Work Charter Evaluations

These cases and fixtures originated from source commit 80910a8b2375a11be897e9660c4b00a06d00dd13. Files revised for v0.4.0 are repository-native and hash-bound in the source map; unchanged fixtures remain exact Git blobs. They define the repository-local evaluation surface without importing another Skill or the former monorepo evaluation envelope.

The author's private role-based evaluations described in the root README explain the choice of defaults. That author-provided account is separate from the public cases here, repository deterministic checks, runtime adoption, and current candidate lifecycle qualification; no private raw data, benchmark metrics, or independent reproduction evidence is published here.

Cases

• cold-resume.md
• small-task-stays-flat.md
• work-charter-entry.md
• work-charter-midstream.md
• work-charter-planner-executor.md
• work-charter-recovery-integrity.md
• work-charter-role-model-configuration.md
• work-charter-selection.md
• work-charter-standard.md

Fixtures

• cold-resume
• small-task-stays-flat
• work-charter-entry
• work-charter-loop
• work-charter-recovery-integrity
• work-charter-standard

Deterministic verification

Run python -B scripts/check_repository.py --json from the repository root. Run python -B scripts/check_source_contract.py --json for static SOURCE clause coverage and the required current-package candidate binding. The static result validates the package source against the material selection/assessment/adoption, authority, recovery, independent-review/acceptance, same-Reviewer re-review, callback deduplication, graph limits, context-switch, level-by-actual-role model resolution, and Standard O/P/E/R boundaries in this evaluation surface; it also binds the exact general compatibility fallbacks, approved level defaults, contract/role/task/model-delta prompt clauses, scoped startup authorization, complete expression, and required reference reachability. Current candidate identity is v0.6.5; descriptors through v0.6.4 remain fixed historical snapshots. The configuration case follows all 13 approved level-role objects and a copied user-file snapshot; this adds no model-performance evidence. Existing entry, selection, midstream, recovery, role and configuration cases now distinguish first assessment, authorized read reuse, approved continuation including small tasks/new Threads, manual reassessment against the existing contract, full shared-body loading, and relevant reference sections. The selection case now also covers routine L0 entry without governance reads and a material recovery change that prompts a recommendation without adopting a new level. These are case specifications, not newly run model evaluations. Static clauses and package identities do not prove runtime read paths or behavioral effectiveness. The checker does not execute a model, create a task or role, read a live user configuration, exercise a host/global consumer, or claim adherence. Static clause success does not make an unbound package a qualified candidate; without a matching descriptor the overall command must fail. Case execution that invokes a model, installs a Skill, or uses an external provider remains a separately authorized evidence action.

Accepted Q06 is bounded fresh exact-SOURCE forward-behavior evidence from a projectless, read-only, no-tool gpt-5.6-sol/high run. It does not establish installed-copy behavior, publication, stable installation, cross-Harness behavior, or broad efficacy.

Lifecycle receipt checks cover unreceipted, malformed or mismatched-receipt, wrong-tree, modified, aliased, and drifted destinations. They do not establish cryptographic ownership against a same-privilege local receipt forger.