← Decapod Labs

An objective.
A repair.
The receipts.

The case study

The 56-second case study follows the run shown in the native terminal recording, with the original scene timing and soundtrack.

Watch the native Pi terminal recording · 12-second cut

The recorded Pi run

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

Pi 0.73.1 with gpt-6.1-sol and Decapod 0.148.8 executed 43 tool calls. The film renders Pi’s recorded terminal stream with larger type and expanded results. Pi received one objective and no follow-up input.

Failed candidateed086de6691d92da623febfaa6b411514828440e
Validated and published heada063de9040c458204eace1929d1e8536ad048a88
Accepted baseada3eb89c3707801a5b86265dccc899fdabe5e1c
Fresh validation receiptsha256:bcd043bc1cec93ccb51d227c699ef137014bb54ab7714518404c60de8b888542
Independent forge readbackGitHub PR #7

Recording metadata · Frame timing and digests · Observer manifest · Native checks and receipts · Independent GitHub readback · Reproduction and capture limits

Admission grants no merge authorization.

The earlier fixture run

The original launch film was based on this first independently observed run. A read-only Rust observer correlated actual provider tool requests, host execution, kernel records, Git state and independent GitHub readback. Model narration is not completion evidence.

The objective

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

Under the human-authorized launch mission, the controller selected this single engineering objective for Pi. It was supplied once, with zero post-objective input. Direct human authorship of that one-sentence prompt is not attested. Pi discovered the governance workflow through repository entrypoints. The baseline defect and the requirement to retain a failing baseline were explicitly seeded in the repository; the controller supplied no repair.

AgentPi 0.73.1 · openai-codex · gpt-6.1-sol
KernelDecapod 0.148.8, byte-pinned throughout the experiment
Execution458.276 seconds · 42 tool calls · 32 assistant responses
Required proof10 fixed authorization assertions and an independent source guard

What actually happened

  1. The agent consulted Decapod, claimed a task and entered an isolated workspace.
  2. Initial specification/governance rejections provided actionable feedback. These do not count as the engineering failure.
  3. The accepted authorization check executed and failed: a member without a repository grant was incorrectly allowed access. The process exited 1.
  4. After receiving that failure, the agent inspected the actual output and changed repository-scoped access logic.
  5. Fresh execution passed all required checks. Decapod published and verified real GitHub PR #2. Independent readback matched both the validated head and accepted target.

Validated head: f3a7a160946b42fed7eefa3898362f0dbb122fd0
Accepted base: 4b7f3b05637120a2ac6d638c46a13288fa6d74ce
Failed candidate: bb1563c98fa8fb8845a24d76d9c722e9e4078659

Inspect the exact publication-time receipt and canonical completion record. This proof was created inside the observed publisher invocation. Standalone validation and later task-completion records remain distinct.

A green result expires

Separately labelled operator-owned probes followed the completed agent run. They did not change the agent’s source branch or supply its repair.

PR #2’s admission is historical; its current applicability is false. Its exact head is preserved, and it was made draft after the target control. No merge is authorized. The original archive and historical PR #1 remain untouched.

Inspect and reproduce

Observer manifest · Objective provenance disclosure · Original human mission · Capture index · Native checks and receipts · Adversarial results · Reproduction and fixture ledger

Implementation and experiment ownership: Decapod #1757 and #1758. Film production: website #9.

Inspect the immutable GitHub evidence archive ↗

Assurance boundary

Digests bind captured bytes; they do not authenticate a producer. Host provenance is a trusted operator assertion with bounded tool-boundary observation. This is no universal OS sandbox or guarantee about uncaptured activity. The model provider exposes an alias, not immutable weights. This small seeded task is one observed run, not comparative evidence of general agent reliability. Native local checks do not attest a hosted CI matrix. PR admission never authorizes merge.

The kernel enforced admission through its native publisher. The observer only reports what it captured and cannot admit a PR. No independent GitHub enforcement of the fixture’s admission policy or prevention of every direct forge bypass is claimed. Offline observer regressions cover bypass and false-completion classification; the kernel source repository’s required CI policy is a separate control.

No Cloud purchase or hosted onboarding success is depicted. The local-first demonstration requires no Cloud subscription.

Main film · Social cut · Vertical film · Captions · Watch with accessible controls