v0.4.0-draft.1: Real owned use and native recovery

Real Owned Use, Native Recovery, and the Full Manuscript

v0.4.0-draft.1 is a benchmark and evidence prerelease. The Spec and core wire
revision are unchanged; this is not a standards endorsement or product launch.

Actual Results

The deployed model produced six report candidates in the preserved amended
read-only study. Two met the frozen task contract and same-project source review;
four oversized candidates remain failures. This continuation does not rerun the
model, shorten failed answers or relax their grade.

The two eligible reports were delivered to
one dedicated owned issue,
using Octokit and GitHub's official pinned MCP server. Six distinct controlled
faults test current revocation, narrowing, real process kills, acknowledgment
timeout and actual stdio disconnect.

• 20 fresh workers: 8 initial executions, 4 read-only recoveries, 8 duplicate probes.
• 5 native comment calls and 5 objects: 2 ordinary reports and 3 synthetic controls.
• Both current-policy denials send nothing.
• Three delivered-but-unacknowledged effects recover through reads, without resend.
• The pre-send crash retains unknown; absence of evidence is not permission to retry.
• All 8 duplicate probes make no new native comment call.
• 21 journals, 283 records and 78 historical source entries reconstruct successfully.
• Six anonymous GETs match the issue and five exact comment commitments.

The supervisor span is 52.437 seconds including fresh-worker startup, shadow
preparation and networking. Earlier issue creation/setup are separate. These
unequal tasks and fixed-order shadows do not support a latency superiority
claim. The earlier matched offline comparison's 3.97% protocol overhead remains
reported, not replaced by a fast-path-only timing.

Read and Reproduce

• 43-page, 9-figure manuscript PDF, with semantics, algorithms, conditional invariants, counterexamples, baselines, costs, failures and five appendices.
• Readable manuscript.
• Results and exact evidence index.
• Versioned no-send reproduction guide.
• Local verification and manuscript checks.
• Clean-bundle verification, including the local OPA download failure.

Attached source-history bundles preserve the historical blobs needed for replay.
GitHub's automatic source ZIP is not enough for history-dependent reconstruction.
Use the full public clones or the two attached bundles, verify SHA256SUMS,
then install the locked dependencies. Replay needs no model key, provider login
or OSuite account. Do not execute new live writes on the existing evidence issue.

Verification and Scope

Release source: 6d4f3c358ca5f68040a1416d751328e9aab041f5.
Original live execution source: f93b1cfcbf2c2665b4acc7c5dc8fb92d1ba8d6b6.
Spec pin: a9b86f4b529a5e330c2a9bf1b8c4c259a7b36e40.

All 1,623 local software methods pass, with no failures/skips; both
Linux and macOS CI jobs
pass on the release commit. These are software checks, not 1,623 independent
customer workloads. The complete paper was rendered and visually checked.

Both source bundles were cloned into a clean directory and the owned-write
replay, arithmetic audit and two old field replays passed. The local default
OPA download failed; the continuation used an already available binary only
after checking its exact pinned hash. This is not a successful fresh local
binary download. The separate CI installation steps passed.

This is actual same-project use of real GitHub infrastructure, not independent
human evaluation or customer production. Faults were deliberately induced.
Current-content reconciliation is conditional, not global exactly-once delivery,
atomic remote revocation or observation of all notification/webhook descendants.

The first pre-reservation failure, four task-budget failures, old releases and
raw measurements remain available. No Studio/Hook modification, new Azure
resource, protected collaboration/customer data, invitation or formal journal
submission is included.

Public Download Check

All ten attached files, including the checksum manifest, were downloaded
without GitHub credentials at 14:23:20--14:23:35 UTC on 19 October 2026.
Every response was HTTP 200 and every file matched its local SHA-256.
The downloaded manifest also verifies all nine payload files.

An earlier direct CLI attempt timed out. The successful check explicitly used
the machine's already configured system proxy, without changing network
settings or supplying a GitHub login. An additional anonymous API download of
the PDF matched the same hash. These observations demonstrate the published
bytes are publicly retrievable under the stated route, not universal network
availability. No earlier asset, raw result or source tag was replaced.

v0.3.0-draft.1: Real Maintainer Use and Full Manuscript

Real Maintainer Use and Full Manuscript

Bench prerelease v0.3.0-draft.1. Editorial date: 19 October 2026,
America/Toronto. This is a benchmark/evidence release, not a new core wire
version or an endorsed standard.

What Is New

Real model-led maintenance tasks now exercise public GitHub sources through
Octokit and the official GitHub MCP server. A separately recorded follow-up
uses the same six intents after contract and evidence-visibility repairs.
Both runs retain all outcomes; they are internal exposed work, not twelve
independent tasks or an outside evaluation.

The full manuscript explains coupled effects, native/context/profile bindings,
ordered independent checking, conditional invariants and counterexamples,
current authority, bounded observation, durable occurrence handling, refusal
ownership, complete baselines, cost boundaries and the new field findings.
Its appendices map the protocol obligations to implementation and replay.

The new replay reconstructs captured model/tool control flow from retained
journals and reads verified historical Git blobs as data. Source-based quality
review remains distinct from these mechanical checks. Normalized release
metadata without a retained native envelope remains an explicit gap.

Read the Evidence

• Full manuscript and PDF.
• Real-use findings, failures and readiness.
• Read-only replay instructions.
• Release verification.
• Controlled live mutation pilot, unchanged.
• Publication scope.

Compatibility and Limits

Spec remains pinned to a9b86f4b529a5e330c2a9bf1b8c4c259a7b36e40 and its
v0.2.0-draft.1 release. Core wire remains 0.1.0-draft.1. No old expected
case, report, source tag or release asset is replaced. The earlier nine-page
paper remains attached to the earlier Bench release.

New real-use work is read-only. It does not qualify the official MCP mutation
handler, autonomous write completion, hosted MCP authentication, complete
provider descendants, multi-host recovery or sustained customer production.
The older twelve actual comments are not reassigned to this new population.

No Studio, Hook, customer action, protected collaboration data, cloud resource
creation, new external invitation or formal manuscript submission is included.
The release carries no performance-superiority or independent-certification
claim. Replay requires no model/provider credentials; a new live model run
requires separately configured access and incurs costs.

Verified Publication

Source: 86abb1e0c25238a0bbc22c59d45d06e983527bbf. Linux/macOS CI passed before this release. The 39-page PDF, standalone TeX, Markdown manuscript, results index, and verification record are attached with SHA-256 checksums.

Download the full PDF. The automatic source ZIP is not a self-contained historical replay bundle: use a full Git clone and the pinned Spec sibling so replay can read original source blobs. No model or provider credentials are needed for captured-evidence replay.

Action-Effect v0.2.0-draft.1: Runtime Pilot and Working Paper

Action-Effect Runtime Pilot: Public Working Draft

This prerelease adds a runnable MCP/SDK effect-description, real OPA decision,
and controlled GitHub execution/readback path. It does not require OSuite.

• An explicit opt-in contract clarification and bounded GitHub-comment profile
  are published in the companion Spec release.
• Separately written JavaScript/Python paths and a complete structured baseline
  each match 108/108 development mapping checks and 52/52 eligible decisions.
• All 32 planned live checks match, with 12 real public comments, four
  read-only lost-ack recoveries and four suppressed duplicate requests.
• Accepted-input cold median cost is 50.391 ms for the protocol versus
  48.468 ms for the structured baseline. This is not a speed-superiority claim.
• Raw attempts, failures, source pins, read-only replay, a working paper,
  adopter instructions and separate seal/run/score tooling are included.

Results and replay
| Working paper
| Quickstart

These are same-project, bounded development results. External independently
authored evaluations remain zero. Direct comment readback does not establish
complete provider provenance or transitive effects. This is not an adopted
industry standard, certification or qualified production gateway. Original
frozen artifacts and the failed first offline attempt remain preserved.

Action Effect Bench v0.1.0-draft.1

First public working draft of Action Effect Bench.

• Pins Action Effect Spec commit ff5de37da6402c5a0e42d1a9cc7d4880579e374d.
• 80 frozen public-development cases plus 22 explicitly post-review regressions.
• JavaScript and Python each match 102/102 expected results, with 102/102 agreement and zero adapter faults in the recorded release run.
• 144 unit/harness/entry-point tests pass; clean remote-clone reproduction and Linux/macOS CI pass.
• The public report records 52 stable source hashes and 204 raw consumer responses. Earlier published reports and the entry-point correction remain visible.

Start with the report index and the pinned reproduction instructions in README.md.

This is finite, same-project offline development evidence, not an independent institutional evaluation, protected holdout, authorization benchmark, production safety claim or full core-conformance certificate. Timings include process startup and are not a speed ranking. All responses retain authority: not_evaluated.