The Agent Graph Runtime

This repository holds a preprint that defines an execution model for agent systems, in which "static", "dynamic", and "hybrid" stop being runtime kinds and become separately specified coordinates: when a graph is constructed, where it becomes immutable and addressable, how it is authored, and what executes it. It also holds the full provenance record behind that preprint, including six attempted external certification epochs, none of which produced a certifiable review, all of which are preserved here rather than summarized away.

Versions

The public rendition is v1.0, dated 22 August 2026: paper-v1.0-preprint.pdf. Read that one.

It is the same manuscript as the archived paper-v0.9.10-internal.pdf, with three authored front-matter changes and nothing else. RENDITION-V1.0-2026-08-22.md lists them. In short: the cover notice's closing clause, which previously read "has not occurred; not for dissemination", now records that the release act occurred on 2026-08-22 under Decision A1; Appendix F gains a dated status paragraph; the version and date strings change. No epoch, no gate, no disclosure was touched.

Both PDFs carry the words "Internal preprint" at the top of the cover. That is accurate and deliberate. The manuscript was written under an internal certification protocol, six epochs of that protocol ran, and none of them produced a certifiable review. The cover notice discloses that instead of hiding it. What changed on 2026-08-22 is the release act, not the certification status.

v0.9.10-internal remains the internal version identity under the release gate, and it is the identity every hash and provenance record in this repository is written against. It is kept here rather than replaced.

What is in here

Path	What it is
paper-v1.0-preprint.pdf	The paper as released. 39 pages, 4 figures, 45 references. sha256 beabc9be049c62658b8cb91ce2223a0fe25a8dc3c7def8cac2b9b750ee4c6c81 (build-corrected 2026-08-26; see RENDITION-V1.0-2026-08-22.md addendum).
RENDITION-V1.0-2026-08-22.md	The three front-matter changes that make v1.0, and the decision behind them.
paper-v0.9.10-internal.pdf	The archived internal build. The bytes every frozen manifest below describes.
paper-v0.9.10-internal.txt	Deterministic text projection of that exact PDF, produced by pdftotext -layout -enc UTF-8. A navigation aid, not a second source.
REPRODUCIBILITY-SUPPLEMENT.md	Procedures, boundaries, and what each reported number does and does not license.
SOURCE-PROJECTION-V0910.json	Frozen build projection: the sha256 of every input that produced the PDF.
REVIEW-PAIR-MANIFEST-V0910.json	The exact byte identities handed to external review, plus the pinned text extractor.
references.bib	The BibTeX database behind the bibliography.
SHA256SUMS	Hashes of everything shipped here.
RELEASE-DECISION-2026-08-21.md	Why these bytes were released, and one open question the author left visible rather than hidden.
VERSION-IDENTITY-RESOLUTION-2026-08-21.md	How that question was answered, and what the release gate said when it was run through rather than read.
LICENSE	The single dual grant: MIT for code, CC BY 4.0 for documents, with the full text of both.
RELEASE-CLEARANCE.json	The author's publication decisions: licence selection bound to LICENSE by hash, and scoped consent for the six artifacts it names.
provenance/	The dated record: every certification epoch, every stop, every void, every finding disposition.
arxiv/	The LaTeX source that builds the PDF, plus submission notes.

Verify it yourself

Nothing here asks you to take a hash on trust. Every claim of identity below is checkable in a few minutes.

First, one thing to know about the manifests. SHA256SUMS and SOURCE-PROJECTION-V0910.json are frozen records of the v0.9.10-internal build. They were sealed then and are not rewritten. Four files in the current tree therefore no longer match SHA256SUMS, and are supposed not to:

• arxiv/paper.tex, arxiv/build-context.tex, arxiv/abstract.txt, rewritten by the v1.0 rendition.
• README.md, this file, which is living documentation rather than a sealed record. It is listed in the manifest because the manifest hashes everything that shipped on release day, and it has been revised since.

Checking the manifest against the current working tree reports exactly those four and nothing else. Every sealed artifact still verifies.

To check the manifest against the tree it actually describes, use the release commit:

mkdir -p /tmp/agr-v0910
git archive 07f6d98 | tar -x -C /tmp/agr-v0910
cd /tmp/agr-v0910 && shasum -a 256 -c SHA256SUMS

One file fails there too: RELEASE-DECISION-2026-08-21.md. The manifest was computed before a closing "Decision A1" section was appended to that document, and the appended text shipped in the release commit. The section was later lifted back out, so the file at the current tip does match the manifest again. The Decision A1 text itself is preserved in git history at commit 07f6d98, and the decision it records is restated in RENDITION-V1.0-2026-08-22.md and on the v1.0 cover.

1. Check the shipped bytes against the manifest.

shasum -a 256 -c SHA256SUMS

Expect OK on 43 of 47 entries, and FAILED on the four named above.

2. Check that the paper's own frozen projection agrees.

SOURCE-PROJECTION-V0910.json was written at build time and records what the build consumed. Its internal_pdf_sha256 field must equal the hash of the PDF you just checked, and its review_pair_manifest_sha256 must equal the hash of REVIEW-PAIR-MANIFEST-V0910.json:

python3 - <<'PY'
import hashlib, json, pathlib
def h(p): return hashlib.sha256(pathlib.Path(p).read_bytes()).hexdigest()
proj = json.load(open("SOURCE-PROJECTION-V0910.json"))
for field, path in [
    ("internal_pdf_sha256", "paper-v0.9.10-internal.pdf"),
    ("paper_text_projection_sha256", "paper-v0.9.10-internal.txt"),
    ("reproducibility_supplement_sha256", "REPRODUCIBILITY-SUPPLEMENT.md"),
    ("review_pair_manifest_sha256", "REVIEW-PAIR-MANIFEST-V0910.json"),
    ("paper_tex_sha256", "arxiv/paper.tex"),
    ("internal_build_context_sha256", "arxiv/build-context.tex"),
    ("references_bib_sha256", "references.bib"),
]:
    print("OK " if proj[field] == h(path) else "MISMATCH ", field, path)
PY

Two of the seven report MISMATCH in the current tree, for the same reason as above: paper_tex_sha256 and internal_build_context_sha256 describe the v0.9.10 sources, and arxiv/ now holds the v1.0 ones. Both match at the release commit:

git show 07f6d98:arxiv/paper.tex | shasum -a 256
git show 07f6d98:arxiv/build-context.tex | shasum -a 256

3. Rebuild the PDF.

arxiv/ holds the v1.0 sources. arxiv/upload-package-2026-08-22.tar.gz is the same set as submitted, and its paper.tex, build-context.tex and paper.bbl are byte-identical to the loose copies. With Tectonic 0.16.9:

mkdir -p rebuild
cp references.bib arxiv/
cd arxiv
SOURCE_DATE_EPOCH=1787356800 tectonic -X compile -C --reruns 2 \
  --keep-intermediates --keep-logs --outdir ../rebuild paper.tex

The v0.9.10 sources build the same way from git archive 07f6d98, with SOURCE_DATE_EPOCH=1785369600.

On byte identity, read the claim narrowly. The author confirmed on 2026-08-21 that the v0.9.10 sources rebuilt to 352,115 bytes, byte-identical to the released PDF, on that machine on that date. SOURCE_DATE_EPOCH removes the timestamp, but it does not pin the TeX package set: Tectonic resolves its support files from a versioned web bundle, and a later bundle typesets to a different byte stream. Rebuilding the v0.9.10 sources against a 2026 bundle here produced a well-formed 350,889-byte PDF, not the sealed one. Treat the SHA256SUMS and SOURCE-PROJECTION-V0910.json entries as the authority on what was released, and a rebuild as a check on the source, not as a second witness to the bytes.

One caveat about the bibliography. Tectonic re-runs BibTeX on every build and rewrites paper.bbl in place, so the rebuild needs references.bib beside paper.tex. It ships here, at the top level, which is why the command above copies it in. This is a Tectonic behaviour, not an arXiv one: arXiv does not run BibTeX and typesets the supplied paper.bbl directly.

4. Reproduce the text projection.

pdftotext -layout -enc UTF-8 paper-v0.9.10-internal.pdf - | shasum -a 256

Expect d27f06255fe436cf2f676e68db1d3cedf80c5d68a1734bb74758d0f868365dea. The extractor version pinned in REVIEW-PAIR-MANIFEST-V0910.json is pdftotext version 26.03.0. Other versions may differ in whitespace, which is exactly why the version is pinned rather than assumed.

What this record does and does not establish

Read provenance/RELEASE-BOUNDARY-DECISION-2026-08-14.md before drawing conclusions from the certification record. In short:

• Six external certification epochs (r1 through r6) were attempted. None produced a certifiable review. The two that reached adjudication (r3 and r6) returned noncompliant on producer conduct, principally because producer self-reports diverged from the platform-represented trace. In r6, 57 logged network requests appeared against zero represented network events.
• That is a result about the review instrument and the producing platform. It is not a result about this manuscript, which no epoch ever judged.
• Novelty, correctness, and comparative rank are uncertified. The paper is released with that stated plainly rather than waiting for a certificate the platform twice demonstrated it could not produce.
• The paper's own reported numbers are bounded by the supplement. The frozen C30 pilot is a mechanical invariant exercise, not a treatment comparison, and the paper excludes comparative superiority, amortization, and generality claims on that basis.

On the cover notice

Both covers carry an internal preprint notice. The v0.9.10 cover ends "That act has not occurred; not for dissemination". The v1.0 cover ends by recording that the act occurred on 2026-08-22 under Decision A1. That is the only substantive difference between the two notices, and it was made because the release act falsified the old sentence.

The rest of the notice stands in both, because it is still true. The six-epoch record is what it is, and the notice is where the paper says so on its own first page.

Clearing the notice in full would have meant building in release mode, and that was attempted on 2026-08-21 rather than assumed. VERSION-IDENTITY-RESOLUTION-2026-08-21.md records what happened. Release mode requires an external review bundle at the frozen protocol revision with a green validation result. No such bundle exists: the four raw trees that have the required layout declare later protocol revisions, and the validations that were produced returned noncompliant twice, indeterminate once, and nothing once. The gate was left closed and unamended rather than widened to let these bytes through. v1.0 is an authored front-matter rendition, not a gate pass, and RENDITION-V1.0-2026-08-22.md says so in those words.

What is withheld

This repository is a subtree of a larger private research repository. The following are deliberately not here:

• The raw certification trees (operon-v099r*/, operon-certification-v099r*/) and the two private Claude Science session exports they audit. The voided epochs are in private owner custody by design, and the exports were never platform-signed. provenance/ carries the dated record of each one, which is the part that carries evidential weight.
• Internal audit metadata: CLAIMS-LEDGER.md, spec-map.json, SOURCE-MAP.md, and the internal build and release runbooks.
• Client-identifying material from the surrounding private repository, which is permanently unpublishable. Everything shipped here passed a de-identification sweep with zero hits.

Citing

Rashid Azarang. The Agent Graph Runtime: A Unified Model for Static, Dynamic, and Hybrid Execution. Preprint v1.0.1, 2026. doi:10.5281/zenodo.22118742 (concept DOI 10.5281/zenodo.22118741).

@misc{azarang2026agentgraphruntime,
  author    = {Azarang, Rashid},
  title     = {The Agent Graph Runtime: A Unified Model for Static,
               Dynamic, and Hybrid Execution},
  year      = {2026},
  doi       = {10.5281/zenodo.22118742},
  publisher = {Zenodo},
  note      = {Preprint v1.0.1}
}

The release locator is bound by RELEASE-CLEARANCE-ADDENDUM-2026-08-26.md per the clearance record's own pending clause.

Licence

LICENSE is the authoritative file and carries the full official text of both grants. (c) 2026 Rashid Azarang.

• Source code is MIT.
• The paper, documentation, provenance records, and data files are CC BY 4.0.

LICENSE-CODE and LICENSE-DOCS are the same grant split in two, kept for readers who want one licence at a time. RELEASE-CLEARANCE.json records the selection as MIT AND CC-BY-4.0 and binds it to LICENSE by sha256.