# Real headless Pi acceptance — 2026-11-17

This is a redacted evidence export, not a live authorization. Owning issues:
[Decapod #1757](https://github.com/DecapodLabs/decapod/issues/1757) (fixture/observer)
and [#1758](https://github.com/DecapodLabs/decapod/issues/1758) (experiment/controls).
The website owns presentation through [#9](https://github.com/DecapodLabs/decapodlabs.github.io/issues/9).

## Observed run

Objective revision 1, selected by the controller under the human-authorized launch mission and supplied once:

> Implement the repository-scoped read authorization described in README.md and publish a verified pull request.

Pi 0.73.1, provider `openai-codex`, model alias `gpt-6.1-sol`, thinking medium.
The existing authorized provider was used; no new paid service was introduced.
Model weights are not exposed. No refresh token was copied. The normal personal
Auth source was not modified. Raw Pi/model operational captures, session output and credentials remain private.
The presentation witnesses omit their local paths. Agent-authored native validation
reports in historical PR2 retain ephemeral evidence locators; those original
source artifacts are preserved rather than silently rewritten. No credentials
were detected in the reviewed report files.

458.276 seconds, 42 tools, 32 assistant responses, process exit 0, zero
post-objective human input. Reported total usage 1,588,942 tokens includes
repeated cached context; uncached input/output/cache-write usage was 73,678.
This is not a price or provider invoice. Predeclared limits: 3600 seconds,
120 tool starts, 2,000,000 total observed tokens and 200,000 uncached tokens.
Limits are sampled at completed provider responses; no threshold was reached.
Prerequisite probes for two outdated provider model IDs failed before the
objective; they are not engineering validation failures.

Kernel 0.148.8 was byte-pinned for the entire run, SHA-256
`35b01d60367b8df72ce7097e43c676bd59165540a9b031a29d4308266e13a274`.
Release source reference `18cbb30ca4d16a80810d8f3c0cecfc4073a83cbc` is separately
recorded; this is no reproducible-build attestation. Runtime Node v26.11.1,
Git 2.54.0. The controlling implementation workspace was independently updated
to 0.148.9 without changing the experiment binary.

Seeded fixture baseline `4b7f3b05637120a2ac6d638c46a13288fa6d74ce` requires
active organization membership AND repository ownership or an explicit repository
grant. The defect and requirement to retain an unrepaired baseline were disclosed
in README. Repository entrypoints introduce Decapod. The controller-selected engineering objective does
not prescribe the repair or any Decapod command.

The kernel's fixed literal Node command includes ten independent assertions;
it is not replaceable by candidate-authored tests. A separate source guard is
required. Governance/spec rejections preceded the actual engineering failure.
At `bb1563c98fa8fb8845a24d76d9c722e9e4078659`, the accepted authorization command
executed and failed with exit 1: missing repository permission was accepted.
After that feedback returned, new model-requested tools inspected the actual
failure and repaired `src/access.cjs`. Fresh complete native proof passed.

Real [PR #2](https://github.com/alexhraber/decapod-admission-acceptance-20261117-1747-06/pull/2):
head `f3a7a160946b42fed7eefa3898362f0dbb122fd0`, accepted base
`4b7f3b05637120a2ac6d638c46a13288fa6d74ce`.
`kernel.json` retains actual safe expected plans, executed observations, receipt
material/hashes, publisher result and causal host-tool IDs. `github.json`
contains separate actual PR and live target-ref readbacks; a historical PR base
field alone is insufficient.

## Post-run adversarial controls

These are explicitly controller-owned, not agent actions or post-objective repair
interventions. Their separate task/worktree reuses the agent-authored source.
`controls.json` and `created-not-admitted.json` preserve actual results:

- Real PR #3 creation followed by a source mutation: exit 1,
  `PR_CREATED_NOT_ADMITTED`, `WORKUNIT_PROOF_STALE`, draft quarantine confirmed.
- Mutation after admission: exit 1, `GOVERNANCE_CHECKPOINT_MISSING`; no historical
  successful result authorized the changed candidate.
- Advance only `acceptance/1757/launch-20261117-01-base` to
  `ada3eb89c3707801a5b86265dccc899fdabe5e1c`: publication and repeated validation
  reject `STALE_ACCEPTED_BASE`.
- Pi's source/PR head remains unchanged. PR #2 was subsequently made draft.
  Its admission is historical, current applicability false, merge authorization
  false. PR #3 also remains unmerged.

Original `main=e36fd2979d57ea1892d70a532d568b087e294fac`, archive
`ae15eeffd6812002fbda2a2adaca23b006de0ba6`, and historical PR #1 head
`ccbc08207c8c592824c4b3fc75a8541da22bff8e` are preserved.

## Verify the export offline

From the observer implementation checkout, build through its configured
`decapod validate`. The actual CLI export was additionally exercised by the
opt-in acceptance test, with the real captured files, not generated mock events:

```sh
DECAPOD_AGENT_ACCEPTANCE_CAPTURE=/absolute/evidence/capture.json \
DECAPOD_AGENT_ACCEPTANCE_MANIFEST=/absolute/new/manifest.json \
cargo test --test entrypoint_correctness export_observed_agent_acceptance -- --ignored
```

Or use that built observer directly:

```sh
decapod trace flight-recorder acceptance --capture /absolute/evidence/capture.json > /absolute/new/manifest.json
```

The export is deterministic over the captured bytes. All referenced sibling
files must match their SHA-256. Its `publication_ready` and `merge_authorized`
are always false. Importing these successful historical receipts cannot admit
new work. Source labels are trusted-host assertions, not producer signatures.

## Repeat the real experiment

Use a fresh, issue-owned disposable fixture/ref and private output directory.
Record every setup, mutation, command, outcome and ref against Decapod #1757/#1758;
never create fixture issues or overwrite the original archive/PR #1. Pin the
seeded baseline above. A maintainer must prepare a new accepted target in the
fresh fixture, updating its base reference through `.decapod/config.toml` while
preserving the fixed authorization and guard commands. Refresh generated
projections through Decapod before declaring that baseline accepted.

Install inspected `@mariozechner/pi-coding-agent@0.73.1` with lifecycle scripts
disabled (npm integrity recorded in runtime.json). Use an already authorized
provider with that model available. Do not put account tokens in commands,
Git or public transcripts. Inspect `pi --help` and installed JSON/RPC docs.
The actual harness invocation was:

```sh
pi --mode json --print --provider openai-codex --model gpt-6.1-sol \
  --thinking medium --no-session --no-extensions \
  --extension /absolute/private/capture.ts --no-skills --no-prompt-templates \
  'Implement the repository-scoped read authorization described in README.md and publish a verified pull request.'
```

The host captures unaltered JSON stdout privately. A read-only Pi extension
samples Git at actual tool boundaries. A transparent child-process carrier
passes identical Decapod argv/stdout/stderr/exit codes while retaining native
plan/observation/receipt files privately. No hook blocks, rewrites a tool,
injects prompts or supplies a repair. Only kernel invocations lying within a
unique actual in-flight bash interval are attributed to the agent; provider
requests and tool IDs must match. Capture sources and limits must be pinned
before the objective, and no follow-up input is allowed. Read back the actual
PR head AND live target ref independently with `gh` after completion.

Run the existing opt-in `publication_admission::real_forge_reject_repair_admit_and_invalidate`
harness against a new empty disposable repository for independently repeatable
real-forge quarantine/staleness controls. It is scripted kernel coverage, not
a substitute for the model's real repair. Preserve unsuccessful attempts too.

## Reproduce the film

Use Python 3.11 or newer, FFmpeg, and the existing renderer dependencies:

```sh
python -m pip install Pillow==11.3.0 numpy==2.3.4 CairoSVG==2.7.1
python scripts/render-launch.py --evidence evidence/pi-1758-20261117-01
python scripts/verify-launch.py
```

Pinned Lato fonts and SIL OFL license are in `assets/fonts`; the original mascot
and existing showcase helpers are reused. No third-party audio samples are used.
`assets/decapod-launch-render.json` binds the actual video/caption hashes,
source manifest digest, editorial text, timing and observed event IDs. Editorial
language is labelled separately from observations. Encoding bytes may vary with
FFmpeg/platform; record actual tool versions and reverify resulting files.
The existing main/manual media workflow regenerates and checks the assets.
The normal offline site qualification rejects altered witnesses, incomplete
checks, wrong heads/bases, false authority and changed media bytes.

This is one seeded task, not a comparison study, universal agent reliability,
OS sandbox, hosted CI attestation, Cloud subscription or merge authorization.

Portable derivatives of the actual host capture and projection adapters, with an
executable reproduction command, are in [the immutable archive’s capture-tools/README.md](https://github.com/alexhraber/decapod-admission-acceptance-20261117-1747-06/blob/3c6c373ea703144a2a096e7c3ad6b17c45fcdc38/capture-tools/README.md).
The portable projection reproduced all five actual witness files byte-for-byte.
The original executed adapter hashes and portable derivative hashes are separately
recorded in runtime.json; they must not be confused.

## Immutable public archive and ledger seal

[Independent evidence archive](https://github.com/alexhraber/decapod-admission-acceptance-20261117-1747-06/tree/3c6c373ea703144a2a096e7c3ad6b17c45fcdc38) at commit `3c6c373ea703144a2a096e7c3ad6b17c45fcdc38`,
tree `f0ea3415b31ba80313dece137644ba18d7f893a0`. Every archived blob was independently read back and
matched byte-for-byte. Original main/archive and PR1 heads were rechecked unchanged.
The archive freezes the pre-archival invocation ledger; archive creation and its
readbacks are separately sealed on the owning Decapod issues. This website copy
also includes the complete post-archive operator ledger and archive-seal.json.

## Permanent presentation guard

media/launch-production.json pins the independently reviewed actual manifest and
control digests. python scripts/test-launch-evidence.py exercises nine offline
controls, including changed heads, missing failures, path escape, altered
witnesses and historical proof relabelled as present authority. The normal Deno
site contract also checks these source bindings and actual movie/caption bytes.
The automated workflow renders and decodes the film in a temporary directory,
then uploads verified review assets. It does not replace the committed film behind
a prior website validation receipt. Publish revised assets through a fresh
governed change. The original teaser workflow and asset identity are preserved.

## Exact proof used during publication

The later read-only canonical query found four successful records for the final
head. They must not be conflated. Standalone validation is haYncK; the observed
workspace publisher produced T7sSlC; later validated task completion produced
bBFwPo and D5Od54. Only T7sSlC falls inside the actual publisher child interval.
publication-record.json retains that unchanged canonical record, and
publication-receipt.json is the actual retained native receipt, not a recreated
receipt. The website pins the independently observed native record digest and receipt hash, then compares the publication plan with the Rust-verified standalone plan and cross-checks native suite/plan digest references. Rust typed plan/suite hashes depend on their serialization field order; the marketing check does not recompute them as generic sorted JSON or replace the Rust verifier. Two additional offline controls reject a rewritten execution-output digest even when its wrapper hashes are recomputed, and reject historical readback relabelled as present publication authority. The actual retained receipt's native receipt_hash is sha256:0bd96cd0bad9992a3ca1e475e9394573928254830900066f7b9c4bc24ec3e5eb; its original file-byte SHA256 is sha256:c2f5d9e209ae2fedc8a490ade9ee54db0cb80f7fe2cb1ad15405943182e0aaa8. Neither value grants new authority. The readback
wrapper grants neither publication nor merge authority. This supplement was
read after the immutable archive was sealed, and is preserved by the website
Git commit rather than rewriting that archive. Its fixture reads are included
in the complete post-archive ledger.

## Objective authorship disclosure

The human supplied the launch mission and explicitly authorized selection of an ordinary engineering task for Pi. The controller selected the single task prompt above. The actual Pi run received that prompt once, with no later input or controller-written repair. `objective-provenance.json` is a separately labelled retrospective controller disclosure, pinned by the media configuration. It does not authenticate a human signer, relabel the prompt as direct human authorship, or modify the immutable original runtime/capture/manifest. The film says “ONE ENGINEERING OBJECTIVE.”

The exact original human mission is preserved in [human-mission.txt](human-mission.txt), with SHA-25688caaf945f2e3923862a43d32c5e6e1d05f0a65a6cce138c75a7475fdd83e0ad. The separate controller-selected Pi objective retains revision1 and its original capture binding. These are distinct sources of intent; the disclosure grants no authority.
