Work Atlas / Field Log

Field notes

A curated public record of discoveries, corrections, and methods changing under evidence.

Browse two or three compact notes at a time on desktop, one at a time on mobile. Open any note for the complete reasoning sequence and evidence.

02 Oct 2026Method Change
Request construction became part of the treatment
A request-fidelity audit moved authoritative task context out of agent-written summaries and into deterministic host code.
Agentic-Jev · request fidelityv2 live-qualified · fresh effectiveness cohort preregistered, not run
Read detailed note
02 Oct 2026Study Result
The first fresh manager study found zero observed lift — and a treatment flaw
Across 30 fresh paired SWE-Lancer manager tasks, control and treatment both scored 15/30; the request audit then found that 0/20 Jev calls met the frozen context-completeness threshold.
Evaluation · SWE-Lancer manager decisions30-pair v1 completed · sanitized public result bundle pending
Read detailed note

01 Oct 2026Boundary Change
Independent review cannot be self-gated by the decision it checks
A semantic decision skill worked in controlled qualification, then made zero calls across six real manager decisions.
Portfolio project · Agent-Workflow / JevMechanism prototype merged · effectiveness unmeasured
Read detailed note
01 Oct 2026Method Change
More context is not the same as cleaner evidence
Controlled review work showed that semantic judgments can become sharper without becoming more discriminating, and that agent-authored framing can move the result.
Semantic assurance · context discipline
Read detailed note

29 Sept 2026Boundary Change
A passing qualification that did not authorize the next run
The first full v2 qualification passed, then a real-scale representation review changed the model-facing schema before any real label was produced.
Read detailed note
27 Sept 2026Correction
A regression test that only correlated with the bug
Upstream review changed what the regression evidence could support.
External upstream · Codex-LBPR #2506 · under review
Read detailed note

Boundary Change
A runtime lock that did not lock the runtime
A pre-live review found that version labels alone did not bind every result-affecting runtime surface.
Read detailed note
Method Change
Logging existed; observability did not
A logging setting and access records did not establish that the important production behaviors were observable.
Read detailed note

Method Change
Four investigators, three corrections
Separate methods and permission to falsify prior conclusions mattered more than the number of investigators.
External upstream · Codex-LB
Read detailed note
Implementation Change
Dogfooding becomes product correction
Four narrow mismatches found in real use became bounded product corrections without broadening the historical claim.
Portfolio project · Agent-Workflow
Read detailed note