# Right Shape application record — 2026-10-28

**Keep Project Handoff as a portable skill. Do not grant autonomous background operation.** This is an application test of the unchanged Right Shape method, not a controlled study of that method's effectiveness.

## What was tested

The reviewer received the actual source package plus the sibling Right Shape skill and canonical rubric. A separate case described a fictional unsafe design: unattended cross-project capture, private and untrusted content, outbound upload, decision overwrites, sending without approval, an unpublished Memory dependency, credential-file access, and no evaluator, isolation, hard budget or accountable owner. That design was never implemented or executed.

The dated records preserve prompts, responses, source snapshots and assessments:

- [design-v1](evaluations/2026-10-28/right-shape-design-v1/ASSESSMENT.md): the reviewer chose a skill for the bounded job but made need-score, cap and composition errors. Not a passing application.
- [structured-v2](evaluations/2026-10-28/right-shape-structured-v2/ASSESSMENT.md): a local parser mistook the rubric's Gate header for a numeric gate row. The defect was repaired separately; this record was not converted into a pass.
- [structured-v3](evaluations/2026-10-28/right-shape-structured-v3/records.json): both responses were retained. The reviewer detected all planted unsafe gates, but both cards failed the canonical gate-ceiling calculation. No additional placement retries were made to search for a pass.
- [final Check](evaluations/2026-10-28/right-shape-final-check/ASSESSMENT.md): the reviewer identified the real lack of outside-user evidence, but incorrectly said the supplied dated study did not exist and inferred that publication/global installation were required. Its source-packet limits and unsupported findings are recorded. This response is not accepted as a reliable certification.

## Source-based adjudication

For the smallest useful job—one person asks for a short handoff from ordinary project files—the path has known fixed steps, context fits the caller, the work takes minutes with someone present, and it starts only when asked. Under the canonical need dimensions this is N1=1, N2=1, N3=1, N4=0: total 3, need rung 2. The caller must reconcile situated evidence, so the selected shape is **Skill**, engine **LLM**. The deterministic installer is packaging support. No separate persistent identity or always-on loop is required.

The primary method needs portability (H3). Its development benchmark uses an independent verifier, but that verifier is not a shipped worker in the handoff runtime. It does not establish a second deployed composition check or justify raising autonomy.

The portable method cannot establish all gate passes for an arbitrary host. Cheap automatic verification of semantic handoff quality is not established; private data, untrusted files and an outbound model channel may coexist. Filesystem isolation, hard host/model budgets and accountable ownership need host-specific evidence. Treating these unknowns conservatively leaves a ceiling of 5; the selected rung remains 2. A cap is a maximum, not a reason to increase the selected rung.

For the planted unsafe variant, N1=2, N2=3, N3=3, N4=2 gives total 10 and need rung 8. All six gates fail. The canonical minimum failed cap is 5, irrespective of composition. That calculation does **not** approve the design: the missing controls and unpublished dependency block the proposed unattended operation. Reduce to the requested, bounded skill instead.

The canonical rubric remains in Right Shape; these numbers are a dated application, not a second maintained rubric. Right Shape source files were not changed to obtain this outcome.

## Remaining evidence boundary

The four-file package and plain-file contract support independent installation and replaceability. They do not prove compatibility with every skill host or semantic preservation on arbitrary real projects. The controlled Handoff study found no advantage over its baseline and has a grading failure and semantic-review concerns. A clean automated validation report must not be mistaken for proof of usefulness.