Skip to content

Streamlined execution amendment

Control: MP-STREAMLINED-EXECUTION-001
Authority: Human Steward directive, 2026-09-01
Status: binding on protected merge; supersedes contrary routine-process requirements

Purpose

Remove administrative blockers that do not protect a material mathematical, certification, security, provenance, or authority boundary. This amendment is deliberately narrow. It does not create a new control plane.

1. Delegated execution replaces routine approval ceremony

For bounded administrative, documentation, engineering, workflow, maintenance, routing, synchronization, and campaign-execution changes within already-authorized scope, standing delegated authority is sufficient to execute the transaction through protected merge and readback.

Routine work does not require a fresh Human Steward approval or a fresh independent-review approval merely because a new commit, synchronization commit, or protected-main commit exists.

Where separation of duties is materially required, the review function may be performed by an appropriately skilled non-author delegated agent. This is a delegated operating role, not a per-transition Human Steward approval gate. A Human Steward disposition is required only where authority is expressly reserved and cannot be delegated by the governing instrument.

Independent specialist review remains appropriate for substantive mathematical certification, source-semantic adjudication, constitutional authority expansion, security-sensitive protection weakening, or external claim promotion. Such review is attached to the material object under review; it is not invalidated by unrelated repository activity.

2. Evidence-closure identity

Governed evidence binds to a material evidence closure, not indiscriminately to the whole repository head.

The closure consists of:

  • the governed changed bytes;
  • directly consumed protected artifacts and their content identities;
  • validators, schemas, workflows, toolchain pins, and policy bytes that can affect the conclusion;
  • the declared authority and claim scope.

A later commit invalidates prior evidence only when it changes that closure, creates a merge conflict affecting it, or changes the authority/claim boundary relevant to the transaction.

Unrelated movement of main, unrelated documentation, independent campaigns, or byte-identical synchronization merges do not invalidate checks, delegated dispositions, or specialist review. No rebase, synchronization merge, rerun, or re-approval is required solely to make a branch numerically current with main.

The existing administrative-maintenance distinction between repository head and material artifact identity is controlling for concurrent development.

3. Concurrent development rule

Protected branches may advance concurrently. A candidate may merge from an older base when all of the following hold:

  1. GitHub reports the candidate mergeable without a material conflict;
  2. required checks for the candidate's material closure pass;
  3. no relevant protected dependency in that closure changed since validation;
  4. the candidate does not widen its governed scope or authority.

If protected main advances only in disjoint material closures, the candidate remains valid. Agents must not manufacture synchronization commits merely to refresh the base SHA.

4. CI proportionality

Pull-request CI must be impact-routed.

  • Cheap classification and syntax/schema checks may run on every candidate.
  • A substantive shard runs only when its governed inputs or validator dependencies changed.
  • Lean replays, external repository replays, full-estate conformance, publication builds, and other expensive checks do not run on unrelated changes.
  • A required status context for an unaffected expensive lane may complete by a cheap content-identity/impact determination that records UNCHANGED_ATTESTATION_REUSED or equivalent.
  • Full sentinel/replay runs belong on relevant material changes, explicit dispatch, or low-frequency scheduled assurance, not on every tiny PR.
  • Unknown impact fails closed into the smallest defensible affected set; it does not automatically imply a full-estate replay unless the classifier itself or its dependency map is affected.

The following execution rules are now part of that proportionality principle:

  1. Bound commands. A policy-shard command receives the registered default timeout. A longer exception must be bound to the exact shard, command index, argv, timeout, and written reason. An ad hoc long-running command is not an acceptable substitute for a governed exception.
  2. Bound test modules. Aggregated unittest execution must bound both individual modules and the aggregate suite, terminate the process group on timeout, and identify the active module in the live log.
  3. Do not duplicate owned suites. A repository-wide regression lane is complementary residual coverage. Tests already owned by a dedicated shard should be excluded from repository regression and exercised through that shard instead. A full sentinel composes the dedicated shards; it does not prove more by executing the same expensive suite twice.
  4. Separate cheap validation from scientific replay. When a test family contains substantial symbolic reconstruction, bounded search, exact linear algebra, external replay, or independent producer/verifier recomputation, the cheap structural checks and the expensive computational closure should be routed separately by material dependency.
  5. Preserve independence where it matters. Performance optimization must not share producer state with an independent verifier merely to save time. Prefer dependency routing or protected evidence identity over weakening a substantive independence property.
  6. Keep deliberate sentinels. Material-input routing is not permanent exemption. Scheduled or explicitly dispatched full sentinels continue to exercise the complete expensive closure so that dependency maps and replay paths themselves remain tested.
  7. Fail closed on ambiguous computational dependencies. A changed helper or source that cannot safely be assigned to a narrow computational stage must invalidate the conservative downstream replay set.

Odd Zeta operating precedent

Odd Zeta establishes the current reference pattern for expensive computational mathematics in CI.

The OZ lane contains ordinary fast tests and a measured set of expensive modules that reconstruct symbolic objects, execute bounded searches/rank computations, and independently replay mathematical producers and verifiers. Those expensive modules are not ordinary syntax or regression checks.

Accordingly:

  • the fast OZ suite runs whenever the OZ lane owns the transition;
  • expensive OZ modules run only when their material computational dependency closure changes;
  • cumulative stage dependencies propagate downstream conservatively;
  • Odd Zeta campaign-source changes enter the replay router through the campaign lane rather than escaping under a different top-level classifier label;
  • shared computational helpers fail closed to the applicable downstream closure;
  • scheduled/manual sentinels execute the complete expensive replay set;
  • no producer/verifier mathematical state is shared and no cached mathematical conclusion substitutes for required replay.

The admitted implementation reduced an unaffected OZ transition from roughly 1024 seconds to 22.161 seconds while still running 24 fast OZ modules and explicitly recording all 12 expensive modules as materially unchanged. This timing is evidence about CI proportionality only; it does not alter the mathematical status of any OZ result.

The same architecture should be preferred for future computational research lanes when the expensive work has a defensible material dependency closure. Do not copy the specific OZ stage map into unrelated domains; derive the narrowest conservative closure from each domain's actual dependencies.

New required CI must identify its trigger set and expected marginal cost. A new routine check should replace, subsume, or demonstrably justify existing cost rather than simply accumulate.

5. Merge and readback

Routine bounded execution sequence is:

classify material closure -> run affected checks -> exercise delegated disposition -> protected merge -> protected readback.

There is no generic independent-approval stage and no generic Human Steward approval stage.

A protected readback establishes the merged result. It does not trigger a second approval cycle.

6. Supersession

For routine bounded work, this amendment supersedes contrary requirements in earlier maintenance plans, handoff packs, runbooks, issue templates, PR prose, or design documents that require:

  • fresh independent approval on every exact commit;
  • fresh Human Steward approval where standing delegation already covers the action;
  • invalidation solely because protected main advanced;
  • synchronization solely to make a base current;
  • full or expensive CI for an unaffected material closure.

Historical records remain historically accurate; they need not be rewritten. Active automation and current operating documentation should be reconciled mechanically when encountered.

7. Non-expansion boundary

This amendment reduces ceremony; it does not promote mathematics, certify a theorem, alter source provenance, widen a campaign claim, weaken security controls, or authorize external claims. Delegation cannot be used to manufacture substantive mathematical independence where the same actor produced and certifies the same mathematical claim.