STOW v0.4.1

STOW v0.4.1

STOW v0.4.1 operationalises
controlled technical writing,
contextual anti-synthetic prose discipline,
focus and action shaping,
protected content,
and public-documentation
progressive disclosure.

This release completes the post-v0.4.0 absorption and evidence-hardening campaign while preserving the compact ordinary path.

What changed

Anti-synthetic writing

• Expanded the cold descriptive-prose carrier with
  explicit description, rationale,
  applicability, countercontext,
  rewrite principle,
  and owner mapping for
  functionless conclusions, empty process narration,
  hollow evaluation, manufactured contrast and escalation,
  mechanical symmetry, blanket hedging,
  heading opacity, and lexical inflation.
• Preserved
  functional summaries, supported expressive voice,
  genuine correction, required templates,
  justified uncertainty, literal technical senses,
  and protected text.
• Kept generated-marker matching advisory and protected compound identifiers.
  STOW does not infer authorship or impose universal punctuation, vocabulary, density, or burstiness rules.

Focus and action shaping

• Limited action-first ordering to contexts where it helps the reader.
  Explanation, decision support, and safety conditions can precede the action.
• Defined ordered steps by task-complete actions without a fixed item cap.
  Exhaustive inventories and required decision material remain complete.
• Deferred secondary issues without dropping them,
  surfaced only material changed state,
  and prevented pre-send self-review from deleting legitimately discursive or complex output.
• Recorded explicit conflict outcomes for safety/decision context and complete required output.
  Contextual self-review remains G1 guidance, not a delivery gate or final-output custody mechanism.

Controlled language and public documentation

• Preserved the Issue 9-derived requirement accounting,
  bundled 2,198-record cold dictionary, sparse lookup,
  project terminology authority, contextual boundaries,
  and locked strict-conformance profile.
• Preserved the direct cold public-documentation route
  under technical-clarity,
  with canonical-owner composition rather than a second ruleset or an audit subsystem.
• Preserved same-item noun-cluster repair and all targeted follow-up evidence without rescoring frozen Trial 2.

Architecture

• 65 active canonical rules.
• Sixty-one contextual G1 semantic owners and four genuine G2 predicates.
• Ten ordinary always-on G1 owners and fifty-one cold or predicate-loaded G1 owners.
• Ten advisory signals and fifty-seven behaviorally qualified G1 owners.
• Ordinary kernel: 1,084 tokens under the repository's exact tokenizer and 1,488 tokens under its conservative character proxy.
• No new profile, route, runtime helper, advisory detector, model turn, tool call, or ordinary cold-reference read was added by the deep-absorption work.

Behavioral evidence

Frozen comprehensive Trial 2 remains unchanged:

FROZEN TRIAL 2

• STOW, Numbered rules:
  49 PASS / 4 FAIL / 0 NOT_SCORED = 49/53 PASS

• STOW, Complete evaluation:
  80 PASS / 4 FAIL / 2 NOT_SCORED = 80/86 PASS

• name-only, Complete evaluation:
  78 PASS / 6 FAIL / 2 NOT_SCORED = 78/86 PASS.

Targeted later evidence covers all four former STOW failure distinctions.
This does not rescore Trial 2 as 84/86.

CURRENT EVIDENCE OVERLAY
Numbered rules: 53 supported / 0 failed / 0 NOT_SCORED = 53/53 supported
Complete evaluation: 84 supported / 0 failed / 2 NOT_SCORED = 84/86 supported

The current overlay is not a Trial-2 rescore.

The public-documentation V5 suite passed seven pathology cases and seven already-good controls.
The deep-absorption prose suite passed six pathology cases and six legitimate controls;
the action suite passed eight pathology cases and eight controls.

Every declared literal passed mechanical custody, and two blinded Luna Max reviewers per deep-absorption suite accepted every case
without retrying or tuning the candidate.

Known boundaries

• STOW-PRO-005 remains an open contextual limitation.
• STOW-WRD-001 remains an external project-authority boundary.
• STOW-WRD-003 and STOW-VRB-005 remain contextual deferrals.
• Strict controlled-language conformance stays locked.
• Universal automatic activation, cross-host behavioral equivalence,
  cross-model proof, rendered-consumer equivalence, and G3 final-output custody
  are not established.
• STOW does not claim universal output superiority or AI-authorship detection.

Release verification

• Final main commit:
  71f0955595cd04ff119a632e4faae55cc0c4b50b.
• Final tree:
  15156845fdcdb6231f4fe54c019ea441d2b635b8.
• Full suite: 1,200 passed / 39 skipped.
• All five generators: current.
• Strong local provenance gate: passed.
• Controlled-language crosswalk, dictionary,
  project terminology, public claims, self-dogfood,
  package, install, and reproducible-build checks: passed.
• Fresh Luna Max exact-tree release review:
  REVIEW_PASS after preserving and repairing one false-closure test gap.
• STOW.skill SHA-256:
  24f1d5619ef67d09ab4052f8dc087f1e513420669122f18b39089e561dd42b92.
• Artifact size: 690,419 bytes; 78 members.

STOW v0.4.0

STOW v0.4.0

STOW v0.4.0 completes the instrumentality reconciliation: it keeps a small ordinary-writing path, loads specialist controlled-writing material only when applicable, and reserves deterministic claims for properties that are actually closed and checked.

What ships

• 65 active canonical rules: 61 contextual G1 semantic owners and four genuine G2 predicates.
• Sixteen ordinary always-on G1 owners and 45 cold or predicate-loaded G1 owners.
• Ten advisory signals, kept distinct from rule-level semantic compliance.
• A bundled 2,198-record controlled dictionary with sparse retrieval.
• A project terminology-authority overlay for approved technical nouns, technical verbs, abbreviations, and canonical terms.
• Protected-region handling and deterministic validators for closed output contracts.
• Fifty-seven behaviorally qualified G1 owners.

Measured operationalisation

In the frozen crosswalk-derived comparison, name-only controlled-language conditioning passed 78 of 86 scored requirements; STOW passed 80 of 86. The result supports a bounded claim: STOW operationalised more of the tested requirement surface by making applicability, controlled-dictionary records, project terminology authority, protected regions, contextual boundaries, and mechanical checks explicit and inspectable. It does not prove universal output superiority or cross-model equivalence.

The comparison exposed operation- and force-preservation damage in the candidate. The cause was narrowed, the guidance was repaired, and a targeted Luna Max regression passed. The complete four-arm comparison was not rerun after that repair.

Runtime instrumentality

The audit also found and removed a genuinely overengineered execution path involving reference walks, runtime discovery, --help probes, temporary candidates, advisory lint calls, and repeated model turns. Normal installed-skill measurements after that repair produced these STOW/name-only cumulative logical-input ratios:

• ordinary cold path: 3.2044x
• technical cold path: 3.1848x
• controlled cold path: 5.7776x
• two-turn workflow: 1.5651x
• second-turn steady state: 1.0651x

These are logical model-input ratios, not billing multipliers. Cached and uncached input were accounted for separately; the host exposed no authoritative applicable cached-input price, so monetary input cost remains NOT_DERIVED. The historical approximately 29x Trial 2 figure remains evaluation-and-custody evidence and is not presented as normal STOW runtime cost.

Known boundaries

• STOW-PRO-005 remains an accepted open contextual limitation.
• STOW-WRD-001 remains an external-authority boundary.
• STOW-WRD-003 and STOW-VRB-005 remain explicit contextual deferrals.
• The strict controlled-language profile remains locked.
• Universal automatic activation is not proved.
• Cross-host behavioral equivalence is not proved.
• G3 final-output custody is not shipped.

Release verification

• Final merge commit: f9f132561df2c283331bbe09e357605206a06421
• Final tree: 19e18a407d4f011e0a70b22164c4f7212818662d
• Full test suite: 1,195 passed.
• All five generators: current.
• Strong local provenance/leak gate: passed over 196 files and 40 private patterns.
• Controlled-language crosswalk, dictionary, package, install, and runtime checks: passed.
• Two independent package builds: byte-identical.
• STOW.skill SHA-256: 5aee2140331a0fad9bc21422d860594d4af78d6b1755ad574603071386491751
• Artifact size: 678,821 bytes; 77 members.

STOW v0.3.6: a writing-discipline skill that governs every LLM response

What STOW is

STOW is a writing-discipline skill for Agents.

When it is active, it governs every user-facing response.

It obeys the exact output contract.

It keeps supplied literals byte-for-byte exact.

It validates structured regions before delivery.

It holds terminology steady.

It shapes clear prose and adds no fabricated detail.

It ships as a filesystem skill with a compact kernel and routed references.

A turn loads only the material its request needs.

What is in this release

• A kernel (SKILL.md) that sets the precedence bands, the region classification, the always-on rules, and the activation map.
• The primary rule set in twenty grouped corpus modules. Each rule has a stable ## STOW-XXX-NNN heading anchor. The README carries the full catalog.
• Four writing profiles, a machine-readable conflict registry, and a routing map checked against the kernel.
• Meta-code schemas and templates for handoffs, plans, audits, runbooks, state records, task packets, and event streams.
• A structured-output validator, a prose linter, a profile resolver, and a rule lookup helper.

What changed in 0.3.6

This release changes version metadata only.

No functional file changed from 0.3.5. The packaged payload stays byte-identical.

It promotes the exact package that passed the deployment campaign.

That campaign verified the installed bytes.

It ran an installed dogfood with no package fault. It proved every primary rule usable through the skill.

Each rule was engaged, retrieved through the bounded lookup, and applied to a planted violation.

The skill identified and corrected each violation.

Every callable check fired on the violation and cleared on the correction.

Rule meanings, corpus wording, and profile and conflict semantics stay unchanged.

Provenance

• Commit: 173df2cbc64b653c24198cfa5d5aedd888e0ce0d
• Asset: STOW.skill
• Asset SHA-256: 04cd2ddd64177caaee71883f29bbbf4477ee8be73dbb50a1b57b4b12768912f4

Verify the asset before you install it:

shasum -a 256 STOW.skill

Install

STOW.skill is a ZIP whose single top-level directory is stow/.

Extract it into a host skills directory.

The skill then resolves at <skills-dir>/stow/SKILL.md.

The runtime helpers need two packages:

pip install ruamel.yaml jsonschema

This release note was written under the STOW controlled-technical-guided profile.