# Evidence extracts: Shed Happens OS + APEX

**Evidence snapshot · October 27, 2026 · Public sharing approved**

These short extracts accompany the case study. They preserve the meaning of saved records while excluding private paths, accounts, coordinates, infrastructure and customer identifiers. They are not new tests. Originals and their SHA-256 hashes are retained in a separate private manifest.

## E1 — Approval isolation: a behavior that failed, then passed

The stored pre-repair output includes this direct behavioral reproduction after an earlier fixture-method error:

> pending BEFORE poll(): ['Escalation A (never replied to)', 'Escalation B (never replied to)']
>
> pending AFTER poll(): []
>
> DEFECT CONFIRMED: one top-level reply resolved 2 of 2 stacked escalations non-selectively (neither was actually replied to in its own thread)

The stored corrected fixture output records:

> PASS LEG A: Escalation B (the one actually replied-to) resolved
>
> PASS LEG A: Escalation A (no reply in ITS thread) remains pending
>
> PASS LEG A: exactly ONE subprocess action ran (B's, not A's)

The corrected fixture also exercises the opposite request order, stray top-level replies, later rejection on another thread page, and an incomplete read that must not decide. Its saved summary reports all 21 checks passed. Color-control characters and padding are omitted above. These were synthetic requests; the outputs were read for this exhibit, not rerun.

## E2 — Hardening is a sequence, not a blanket guarantee

The July 27 review records **APPROVED-WITH-NOTES** after findings concerning trusted execution location, a shell fallback and evidence handling were addressed. The saved old-code failure log stops at a missing function; it alone would not demonstrate the original exploit behavior.

The September 23 red-team record includes these test outcomes in synthetic fixtures:

| Case | Recorded outcome |
|---|---|
| Unsigned approval record plus a replayed reply | Held; action not executed |
| A reply containing a rejection | Rejected rather than accepted as a substring approval |
| An approval-shaped question | Not approved |
| Forged lifecycle text | Did not unlock the guarded production operation |

These outcomes support specific controls and regressions. They do not certify every execution path, permission boundary or future version. No exploit commands or private infrastructure details are reproduced here.

## E3 — The fair test kept the losing result

**October 20, 2026.** Frozen-code comparison on 26 new deer/elk sides from 18 cards. Two antelope sides are outside the decision population. The reported clustered interval resamples by card.

| Measure | App | Candidate |
|---|---:|---:|
| Independently placed tine tips within half an inch | 21/41 (51.2%) | 26/41 (63.4%) |
| Median stray proposals per side | 0 | 1 |
| All tips within half an inch, including anchored taps | 100/122 (82.0%) | 94/122 (77.0%) |

Pre-stated rule: at least 10 percentage points better on independent tips **and** no increase in median strays.

Recorded result: +12.2 percentage points, 95% interval **-3.2 to +29.3**; median strays increased. **Rule not met; diagnostic/inconclusive; candidate not adopted.**

The report identified 55 of 96 tine taps as exact copies of an app proposal. This is why its independent-tap comparison is shown separately. Small samples, shared rack sides and anchoring limit the interpretation.

## E4 — A small apparent gain stayed diagnostic

**October 20, 2026.** Learned-lengths study on 62 carded elk, using leave-one-antler-out evaluation.

| Principal comparison | Within five inches |
|---|---:|
| Baseline rule on human labels | 26/62 |
| Learned principal arm on human labels | 28/62 |

The source reports a 95% interval for the difference of approximately -11 to +18 percentage points. The source labels the result diagnostic/inconclusive, with its pre-stated under-50% reading rule met. The reference labels were made with tape measurements visible, and scale calibration had used 11 of the 62 examples. These results do not establish a production accuracy gain.

## E5 — Current workflow, with its remaining breaks

October records document Friday planning, a new-scan fair test, an authorized Saturday work ladder, budget thresholds and separate GPU/promotion gates. The label records report 70 approved pilot elk labels.

The same record preserves the October 24 missed scheduler start and catch-up problem. A watchdog was built but not installed when checked. A later October 27 entry records that Chuck stopped the retrain during rendering; its manifest is aborted and no model was produced. The same-day audit reported training/test overlap and missing retained renders in the earlier lineage. Chuck put training on hold and ordered a reset from public base weights, with every training image retained. Earlier training is treated as unproven. These audit findings are attributed to the saved record, not independently re-audited here.

The source records themselves need care: the living state document contains newer dated entries beneath an older top-level stamp. The exhibit uses the dated entries and progress record rather than assuming the header proves currency.

## E6 — Cost estimate arithmetic

The user-supplied October 27 report gives rounded monthly Claude API-equivalent estimates:

| Period | Estimate |
|---|---:|
| June | $182 |
| July | $3,011 |
| August | $6,405 |
| September | $12,633 |
| October 1-27 | $4,071 |
| **Sum** | **$26,302** |

The stated subscription is $200/month. September's equivalent estimate divided by $200 is 63.165; October-to-date divided by $200 is 20.355. These ratios compare an estimated list-price counterfactual with a subscription fee; they are not measured productivity multipliers or realized savings. Codex is separately reported.

The printed commands do not pin a tool version or calculation mode. Current ccusage documentation distinguishes default mixed stored/calculated cost handling from explicit repricing. The arithmetic above was checked; the source logs, pricing categories and model-rate assignments were not independently reconstructed.

## Release boundary

The exhibit and these extracts were prepared locally. Chuck McFarland authorized this public release. The private manifest and internal recommendation report are excluded.