Making AI work accountable.
A practical operating system for AI-assisted work, and an experiment workflow that lets evidence refuse a candidate.
Shed Happens is a wildlife data company that ships an app. The work behind it serves one question: does this grow the dataset?
Construction discipline, encoded.
Chuck brought the structure of a construction job to AI work: a written scope, assigned labor, proof, separate inspection and owner authority.
The implementation was built with Claude and Codex. Chuck directs the work, defines measurement rules, reviews labels and makes the consequential decisions.
A record the next session can inspect
Scoped work moves through probing, implementation, evidence, review, authorized delivery and a verified memory handoff. Each stage has a record and an authority boundary.
Production changes, external distribution, new spending and public or mass communications require specific founder approval.
An architecture is not proof of perfect enforcement. The historical audit found same-family self-review behind different role labels. Reviewer identity matters alongside role separation.
Approval belongs to one request.
A saved July test reproduction shows one top-level approval incorrectly clearing two unrelated pending requests.
The repair bound a reply to the request thread. Later work added sanctioned action execution and signed, single-use approval handling.
The display normalizes stored fixture outputs into request A/B labels. These were synthetic test requests. The original records were inspected for this exhibit, not rerun. [E1–E2]
Requests A and B are pending.
A reply is posted in B’s thread.
B resolves. A remains pending.
A stray top-level reply resolves neither request.
What the test can claim
The saved fixtures show selective resolution, later rejection and refusal to decide from an incomplete thread read. They support a specific regression story. They do not establish a universal security guarantee.
Let the test refuse the candidate.
The documented APEX workflow combines AI-drafted labels, human corrections, certified 3D measurements, model training and controlled evaluation.
Records report 70 human-approved pilot elk labels. Weekly planning sets the work ladder. Authorized batches use recorded usage thresholds; GPU work and model promotion have separate human gates.
This historical diagnostic illustrates evaluation, not validation of the current model. A tip proposer looked promising on seen scans. The code and decision rule were then frozen for a new-scan test: 26 deer/elk sides from 18 cards. Two antelope sides were reported separately. [E3]
| Decision measure | App | Candidate |
|---|---|---|
| Within 0.5 inch | 21/41 51.2% | 26/41 63.4% |
| Median strays / side | 0 | 1 |
The rule required at least a 10-point improvement and no increase in median strays. The candidate failed the second condition.
95% interval spans no improvement. The estimate is uncertain.
Why the independent-tap distinction matters
Some human taps were exact copies of the app’s visible proposals. Counting those as independent reference labels would partly reward the app for anchoring the human. The report separates independently placed taps and keeps the limitations explicit.
Why training was stopped
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 retrain during rendering: no model was produced. A same-day audit reported training/test overlap and missing retained renders in the earlier lineage. Chuck held training and ordered a reset from public base weights with every training image retained. The audit findings are attributed to the saved record, not independently re-audited here. Earlier training is treated as unproven. [E4–E5]
Where the weekly automation failed
The October 24 scheduled run missed its start. A watchdog was built afterward, but was not installed at the time checked. The documented weekly rhythm should not be described as uninterrupted execution or a guaranteed weekly accuracy improvement. [E5]
Measure the work. Then reduce the waste.
The operating rules address session length, context size, model selection and weekend batch scheduling. A local-log report tracks their resource cost.
The report associates those changes with lower usage rates. It does not isolate causation or hold workload and quality constant. Most processed tokens were cache rereads.
| Period | Estimate |
|---|---|
| September 2026 | $12,633 |
| October 1–27 | $4,071 |
| June–October 27 | ≈ $26,300 |
Stated subscription cost: $200/month. Codex usage is separate. These are estimates, not invoices or realized cash savings. Monthly-row arithmetic was checked; raw usage events and price assignments were not independently reconstructed. [E6]
A better efficiency denominator is accepted labels, validated experiments or certification minutes per unit of usage, with quality held constant.
Keep the limits with the evidence.
This case demonstrates practical experience making AI-assisted work observable, bounded and testable. It does not establish a general alignment result, a formal privacy guarantee or continuous production improvement.
Private user pins, coordinates, customer records, repository paths and infrastructure details are excluded. Originals and source hashes remain in a separate private manifest.
Prepared from saved records and selected source. Historical experiments were not rerun; model quality and production state were not revalidated. Prepared offline and approved for public sharing. No training or product production-system changes were performed.
Public context: Shed Happens · Founder
Method: ccusage cost modes
- E1
- July 23 approval repair. Stored broad-approval reproduction and corrected thread-selection fixtures.
- E2
- July 27 and September 23 hardening. Stored tests and review records; selected current source inspected.
- E3
- October 20 tip-proposer fair test. Frozen-code comparison, per-tap outputs, clustered interval and recorded non-adoption.
- E4
- October 20 learned-lengths study. Leave-one-antler-out evaluation on 62 carded elk; diagnostic/inconclusive.
- E5
- October 20–27 APEX planning, progress, data-flow and retrain records. Workflow, manual steps and gaps.
- E6
- October 27 local-log cost estimate. User-supplied summary; estimate arithmetic checked.