- From: Abisheak P <abisheakp197@gmail.com>
- Date: Fri, 30 Oct 2026 19:37:16 +0530
- To: llewis <llewis@sovrlabs.com>
- Cc: ShaoBao Li <lee@aicent.com>, public-agentprotocol <public-agentprotocol@w3.org>, Alan Karp <alanhkarp@gmail.com>, "chopmob@gmail.com" <chopmob@gmail.com>, Lars Kersten Kroehl <lars@moltrust.ch>, ShaoBao Li <lee@iqa.org>, Xinghua Zhao <zhaoxinghua07@126.com>
- Message-ID: <CALwwdfobZ4KXh6f-Ey6YdxmNZxLP5AHkNxagh+k3kkdjpbT2tw@mail.gmail.com>
Hi Larry, Alan, Steven, AlgoVoi, Lee, Tymofii, all, I am writing to officially declare the NTI-1 specification as stable and production-ready for the AIW Book of Proceedings. The nti-spec repository is now fully green. We have integrated Steven's LGD transition contract, AlgoVoi's cryptographic canonicalization, and Alan's VOC model. The Python sidecar harness is live and running automatically in our CI/CD pipeline. From this moment, NTI is no longer a theoretical debate. It is an open-source standard. I am officially opening the repository for contributor onboarding. Here is how you can get involved immediately: 1. Become an Official Contributor: If you have test vectors, Rust matrices, or schema definitions (like Larry's Rust tests or AlgoVoi's preimage defenses), please submit them as a Pull Request to nti-spec or Neutral-Trust-Infrastructure. I review and merge daily. 2. Fork and Build: NTI is license-free and framework-agnostic. You are welcome to fork the repo, build your own custom features, and create your own NTI-compliant versions. 3. Integrate with Any AI Agent: NTI works with Codex, Claude, Gemini, or any other third-party agent. If you are building an agent framework, integrate the NTI sidecar and run the CI. We are submitting this consensus to Kaliya for the AIW Book of Proceedings. The window for getting your name and your work into the official standard is now. The repo is here: https://github.com/abisheakp197/nti-spec Let's build the Agentic Web. Best, Abisheak Architect & Maintainer, NTI-1 On Fri, 30 Oct, 2026, 11:04 am Xinghua Zhao, <zhaoxinghua07@126.com> wrote: > Larry, > > You asked for seven things: a runnable sidecar, a PAUSE-receipt entry > point, > one non-interruptible and one checkpointable operation, each operation's > finish-or-stop contract fixed before the run, admission scored separately > from > effects, a receipt checked against an independent observer, and an > independent > resource log. All seven are in the attached single file. Stdlib only, no > network, no keys: > > python transition_sidecar.py # human summary > python transition_sidecar.py --json out.json # machine report > python transition_sidecar.py --selftest # three registry modes > > Virtual clock throughout, so two runs on different machines produce > identical > bytes. Nothing here claims authenticity; no signatures are used at all. > What > it claims is that every receipt field is recomputable from the run inputs. > > THE TWO CONTRACTS, FIXED BEFORE THE RUN > > FINISH_THEN_STOP an admitted non-interruptible call runs to > completion. > The halt applies to the next call. The sidecar never > claims to have stopped it. > STOP_AT_CHECKPOINT an admitted checkpointable call stops at the first > checkpoint at or after the halt is known; work up to > that checkpoint is kept. > > Why two contracts rather than one policy: a receipt that says "stopped" > over > an operation that actually finished is not an approximation, it is a false > record -- and it is precisely the false record a test row is looking for. > For > the non-interruptible case there is no honest alternative to > FINISH_THEN_STOP, > and declaring that in the contract is cheaper than discovering it in the > results. > > FOUR QUESTIONS, SCORED SEPARATELY > > Q1 admission of the subsequent call DENIED after revocation > Q2 effects, non-interruptible completed = true > Q3 effects, checkpointable completed = false at cp5 > Q4 PAUSE receipt vs independent observer agree on both completed and cp > > Q2 and Q3 are expected to differ, and the difference is the finding rather > than a defect in either: on one timeline, with the halt landing while both > calls are already admitted, one of them must finish and the other must > stop at > a checkpoint. A suite that expected both to stop would be scoring an > implementation that lies, or one that cannot admit a non-interruptible > operation at all. > > On Q4: the observer log is written by the operation, and the sidecar does > not > read it during the run. The cross-check runs afterwards, on bytes. If the > guard could read the observer, "agrees with an independent observer" would > be > circular, so the arrangement makes that impossible rather than merely > inadvisable. > > REGISTRY FAILURE BEHAVIOUR > > reachable revocation at T=10 -> PAUSE, trigger=revocation > unreachable registry not asked -> PAUSE, trigger=registry_unreachable > stale_root answers an old root -> PAUSE, trigger=stale_root > > The latter two report authority_not_determined, kept distinct from > authority_revoked on purpose: a run that could not reach the registry must > not > score the same as a run that was told "not revoked." No mode stays ACTIVE. > > RECEIPT FIELDS (tier 1, all recomputable from the run inputs) > > receipt_id hash of the body with id and chain link removed > chain_id, prev_receipt, chain_link > transition PAUSE | ROLLBACK | RESUME > trigger revocation | expiry | anomaly | > registry_unreachable | stale_root > call_id, operation_kind, finish_or_stop, checkpoint_id > operation_completed boolean, must match the observer > t_effective when the authority-side fact took effect > t_observed when the runtime saw it > registry_root_version, registry_root_seen > resume_basis evidence | exemption (RESUME only) > exemption_covers, exemption_does_not_cover > issued_at, spec, scenario > > t_effective and t_observed as separate fields are Lars's point. > resume_basis > is the one I have proposed to Abisheak for the VOC path: both values may > end > in RESUMED, but the receipt says which door was used, so "the condition was > restored" and "someone accepted the risk" stop looking identical in the > record. > > Current result: 4/4 criteria and 4/4 receipts replay, including a negative > control that mutates one field and requires the receipt id to change. A > replay > check that cannot fail is not a replay check, so the control is part of the > report. > > LIMITS > > No action-time authorization, no signing, no Merkle tree beyond a hash > chain. > This is a row to run, break and score, not an implementation of record. > Take > the field list and the two contracts and discard the code if that is more > useful -- the contracts are the part I would defend. > > I can fit a session to your Central Time; my mornings are your evenings. > > Best regards, > > Xinghua Zhao > Independent researcher (SynomosAI initiative) > ORCID: 0009-0001-0512-1237 > > [Disclosure: drafting of this message was assisted by AI; all positions > reviewed and approved by me.] >
Received on Friday, 30 October 2026 14:07:34 UTC