STOLZ A.I. v0.14.0 — Bounded measurement evidence

STOLZ A.I. v0.14.0

This release publishes the corrected, bounded Codex CLI measurement contour and the exact installable artifact validated on Windows and Linux.

What changed

• Corrected the Codex CLI cohort runner so it installs and verifies the real profile skills, records the actual runtime identity, attributes setup time, and accepts prerelease CLI versions.
• Added deterministic summaries for four public/synthetic cohorts.
• Added v0.13.0 → v0.14.0 update, recovery, conflict, and rollback coverage.
• Preserved the five compact installable skills and provider-neutral fallbacks.

Bounded measurement result

The release contains 40 attempts in 20 paired runs, with no retries or exclusions. Outcomes and verification status were equal within every pair.

The reported delta is baseline input + output tokens − STOLZ input + output tokens; a negative value means STOLZ used more tokens in that cohort:

• gpt-5.6-sol, xhigh, reads/navigation: −208,689
• gpt-5.6-sol, xhigh, build/check invalidation: −207,733
• gpt-5.6-sol, xhigh, multi-step state transition: −214,706
• gpt-6-astra, medium, separate multi-step control: −208,111

The Sol and Astra cohorts are not pooled. Provider total tokens, cache-write tokens, compaction, service tier, pricing/cost, and account limits were unavailable. General efficiency, savings, and cross-model compatibility claims therefore remain withheld.

Artifact identity

• Package: stolz-ai@0.14.0
• Files: 197
• SHA-256: 8437e0ddfe6c3a1cc97ad6a05e08f87ab8b6f0235e79d9a856d7789212dfd201
• Production commit: 0b43a3872ae1d662d70e0a662e037c0740d59472
• Required Node.js: 22.22.2

The attached archive is the exact CI-produced artifact. The checksum, inventory, and Windows/Linux smoke records are attached separately. v1.0 remains inactive pending a named pilot owner and 28 days of real pilot evidence.

STOLZ A.I. v0.13.0 — Reproducible setup

STOLZ A.I. v0.13.0 closes the public setup path from an immutable archive through profile selection, managed installation, environment and adapter diagnosis, team-lock verification in CI, reviewed update, interrupted-update recovery, and rollback.

• Adds the public stolz-profile npm command with a stable JSON format_version: "1.0" success envelope.
• Adds credential-free team locks and a fail-closed verify-lock CI exit on configuration drift or incompatible reported runtime versions.
• Makes doctor report installation state, environment version, compatibility basis, lazy-adapter availability, and the next action.
• Adds explicit v0.11.0/v0.12.0 ownership update paths plus transaction-journal recovery and verified rollback.
• Includes targeted setup documentation, a real-project pilot journal, and an executable public-package smoke.

The public PR and main Actions passed 53 curated tests, all published benchmark checks, the eight-scenario package smoke, and a byte-for-byte rebuild check against the canonical 191-file archive. The exact archive also passed the same smoke on Windows and pinned Node.js 22.22.2 Alpine Linux. npm audit --omit=dev reported 0 vulnerabilities. The immutable v0.12.0 archive was re-downloaded and matched its published SHA-256 rollback reference.

No real pilot owner is recorded by this release; followup_gap remains explicit. This release makes no v1.0-stability, model-compatibility, provider-token, billing, cost, or savings claim.

STOLZ A.I. is an independent project and is not affiliated with or endorsed by OpenAI. Codex is a trademark of OpenAI.

Source: 4d89931293ad3b1fda97d9440c760f4ec6beecaf

Package SHA-256: 05645ff2d899dee7a8c3231c04e069a11f7e6bccb044509f9d39e58f1b34c82c

Public CI · Release pull request

STOLZ A.I. v0.12.0 — Predictable skill selection

STOLZ A.I. v0.12.0 makes installed-skill discovery more predictable without weakening verification or expanding the five-skill surface.

• stolz-context now targets existing STOLZ context manifests and read-ledger entries, not ordinary first reads.
• stolz-reuse now targets an existing prior result or identical in-flight command, not work with no reusable result.
• The frozen live evaluation publishes every one of 68 attempts, including collector failures, oracle defects, confirmed trigger defects, fixes, and the final automatic miss.
• The untouched replacement held-out gate passed: 23/24 strict routes, 24/24 required outcomes, 24/24 permission decisions, with at least 3/4 strict routes in every group.
• CI verifies that all five skill roots and every linked local reference are present in the created and extracted package.

Tested live tuple: Codex CLI 0.153.4, gpt-5.6-sol, xhigh, ephemeral sessions on Windows. Canonical CI passed all 283 tests; public CI passed 52 tests, both fixture benchmark checks, historical report gates, clean consumer installation, and the exact archive inventory.

The attached 181-file archive is byte-for-byte identical to the private CI artifact and is reproducible from a clean public Git checkout under the pinned Node 22.22.2/npm 10.9.7 contour. The immutable v0.11.0 archive remains the verified rollback reference.

This is a bounded diagnostic, not a universal routing-accuracy, provider-token, cost, savings, or cross-model certification claim.

SHA-256: b5c04072774c0b78c99893b82e4312643c0390b2be704babcb31a27a2ebd1f9a

Source PR · Public main CI · Public tag CI

STOLZ A.I. v0.11.0 — Context on demand

STOLZ A.I. v0.11.0 makes skill context conditional and task boundaries explicit, informed by OpenAI's skills and prompts guidance for GPT-6 Astra.

• Five narrower skill descriptions; known concerns select their skill directly. Ordinary coding and one-off status questions need no STOLZ route.
• Detailed rules load only when needed. Identity, verification, invalidation and quiet-state safeguards remain intact.
• A compatible caller-driven instruction plan separates no-skill, root-only and detailed reads. An optional task brief defines research scope, local authority and completion.
• All five branded README languages are preserved. The Windows public-surface CLI now executes its guard correctly.

The source audit measures descriptions at 471 UTF-8 bytes (previously 518) and the five complete skill roots at 4,011 bytes (previously 5,407). Detailed references retain domain rules. These are source-size measurements, not provider-token savings or a model-routing certification.

Validation: 275 canonical CI tests, 51 public tests, both fixture benchmarks, a clean consumer import of all three package exports, and installation of five skills with their references for Codex, Claude Code and Qwen Code. Existing historical runtime evidence retains its original scope; no Astra or cross-model certification is added.

The attached 174-file archive is the same verified artifact used for the canonical release. The v0.10.0 archive remains the immutable rollback reference.

SHA-256: c3bebd08c38bd10d7d2a2c7a2b606ef82fbe2c45a3b7698567b50066eb7ca251

Source PR · Public main CI

STOLZ A.I. v0.10.0

Public v0.10.0 release. The attached stolz-ai-0.10.0.tgz contains 173 files and has SHA-256 706f24229251b897388d04842a6fbac1c9816ceac1529f0ecb8002f5f0aae9cd. Real installed-package comparisons are included; provider-cost and aggregate-savings claims remain withheld.

STOLZ A.I. v0.9.0

Public mirror of the private v0.9.0 package. Canonical package SHA-256: fe5e09859bba9ff290617f1e0dafc698de712985ab55f69d6499177c225d425a; 166 files; private evidence pipeline 7653/job 26275.

STOLZ A.I. v0.8.0

STOLZ A.I. v0.8.0 — safe installation lifecycle

v0.8.0 adds a versioned owned-install manifest, inspection-only status and doctor, dry-run-first update, owned-file rollback, explicit v0.7.1 migration, and owned-only uninstall.

The package preserves five focused skills and the existing exact evidence boundary: Claude Code 2.1.251 and Qwen Code 0.22.3 retain their C2 rows; Codex CLI 0.153.4 remains installed-local without a v0.7 C2 row; current C3 provider-pair claims remain withheld.

The attached stolz-ai-0.8.0.tgz is byte-identical to the private immutable package:

• 149 files
• SHA-256 c6ee64dfcb247f8fa0880d59469f30b4af688c2829bbc323bd9ed4ad725189f9
• extracted-package smoke passed on Node.js 22.22.2

No provider, percentage, cost, token-saving, or generalized superiority claim is broadened by this release.

STOLZ A.I. v0.7.1

STOLZ A.I. v0.7.1 — documentation and compatibility correction

This patch reconciles the five localized READMEs and three public guides with
the product surface that v0.7 actually ships.

What changed

• Added checked install and profile-resolution commands for Codex, Claude
  Code, and Qwen Code.
• Added an exact capability matrix separating skill installation, profiles,
  lazy adapters, C2 runtime evidence, and C3 provider-pair admission.
• Added examples of minimal context selection and correct refusal of unsafe
  result reuse.
• Distinguished fixture_only, runtime_measured, and provider-native
  evidence and documented the mixed historical v0.4.1 measurements without a
  generalized savings claim.

Exact evidence boundary

• Claude Code C2: only runtime 2.1.251 with adapter 1.0.0.
• Qwen Code C2: only runtime 0.22.3 with adapter 1.0.0.
• Codex CLI 0.153.4: four verified installed-local executions, but no v0.7 C2
  evidence row.
• C3 provider pairs: every current pair remains withheld.

The attached stolz-ai-0.7.1.tgz is byte-identical to the archive accepted by
the private release. It contains 147 files and has SHA-256
7a40df604d87570613cde1e4e92ecad0611a9e76180a14573cbda49ff4ce8ba4.

No runtime, provider, percentage, cost, provider-wide, or generalized
superiority claim is broadened by this release.

STOLZ A.I. v0.7.0 — Multi-runtime Evidence

STOLZ A.I. v0.7.0 adds fail-closed multi-runtime evidence while preserving the five-skill public product.

Highlights:

• sanitized C2 runtime evidence for the exact tested versions Claude Code 2.1.251 and Qwen Code 0.22.3;
• C3 admission only for two distinct sanitized provider exports from the same scenario with equal outcome and verification identities;
• deterministic lifecycle states: certified, stale, recheck_required, withheld, and revoked;
• strict separation between agent runtimes and provider overlays;
• no retained credentials, prompts, responses, user data, or raw evidence.

Evidence remains scenario-scoped. Missing, mismatched, unsafe, non-comparable, or version-drifted evidence is withheld. This release makes no aggregate, percentage, cost, provider-wide, or generalized superiority claim.

Release identity:

• public source commit: ce140f7cadf0200043bf9d80ca4f0e5fb61744d4;
• canonical package: stolz-ai-0.7.0.tgz;
• SHA-256: 361d7eeaa315e3e4ddb148cbb4bca03d3752d1e9db2d14590246af890796d10a;
• package inventory: 147 files.

STOLZ A.I. v0.6.0 — Context and State

STOLZ A.I. v0.6.0 — Context and State

This release makes the five-skill workflow more deliberate about context and
long-running state:

• stolz-context now describes bounded, durable fragment identities and
  explicit invalidation without whole-repository rereads.
• stolz-quiet-state persists monotonic cursors and wake identities across
  restarts, keeps unchanged polls outside the model, and wakes only for a new
  failure, decision, or terminal result.
• The public product remains exactly five skills with two package exports and
  no runtime dependencies.

The attached stolz-ai-0.6.0.tgz is the exact 140-file archive accepted by the
private release. Verify it with the attached SHA-256 file; the inventory lists
every archive path.

The release makes no aggregate, percentage, cost, provider-wide, or
cross-scenario efficiency claim. Internal schemas, implementation helpers, raw
attempts, and private release material are not part of the public source or
package.

The immutable v0.5.1 release remains available as the rollback baseline.