# v1.3.0 candidate checkpoint

This record binds the untagged branch candidate to the exact fifteen-skill
portfolio and the final5 Autoimprove receipt. It is not a publication record.

```text
VERSION: 1.3.0
STATUS: RELEASE_CANDIDATE
VERIFIED_DATE: 2026-09-01
SKILL_COUNT: 15
PORTFOLIO: EXACT_FIFTEEN_CANONICAL_NOBRAINER_SKILLS
AUTOIMPROVE_RESULT: PROMOTED_FINAL5
AUTOIMPROVE_BASELINE_SHA256: 574db22aa0f809febc9bd29324e40308c76013cfce35a8141e637e42f1143529
AUTOIMPROVE_CANDIDATE_SHA256: 223702bb56cd66448be4c53cc31b4f321f8c47fc736e18e3a9d2f67cc5f9233c
AUTOIMPROVE_HOLDOUT: PASS
AUTOIMPROVE_EVALUATOR: CALIBRATED
AUTOIMPROVE_RECEIPT: docs/evals/artifacts/v1.3.0-autoimprove-integrity-final5-receipt.md
DETERMINISTIC_TESTS_TOTAL: 102
DETERMINISTIC_TESTS_STATUS: PASS
SECRET_SCAN: PASS
PUBLIC_SYNC: FETCH_ONLY
AUTOMATED_PUBLIC_COMMIT_PUSH: BLOCKED
MERGE: NOT_PERFORMED
TAG: NOT_CREATED
DISTRIBUTION: NOT_PUBLISHED
ACCEPTANCE: PASS_LOCAL_PR_CANDIDATE
```

## What changed

- Ultra binds one portable goal and definition of done to the canonical plan,
  with host-native goal readback only after scope is frozen.
- Ordinary work gets a concise outcome-first Progress view; a durable ledger
  appears only when dependency, handoff, recovery, multiple-writer or
  consequential-effect risk justifies it.
- Build, Dispatcher and Sessions now expose minimum-solution, critical-path,
  context-receipt and delivery-acknowledgement boundaries without duplicating
  ownership.
- Autoimprove protects the evaluator from candidate writes, calibrates controls,
  binds immutable score receipts, handles stochastic noise and continues with
  Pareto-sized fresh experiments after a rejected candidate.
- The portfolio remains exactly fifteen skills. Caveman's useful
  evidence-budgeted brevity was adopted in the host-compatible workflow layer;
  no sixteenth skill or client fork was introduced.

## Behavioral evidence

The final5 experiment used a frozen contract, a fresh sealed holdout, three paired
repetitions, equal conditions and two blinded independent judges. Candidate
outputs were 255/255/255 words; baseline outputs were 248/253/252. The candidate
passed every hard gate and scored 8/8 in every valid judge repetition; the
baseline scored 8/8 in two of three repetitions from each judge. Earlier
rejected, invalid or infrastructure-only rounds were not reused as holdouts.

See the [final5 receipt](../evals/artifacts/v1.3.0-autoimprove-integrity-final5-receipt.md),
[behavioral evaluation](../evals/v1.3.0-autoimprove-integrity-2026-09-01.md),
[compatibility boundary](../COMPATIBILITY.md) and
[testing contract](../TESTING.md). The exact v1.2.1 commit remains the rollback
anchor recorded in [its release readback](v1.2.1.md).

## Runtime boundary

This candidate record does not upgrade client-specific compatibility by
inference. The Codex and Claude rows in [COMPATIBILITY.md](../COMPATIBILITY.md)
remain scoped to their own clean-session evidence. Merge, tag and distribution
require their own owner-gated readback after this candidate review.