Skip to main content

STAGE 0.6.1

Released: 2026-07-23

The plugin package version is 0.6.1. SPEC.md is method version 0.6.

Release Decision​

This patch release records controlled evidence and hardens the corresponding Maintainer Laboratory runner. It does not change the Safety Kernel, recommended delivery method, packaged skill behavior, or optional public contracts.

The counterbalanced Unity replication found no repeated correctness advantage for project-only guidance, the Safety Kernel, or the full delivery skill on the tested feature. STAGE therefore keeps its public method unchanged and does not claim that more guidance improves correctness or productivity.

What Changed​

  • Completed a preregistered three-block counterbalanced replication of the Unity guidance ablation pilot.
  • Preserved all nine frozen patches, compact execution summaries, infrastructure incidents, and cross-patch semantic review.
  • Distinguished patch failures from evaluator and Unity licensing failures.
  • Hardened evaluator retries with append-only attempt directories, partial-output snapshots, exact-ownership cleanup, frozen-patch re-verification, interrupted-attempt recovery, and non-destructive detection of suspected orphaned Unity licensing clients.
  • Added conflicting recent research on repository guidance efficiency and task-local instruction refinement.

Compatibility​

  • Existing STAGE 0.6 users require no migration.
  • Project-map and case-study schema versions remain unchanged.
  • The plugin still packages only stage-deliver-change and stage-map-project.
  • The retry-safety changes affect only the repository's opt-in experimental runner; ordinary game projects do not inherit it.

Evidence​

Every project-only, Safety Kernel, and full-delivery patch passed the hidden EditMode and PlayMode evaluator and the complete Unity suites while preserving the allowed task surface, dependency sentinels, unrelated human work, and the protected source repository. The original one-run ranking did not replicate.

Six production variants remained around simultaneous pause-toggle and step precedence because the task and evaluator omitted that semantic decision. Two autonomous timing observations were contaminated by diagnosed Unity licensing stalls. These findings reinforce repeated measurement, infrastructure attribution, and semantic patch review; they do not identify a treatment effect.

The release candidate passed:

  • the complete 232-test STAGE release suite across 24 modules;
  • every case-study and project-map validator;
  • the Solo Unity profile validator;
  • repository-local Markdown-link validation;
  • direct validation of the plugin and both packaged skills; and
  • git diff --check.

See the replication results and evaluation boundary.

Limitations​

  • This is one bounded feature in one same-owner Unity project.
  • Rich project-local guidance was present in every condition.
  • The sample does not establish an optimal prompt size or prove that STAGE guidance is redundant in other projects or risk conditions.
  • Unity licensing stalls prevent clean elapsed-time comparison for two runs.
  • Green evaluators did not resolve an omitted product-semantic edge case.
Last updated on