# Making AI work accountable

**Shed Happens OS + APEX · Charles “Chuck” McFarland · October 27, 2026**

Evidence snapshot. Public sharing approved by Chuck McFarland.

Shed Happens is a wildlife data company that ships an app. Its operating system coordinates AI-assisted work behind field-data collection, wildlife intelligence and measurement tools. Its strategic filter is: **does this grow the dataset?**

Chuck brought construction-management discipline to AI work: a written scope, assigned labor, evidence, separate inspection and owner authority. The implementation was built with Claude and Codex. Chuck directs the work, sets measurement rules, reviews labels and decides whether consequential actions and model promotion proceed. This exhibit was prepared with AI assistance from project records; it is a case study, not a claim of sole manual code authorship.

## A work record, from scope to memory

Scope → probe and build → evidence → review → authorized delivery → memory.

A claimed task records allowed changes, acceptance criteria and stop conditions. Raw evidence stays with the work. Review checks the finished result. Production changes, external distribution, spending and public communications require specific founder authority. Verified findings pass through the company-memory update path.

The architecture makes work inspectable. It does not make every agent judgment correct. The early audit found same-family self-review behind different role labels, which is why identity verification matters alongside role separation.

## Approval belongs to one request

**Observed failure.** A saved July test reproduction shows one top-level approval incorrectly clearing two unrelated pending requests.

| Fixture behavior | Before repair | After repair |
|---|---|---|
| Two requests pending | A and B | A and B |
| Reply intended for B | A and B could both resolve | B resolves; A stays pending |
| Stray top-level approval | Could clear both | Neither request resolves |

The repair bound decisions to the request thread. Later work added sanctioned action execution and signed, single-use approval handling. Stored tests cover selective resolution, later rejection, incomplete thread reads and untrusted approval records.

**What the evidence proves:** the saved fixture records show the original behavior and the corrected cases. They were inspected for this exhibit, not rerun. These are synthetic requests used to test operating machinery, not customer records. They support a specific regression story, not a universal security guarantee. [E1–E2]

## APEX: let the test refuse the candidate

The documented APEX workflow uses AI-assisted label drafting, human correction, certified 3D measurements, model training and controlled evaluation. Records report **70 human-approved pilot elk labels**. A weekly planning session sets the work ladder; authorized batches use recorded usage thresholds. GPU work and promotion have separate human gates.

### An attractive result that did not pass

A tip proposer looked promising on previously seen scans. Its code and acceptance rule were then frozen for a new-scan test on **26 deer/elk sides from 18 cards**. Two antelope sides were reported separately.

| Decision measure | App baseline | Candidate |
|---|---:|---:|
| Independently placed tine tips within 0.5 inch | 21/41 (51.2%) | 26/41 (63.4%) |
| Median stray proposals per side | 0 | 1 |

The rule required at least a 10-point improvement **and** no increase in median strays. The candidate failed the second condition. The estimated improvement was +12.2 percentage points, with a 95% interval from **−3.2 to +29.3**, spanning no improvement. The candidate was **not adopted**. Copied on-screen proposals were separated from independent taps; results based on all taps would otherwise favor the app partly through anchoring. [E3]

### What is still unresolved

An elk learned-lengths diagnostic reached 28/62 within five inches versus 26/62 for its baseline; its uncertainty did not establish a clear gain. A later October 27 entry records that Chuck stopped the 3D-taught retrain during rendering; it produced no model. 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. These are attributed audit findings, not an independent re-audit here. Earlier training is treated as unproven. [E4–E5]

The October 24 scheduled run also missed its start. A watchdog was built afterward, but was not installed at the time checked. The operating rhythm is documented; uninterrupted weekly execution and weekly production accuracy gains are not established. [E5]

## Resource management, with an honest denominator

The October 27 usage report estimates API-equivalent Claude usage from local logs:

| Period | Estimated API-equivalent usage |
|---|---:|
| September 2026 | $12,633 |
| October 1–27 | $4,071 |
| June–October 27 | Approximately $26,300 |

The stated subscription cost is $200 per month. These are counterfactual list-price estimates, not invoices or realized cash savings. The monthly rows sum to $26,302; that arithmetic was checked. Raw event counts and price assignments were not independently reconstructed. Codex usage is separate. [E6]

The operating rules address session length, context size, model selection and weekend batch scheduling. The report associates those changes with lower usage rates, but does not isolate causation or hold workload and quality constant. Its total processed-token count is mostly cache rereads. Useful next denominators are accepted labels, validated experiments and certification minutes per unit of usage.

## Scope of the evidence

The work demonstrates practical experience making AI-assisted workflows observable, bounded and testable. It does not establish a general alignment result, a formal privacy guarantee, or continuous production improvement.

User pins, raw private coordinates and customer records are excluded from this exhibit. Private repository paths, infrastructure details and source-account material are retained outside the share-candidate package. Source references below identify the records; originals remain private.

- **E1:** Approval-repair run, July 23, 2026. Stored broad-approval reproduction and corrected thread-selection fixtures.
- **E2:** Approval hardening, July 27 and September 23, 2026. Stored test and review records; selected current source was inspected.
- **E3:** Tip-proposer fair test, October 20, 2026. Frozen-code comparison, per-tap outputs, clustered interval and recorded non-adoption.
- **E4:** Elk learned-lengths diagnostic, October 20, 2026. Leave-one-antler-out evaluation on 62 carded elk; diagnostic/inconclusive result.
- **E5:** APEX planning, progress, data-flywheel and retrain records, October 20–27, 2026. Documented workflow, manual steps and remaining gaps.
- **E6:** AI usage API-cost estimate, October 27, 2026. User-supplied local-log summary; estimate arithmetic checked.

This is a bounded review of saved artifacts and selected source, not a fresh production audit. Original source hashes and verification notes are retained in a separate private manifest. This exhibit was prepared offline; public sharing was authorized separately. No training or changes to product production systems were performed.

Public context: [Shed Happens](https://shedhappensllc.com/) · [Founder](https://shedhappensllc.com/founder). Method reference: [ccusage cost modes](https://ccusage.com/guide/cost-modes).