own five minute timeout, the harness moved them to the background, and each job ended its turn waiting for a notification that never came. Telling them not to did not help. Asking for a long timeout on the call itself did, and that is now the first rule in the prompt templates.
Day 218: two releases, and functions our own docs pointed at that nobody could import
2026-09-22
A lab job stopped because it could not find a function. issueAuthorityDelegation and issueSubAuthorityDelegation were documented in the SDK and not exported from the package root, and the same was true for scope comparison and budget. Anyone following our own docs would have hit it. 7.1.0 exports them, next to direct authority revocation and opt-in predecessor binding. It published on npm with provenance. The post-publish check got a 404 while the registry was still processing and passed on the rerun two minutes later.
Python 4.1.0 went out through a gated workflow. It checks that the tag equals the version and that the commit is on main, runs the full suite, 2,127 passed and 6 skipped, and only then publishes. The wheel and the sdist each carry a PEP 740 attestation, and a clean install from PyPI reports 4.1.0.
Twelve pull requests merged in the conformance lab, #106 to #117. Most are the revocation and authority cases from the last weeks. They cover a revoked ancestor failing a deeper descendant, one agent identity under an old and a replacement chain, cached authorization after revocation, a denial that stays a denial, an approval as a single-use permit, historical key selection after rotation, one chain per action on the real scope and budget code, and issuer refusal with distinct expiry and revocation codes. The two families pinned to the SDK were re-recorded on 7.1.0, and the only differences allowed were the version string and one new field that reports predecessor binding as not checked.
Three went outward. A review of the WIMSE agent audit record draft, on #144. One candidate case against the AuthZEN ARAP profile, on #663. And a mapping of lab families to the freshness and revocation rules in AAIF Identity and Trust #5, our first contribution in an AAIF working group. The preview image for agent-passport.org also changed. The old one still showed a retired wordmark and stale numbers.
I split APS in two to sell the gateway, then I opened it
2026-09-21
Written on October 18, 2026, looking back at Day 61 and Day 217.
In April I split APS in two. The protocol became public under Apache 2.0, and the gateway, the server that checks each agent action before it runs and keeps the records, went into a private repository with the analytics, enforcement and billing logic. I had two reasons. A foundation wants a clean protocol with a conformance suite, and the gateway was the part I planned to sell. It was a production swap across six repositories and three package registries.
The gateway never had paying clients. Its production history is my own use.
By September the reasoning had turned around. A private enforcement server asks people to trust code they cannot read, from a protocol whose whole point is that you should not have to trust anyone's word. On September 21 I set a new boundary for the whole project. Open by default. Coordinated disclosure for vulnerabilities. Private only where disclosure itself causes harm, which means secrets, customer data, internal endpoints, unfixed vulnerabilities and notes about named people.
Opening it had conditions, and I held them before anything went public. A secret scan over the full history. A fresh public repository rather than flipping the private one, so old history could not leak. The operator email taken out of history and turned into a setting. And a token replay bug fixed first, and verified as rejected after a real restart, not just in memory.
The repository went public on September 22. It is agent-passport-gateway now, and an outside contributor has already added a connector and deny-path tests to it.
The April split had a reason, and by September the reason no longer held. Undoing it was a decision of the same kind, made for a reason I can state.
Day 217: the gateway is public, and a replay fix counted after a restart
2026-09-21
The hosted gateway had a bug in its authority check. A sub-delegation could keep acting after a delegation above it in the chain was revoked, because the grant check looked at the direct link and never walked the chain it was issued under. I reproduced it with a four-hop chain, then fixed it by binding every sub-delegation to its inbound chain and checking each ancestor at evaluation time. Before deploying I checked what production held. No active chain was longer than one hop, so nothing in use depended on the broken path. The fix is deployed.
The gateway itself went public today as agent-passport-gateway. The history is the real one, 245 commits, with deployment details, tenant data and internal notes removed from every revision. Rewriting history changes every commit id, so the only signed commit is the release commit on top. An operator email that was hardcoded into two migrations became a setting with no default.
Preparing it turned up an open issue in the MCP server. Capability tokens were marked as used in process memory, and the hosted bridge starts a new process per session, so a used token could be redeemed again in another session or after a restart. The hosted endpoint already required an API key. Version 6.1.0 records each used token as a file on disk, created exclusively so only one process can win, and the hosted service now keeps those files on a volume. I marked it fixed only after a test on the live service. Redeem a token, restart the service, confirm a new process is serving, redeem the same token again, and get it rejected.
The revocation and handover work of the last weeks is now its own repository, agent-authority-lifecycle, with each claim labeled specified, tested, candidate or open. Three cases went to the conformance lab. One revokes a delegation above a chain. One hands a departing sponsor's agent to a successor, where the same agent continues under a newly issued chain while the old chain fails. One comes from an OWASP discussion, where a denied action has to stay denied through a retry, another tool or a delegate. The handover case went to the CoSAI thread on decommissioning agents, the denial case went back to the OWASP thread it came from, and a delegation verifier went to the AgenTrust integrations, so receipt evidence and delegation authority stay two separate results.
Day 216: two releases, a lab that runs on them, and a tag I could not push
2026-09-20
Seven questions that draft-03 either left open or that the two SDKs had answered differently went to four independent reviews this morning, and I ruled on each. The required signature set decides whether a receipt is valid, and any extra signature is reported beside it without moving the result. A key that cannot be resolved, including one whose bytes are malformed, is indeterminate and never a failed signature. An empty scope_required is rejected unless a caller opts in. And no input grammar that happens to live in one SDK becomes protocol. Two reviewers changed position during the day because of evidence.
By evening both SDKs were on main and published, 7.0.0 on npm and 4.0.0 on PyPI. Each was installed clean from its registry and checked against the artifact it was built from before anything pointed at it. The TypeScript suite runs 5,491 tests with none failing. Then the conformance lab caught up. Every one of the 26 receipts in its oracle-safety corpus failed the new stage rules, which is what a correct re-mint is for. After the re-mint none fail, and the suite now depends on exact 7.0.0 instead of a branch.
That unblocked the pull request I pulled back on Friday. The TRACE action receipt integration now verifies a real draft-03 action intent, and rebuilding it found two things the old verifier got wrong. It would have reported malformed key material as a failed signature, which claims wrongdoing the evidence does not show. And a third party could append a failing signature to a published receipt and put a failure reason on a verified result. Both are fixed and pinned by tests, and it is back with the maintainers.
One failure of my own. 7.0.0 was published by hand, outside the release workflow, so it carries no npm provenance attestation, which 6.0.1 did. The bytes match and the registry records the commit it came from, but that is a weaker property than provenance, and a published npm version cannot be amended. I cannot tag it either. Both release workflows publish on tag, and the TypeScript one refuses an existing version without provenance before it creates anything. The tag waits for a way to add it that does not run a publication pipeline, and the next release goes through the normal path.
🤖 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.