# Engineering demonstrations

The folk change; the work needs to remain understandable. Corpus connects temporary agents to explicit responsibilities, identifiable changes and evidence another holder can inspect. These examples show particular pieces of that machinery at work: a reviewed change reaching main, unfinished work reaching another holder, and an archive import recovering after interruption.

<section class="engineering-specimen" aria-labelledby="reviewed-change">
<p class="engineering-label">01 / Recorded integration · 12 October 2026</p>
<h2 id="reviewed-change">A reviewed change in a moving repository</h2>

A review takes time. Meanwhile, other work can reach main. Corpus's job pipeline can construct a landing from the reviewed change and the newer repository, while checking the boundaries of what may change.

On 12 October, a repair to the retirement command passed separate agent review. Its candidate was identified as `7538f805`. The landing record then shows a different commit, `2b6c50d0`, constructed from that candidate on newer main. A separate check followed the landing.

<div class="engineering-receipt">
<p class="engineering-label">Selected public receipt · TRACE-5535 / REQ-856</p>

```text
LANDED REQ-856: refs/heads/main 23f8f4f5ffb4 -> 2b6c50d0afaa (constructed of reviewed 7538f8053a1d)
AFTER-LAND PASS REQ-856: 1 of 1 item(s) passed at 2b6c50d0afaa; the landing stands either way
```

</div>
<figure class="engineering-schematic">
<div class="engineering-merge"><div class="engineering-inputs"><span>Reviewed candidate</span><span>Newer main</span></div><span class="engineering-arrow" aria-hidden="true">↓</span><div class="engineering-node">Constructed landing</div><span class="engineering-arrow" aria-hidden="true">↓</span><div class="engineering-node">Separate later check</div></div>
<figcaption>Schematic of the recorded integration, not a run screenshot. The reviewed candidate and landing are distinct commits.</figcaption>
</figure>

The evidence here concerns this job route and this change. Normal landing does not run a fresh pytest gate before main advances. The separate PATCH route has a recorded concurrency and receipt asymmetry; preserving reviewed changes does not establish that the combined program is correct. This episode does not demonstrate section merging or crash recovery.

[Follow the contract, review and integration receipt](/evidence/#the-contract-req-856).

</section>

<section class="engineering-specimen" aria-labelledby="work-across-holders">
<p class="engineering-label">02 / Observed continuation · 12 October 2026</p>
<h2 id="work-across-holders">An obligation reaches the next holder</h2>

A useful handoff tells the next holder what remains to be done and where the unresolved judgment lies. In one October exercise, a Scout inherited an unfinished revision of a large command survey. The predecessor pointed to the critique before the draft. The successor revised the document, committed the change, recorded closure and notified its owner. A later card change shows that the result was taken up.

The obligation was ORD-scout-20. FIND-5389 reports the successor's sequence; document commit `a25bf77b`, closure FIND-5388 and notification MSG-11061 corroborate the resulting work. These records establish continuation without certifying every row of the final survey.

<figure class="engineering-schematic">
<ol class="engineering-chain"><li>Unfinished draft</li><li>Preserved critique</li><li>Directed handoff</li><li>Successor's document change</li><li>Closure and owner action</li></ol>
<p class="engineering-branch"><span>Alongside the continuation</span> Recovery instruction omitted → separately restored</p>
<figcaption>Schematic reconstruction of recorded events, not an original transcript or a single uninterrupted run capture.</figcaption>
</figure>

The handoff was imperfect. A separate recovery instruction disappeared and had to be restored. FIND-5389 and FIND-5390 retain that loss; commit `01f0de01` records the repair. The record supports one observed continuation of real work during a deliberate exercise, with that loss alongside it. It does not supply a success rate for succession.

[Read the succession account and its corrections](/work/succession-and-retirement/).

</section>

<section class="engineering-specimen" aria-labelledby="interrupted-archive">
<p class="engineering-label">03 / Controlled historical repair · 5 October 2026</p>
<h2 id="interrupted-archive">An archive survives an interrupted import</h2>

A session archive can be changing while another process tries to preserve it. Corpus found a race in which importers shared staging names, and a later import could encounter files that no longer matched their recorded metadata.

The repair serialized cooperating importers, separated their staging files and allowed recovery when the already recorded prefix could still be verified. Its regression tests included killing an importer during publication and checking that the next import completed with matching checksums. A separate agent verified the named repair commit.

<div class="engineering-receipt">
<p class="engineering-label">Historical verification · FIND-4553 / PATCH-1398</p>
<dl><div><dt>Separate verification</dt><dd>TRACE-5092</dd></div><div><dt>Verified commit</dt><dd><code>1cd979125d69f1340d0f90cc8f205ae9d115822e</code></dd></div></dl>
</div>
<figure class="engineering-schematic">
<ol class="engineering-chain"><li>Recorded prefix</li><li>Publication interrupted</li><li>Next import verifies prefix</li><li>Matching metadata published</li></ol>
<p class="engineering-branch"><span>Damage branch</span> Recorded bytes altered → recovery refused</p>
<figcaption>Schematic of the controlled reproduction, not a production incident or run screenshot. Verification belongs to the historical repair.</figcaption>
</figure>

This is evidence from a controlled historical repair. The concurrency lock is Linux-specific, the captured files do not form a globally simultaneous snapshot, and a repository archive is not an off-machine backup. The tests establish the named recovery behavior; they do not supply a long-term reliability rate.

[Read the implemented mechanism and its limits](/readings/implemented-mechanisms/#6-resuming-an-archive-import-while-transcripts-keep-growing).

</section>

<section class="engineering-further" aria-labelledby="more-engineering">
<h2 id="more-engineering">More work to inspect</h2>

- **[Measurement and correction](/work/tests-under-the-floor/).** A faster test setup broke a canary's fixture. The subsequent repair retained the guard the canary was supposed to exercise. The account keeps the regression, mistaken explanations and measurement limits visible. A [distinct study withdrawal](/studies/pilots/agent-instructions/#initial-comparison-and-withdrawal) shows a different kind of correction; neither supports a composite speedup claim.
- **[Native-provider execution](/portfolio/corpus/#provider-architecture).** A common job contract sits above host-specific implementations of permissions, isolation and process lifetime. The providers' different behavior shaped the implementation and the choices behind it. The architectural decision is evidence of a particular boundary, without certifying all providers or establishing completion of the original live-viewing goal.
- **[Publishing this public edition](/readings/implemented-mechanisms/#10-making-the-private-to-public-boundary-executable).** A build can pass and still lose a required file during transport. In a recorded October deployment attempt, a second check refused that incomplete artifact before upload. FIND-6627, PATCH-1691 and completion record FIND-6628 document the refusal, repair and reported completion. This is a recorded account, without retained raw output for the final episode.

</section>

<p class="engineering-source-note">Selected public passages and receipt fields, prepared 20 October 2026. Separate agent review means review within Corpus, not an external audit. The full working records remain private; identifiers preserve provenance without releasing those records.</p>