Research edition: 2 October 2026. Source inspection is pinned to 4938e405. Proposals and standalone simulations are not production qualification. Original Markdown

Pantograph Research Sources

These books and papers inform development; they do not admit implementation work or change Pantograph's supported capabilities. Start with the active domain-architecture plan for scope and acceptance, the ADR index for accepted decisions, and Architecture for current ownership.

Editions And Evidence

All five manuscripts are the 2 October 2026 research editions. Their Pantograph source inspection is pinned to 4938e405c7f656365eefdca492774ccae110c90d. Statements such as "current" or "already implements" describe that snapshot, not later revisions. Source inspection is not build, deployment, crash-test, security-containment, or hardware-performance evidence.

| Reading | Status and use | | --- | --- | | Designing Compatible Inference Backends | Book: source-grounded comparison and proposed compatibility architecture. Chapter 9 audits Pantograph; chapters 10–16 propose contracts, qualification and a native roadmap. | | Pantograph Node Execution and Dynamic Extensions | Paper: section 5 audits the node boundary; sections 6–14 propose managed invocation, lifecycle, extension admission and qualification. | | Pantograph Execution Ledger and Provenance | Paper: section 3 audits existing evidence/artifact paths; later sections propose identity, commitment, retention and evaluation contracts. | | A Resource Safe Learning Scheduler for Pantograph | Revised paper: inspected production constraints, audited v1 simulation findings, and a proposed v2 research contract. It predates the v2 evaluation below. | | Pantograph scheduler v2 evaluation | Later C3-corrected standalone simulation report, with historical C1/O1 results and C3 regression evidence kept separate. No Pantograph implementation or GPU benchmark. |

The source manifest records exact SHA-256 identities, edition information and the separate C3 simulator fingerprint. Manuscripts, their linked figures and the inference book's primary-source index are byte-for-byte copies of the supplied research editions. Bibliographies, pinned source links, original diagram sources and limits remain with each work. No third-party papers, upstream source snapshots or PDF editions are bundled.

Read Against The Active Acceptance Criteria

The following links are design inputs to the plan's existing DA criteria, not new acceptance criteria or claims that an item is satisfied.

| Active obligation | Relevant research | | --- | --- | | DA-01 review and DA-02 domain ownership | Inference chapter 9 source audit and chapter 10 contract owners; node section 5 integration gaps. | | DA-03 real dependent text-to-image workflow | Inference chapter 12 data contracts and chapter 13 qualification. A simulation or independent text/image runs cannot satisfy the real mixed-workflow criterion. | | DA-04 failure, cancellation and resource ownership | Node section 7 lifecycle and section 13 falsifiers; scheduler sections 4–6. | | DA-05 retained-output lifetime and cold reopen | Ledger section 7 commitment, section 8 retention and section 14 validation. | | DA-06 measured orchestration and resource budgets | Scheduler section 9 measurement, section 10 planning cost and v2 real planning overhead. | | DA-07 final integrated evidence | Inference chapter 13, node section 13, and ledger section 14 supply possible falsifiers. Each required production claim still needs its own environment-qualified evidence. |

For the selected Pumas owner/client integration, preserve the plan's producer/consumer contract. The inference book distinguishes Pantograph's consumed Pumas revision f87c3da8276a914a54c6f4f36d617bef9d9f424e from its separately inspected newer upstream snapshot. Neither a newer producer nor a research proposal silently replaces the application's dependency pin, artifact identity, or load-target authority.

Scheduler Results Need Their Denominators

The C3 fresh holdout has 56 runs: 55 terminal and one budget-exceeded run. Search versus earliest-finish has nine wins and one loss across ten complete pairs; the iterative search run did not complete. This is one fresh seed, not broad generalization evidence. The earlier C1/O1 complete-pair comparison has 27 wins, one tie and two losses, with three of 33 search runs budget-exceeded.

C3 search averaged 116.72 seconds of planner CPU and 116.79 seconds of whole-run CPU; earliest-finish averaged 0.19 seconds of whole-run CPU. These are real CPU costs on the evaluation host. Modeled execution seconds are synthetic and are not hardware-calibrated. Planner columns omit interrupted calls and therefore understate planning cost for interrupted runs. The report preserves losing batching/split cases and the tiny oracle's missed batch-of-three optimum. Favorable simulated schedules do not establish useful production throughput.

C2/C3 corrected hidden-demand and forecast-metadata leakage, then cleanup-event ordering that could reveal latent state. The interrupted seed-9001 population, fresh seed-9029 population, and historical regression are separately disclosed. Do not use pre-correction results as evidence for the corrected information boundary or describe historical reruns as unseen holdouts.

Reproduction And Maintenance Boundary

This is a reading collection, not a self-contained experiment package. The scheduler manuscripts refer to their original standalone companions: simulator source, frozen contracts, raw runs, traces, tests and independent review files. Those companions are not bundled here, and no public download location is established by this import. Their reproduction commands and relative evidence filenames refer to those packages, not this repository checkout. This import checks manuscript/asset fidelity and links; it does not rerun their experiments or independently certify all reported results.

The inference book retains its 15 linked PNG diagrams and editable SVG originals; the v2 report retains both linked figures and their SVG originals. PDFs, covers, rendering tools and unused companion figures are omitted. Research recommendations remain proposals until adopted through the owning plan or ADR. When development invalidates a dated observation, update current guides and acceptance evidence rather than silently rewriting the research edition. Git history preserves replaced editions without a parallel archive.