# Coordination Bus Governed Agent Infrastructure - Research Notebook ## Question (pre-registered, dated) [2026-07-06] What does a coordination bus add to agentic engineering governance when it is treated as evidence infrastructure instead of execution authority? ## Method (pre-declared, before any data/derivation) Case study plus systems analysis. Start from public HUMMBL artifacts, then reconstruct the HUMMBL hardening-sprint lessons at a public-safe level. Keep operational details out of the paper unless they are already intentionally public. ## Predictions (timestamped BEFORE observations) [2026-07-06] The strongest contribution will be the evidence/action separation: a bus can coordinate agent work without becoming a tool-invocation authority. [2026-07-06] The most likely reviewer objection is generality. The first draft should not overclaim beyond HUMMBL's own case study. [2026-07-06] The cleanest measurable outcome will be queue reconstruction time: how quickly a reviewer can explain why work is blocked, ready, merged, or rejected. ## Observations (append-only, dated) [2026-07-06] Issue #12 asks for the first hummbl-papers artifact from the HUMMBL hardening sprint. The existing roadmap lists COORDINATION_BUS among the current hardening-sprint papers. [2026-07-06] The repository already contains public source artifacts relevant to the paper: `reports/reports/HUMMBL_BUS_PROTOCOL_SPEC.md`, `reports/reports/multi_agent_coordination_2026.md`, `reports/reports/capability_security_agent_systems_2026.md`, and `frameworks/unified-tier/governance/CAES_SPEC.md`. [2026-07-06] The release-artifact guard currently checks for at least one non-README artifact under both `papers/` and `notebooks/`. A scoped paper plus this notebook should satisfy the guard without pretending the paper is published. ## Interpretation (separated from Observations) [2026-07-06] COORDINATION_BUS is a suitable first seed because it sits between research, runtime governance, and the operational queue-control issues surfaced in recent hardening work. [2026-07-06] The public version should be written as a scoped case study. It can discuss architectural invariants and measurable outcomes while excluding private bus contents, local machine paths, credentials, and live operational state. ## Dead ends (failed attempts, also first-class) - [2026-07-06] Considered seeding a formal KILL_SWITCH paper first. Rejected for this issue because a coordination-bus case study is easier to publish safely without exposing private runtime conditions. - [2026-07-06] Considered adding a finished DOI-ready paper. Rejected because the source material has not yet gone through arXiv/Zenodo release steps, and the repository's method docs distinguish SCOPED/DRAFTING from PUBLISHED. ## Open questions - What replay fixture can be released to make the queue-control claim testable? - Should the first empirical dataset be synthetic, redacted from internal operations, or collected from future public repos only? - Which venue is most appropriate for the first version: arXiv technical report, workshop paper, or Zenodo-first white paper? ## Cross-references - [Paper scaffold](../papers/coordination-bus-governed-agent-infrastructure/README.md) - [Draft manuscript](../papers/coordination-bus-governed-agent-infrastructure/paper.md) - [Roadmap](../ROADMAP.md) ## Revision audit — 2026-10-10 (append-only correction) **Method/provenance:** Inspected the four linked repository source documents at baseline `0b6cd1fa9865605499d0ad11bf96adaf1f26ed68`. No operational bus dataset was collected and no reviewer experiment was run. Earlier headings labelled pre-registration describe a post-hoc reconstruction, as the original comment already states; they are not evidence of a prospectively registered study. **Observations:** The March 24 protocol is a draft JSONL design with event identities and hash chaining, while CAES describes a historical five-column TSV interface. Their co-location does not establish a deployed system with all listed properties. Both are historical materials; their founder-mode location or operational instructions must not be treated as current authority. The case-study file remains a template with no observed incidents. **Interpretation and changes:** Expanded the seed into a systems paper with a clear evidence/action boundary, a threat model, defined queue projection rules, and explicit empirical endpoints. Canonical sender strings are distinguished from authenticated principals; append-only conventions and hash chains are distinguished from externally verifiable integrity. Repeated blockers do not reset an evidenced-progress clock by default. Revision-bound evidence is required to interpret a check as relevant to current work. **Artifacts and verification:** Added `papers/coordination-bus-governed-agent-infrastructure/replay_fixture.json`, containing ten entirely synthetic events, stated replay rules, and expected answers. It contains no redacted or real bus data and is not empirical performance evidence. Revised the case-study template to separate observation from counterfactual hypothesis and document selection, missingness, redaction, and measurement precision. External primary sources checked: Fowler's Event Sourcing article and RFC 9162. Bibliography now includes only sources used in the revised manuscript. **Open evaluation requirements:** Release an ethically and operationally appropriate closed-history dataset; freeze an adjudication rubric and equal-information comparison; preregister future endpoints and sample-size rationale. Review throughput, recovery time, generality, and actual containment remain unmeasured. Repository structure validation alone cannot establish those claims. ## Executable synthetic replay — 2026-10-10 (append-only) **Scope and method:** At 2026-10-10T13:52:38Z, specified the scope of a small offline projection before implementing it. The ten-event example and its answers were already known; this is not preregistration or a held-out benchmark. Preserved `replay_fixture.json` byte-for-byte at SHA-256 `06e4046a40a771e8ef13d0b07e83ff0329437adce9924e7db94a821b01038d68`. Added a standard-library program with no bus/network access or tool execution, versioned instrument rules, tests and recorded JSON output. The sequence cutoff selects a prefix, including when two events share a timestamp. Expected answers are compared only after projection. **Synthetic observations:** The full projection reports revision r2, CI receipt e7/pass, e9 stale on arrival, and one episode from e4/minute 3 to the linked e6 artifact/minute 10: seven constructed minutes. e5 is a repeated blocker, not a new episode or progress. At sequence 5 and minute 9, the same episode is right-censored at six elapsed minutes with no resumption event. A new artifact without a link to the active blocker does not close it, and CI success alone does not substitute for that fixture rule. Receipt results remain reported observations, not authenticated or rerun checks. Permission is explicitly outside the instrument. **Checks:** Fourteen synthetic tests passed, covering oracle independence, cutoff/future-outcome separation, repeated blockers, heartbeat versus progress, exact and conflicting duplicates, invalid sequence/time/causal references, unrelated artifacts and deterministic CLI output. All six full-fixture answer values match the authored oracle. These checks show consistency with declared rules, not operational performance or human reconstruction accuracy. No real bus messages or reviewer outcomes were collected. **Interpretation and correction:** Updated the manuscript's worked example to identify the executable artifact and its limits. Corrected template Incident C: a local claim can report an outcome, and a source-linked receipt can still be false, stale or incomparable. Reported observations and independent checks need distinct provenance; neither alone grants authority. Simultaneous unrelated blockers, cancellation, corrections, real distributed admission and source authentication remain outside this demonstration. **Artifacts:** [Replay method](../papers/coordination-bus-governed-agent-infrastructure/replay_method.md), [program](../papers/coordination-bus-governed-agent-infrastructure/replay_synthetic.py), [tests](../papers/coordination-bus-governed-agent-infrastructure/test_replay_synthetic.py), and [recorded full output](../papers/coordination-bus-governed-agent-infrastructure/replay_result.json). ### Internal replay review correction — 2026-10-10 The native formal-papers reviewer supplied a counterexample: an e6 artifact with occurred time 2 could refer to e4 at time 3 and still close the episode. Under this instrument's declared same-clock causal-link convention, that input should be rejected. Added a cause-occurrence check, defaulting absent occurrence values to recorded time, and regression coverage for the reversed edge and a valid delayed e9 receipt with earlier cause e2. Passing the check establishes clock/reference consistency only, not real causality or timestamp truth. Clarified that `input_duplicate_event_ids` describes full-input validation, including duplicates outside the cutoff, and is not projected work state. The original fixture and recorded full-result bytes are unchanged. This is a documented correction following review, not retroactive preregistration or an independent empirical result.