CREATION-LOG.md

CREATION-LOG — track-record

Packaging fix — 2026-07-18

• Drafted a Claude Code manifest so claude plugin validate can discover the package. (Note: that manifest lives in the not-yet-published packaging workspace and is not part of this release — do not look for a .claude-plugin directory in what you installed; the release ships the root plugin.json as the package manifest.)
• Retained root plugin.json as a compatibility mirror for existing skill-package consumers. Both manifests carry identical metadata, and the operative skill entry remains root SKILL.md for both packaging paths.
• This is packaging-only; no behavior fixture or pressure test was run or newly marked PASS.

Phase 2 contract layer — 2026-07-18

Not published in this repo. The artefacts listed below were drafted during the 2026-07-18 contract-layer pass and deliberately kept internal. None of them ship with this skill, and you will not find them in a clone. They are recorded here for provenance only.

• Drafted an evidence-bearing agent-modpack component manifest, a schema-pinned structural runner, and a machine-readable ledger-entry schema with positive/negative fixtures. Unpublished — which is why item 5 of the v1.0.0 notes below (Zero executable code shipped) still holds for anything you actually receive.
• Drafted SK-TR-P-001 and SK-TR-N-001 as independently executable case specifications. Unpublished, and both remain not-run; static validation is not behavior evidence.

Convention: keep the forging history in the repo. For a calibration skill this is not decoration — the log below includes the design's own bets, written in the skill's own ledger format. If the discipline is any good, it should survive being applied to itself.

v1.0.0 (2026-07) — initial forge

Design inputs (public sources)

• claude-improve — signal taxonomy and the promotion ladder (memory → rule → prominent rule → enforced hook). Also a cautionary tale: its own docs shipped inconsistent promotion thresholds in two places, which is why this skill defines every numeric threshold exactly once (SKILL.md, threshold table) and references it everywhere else.
• claude-reflect — confidence-tiered actions (HIGH/MED/LOW bound to different permissions) and false-positive filters for user-message signals; fingerprint-and-count for "same pattern recurring".
• Tetlock & Gardner, Superforecasting (commandments) — triage to the Goldilocks zone; outside view before inside view; replace vague verbiage with numeric probabilities; balanced under/overreaction; don't treat the commandments as dogma.
• Calibration-training practice (Hubbard-style interval training) — a fixed fast-feedback probability tier lets a beginner accumulate calibration signal quickly with small samples.
• Stateless outcome-grader pattern — the settler sees only the artifact and the standard, never the author's explanations; unmatchable standards are marked unsettleable, not force-graded.
• The superpowers school of discipline-writing — rationalization tables with verbatim excuses, red-flag phrases, violation symptoms, and the meta-principle that violating the letter is violating the spirit; plus the convention of shipping a CREATION-LOG.

Design input (anonymized, internal)

A ledger prototype ran ~3 weeks and accumulated 30+ entries; almost none were settled, the archive stayed empty, and the few completed settlements were nearly all externally driven. This is the single most load-bearing fact in the design: it turned "settlement reminders" into forced settlement.

The timeout three-way mechanism (settle now / mark unsettleable / one explicit extension with a new observable trigger) has, as far as our search found, no external precedent — it is original to this skill and is explicitly flagged as unproven. See "Design bets" below: we eat our own cooking.

Key decisions

1. Learning signal = reality vs. own predictions, not user corrections — deliberate positioning; the README carries a three-row comparison against the correction-harvesting tools rather than pretending they don't exist.
2. Forced settlement over gentle reminders — direct consequence of the prototype evidence.
3. Independent settler, no self-grading — imported from the outcome-grader pattern.
4. Single-source thresholds — one table in SKILL.md; every other mention is a reference.
5. Zero executable code shipped — the optional host hook is documented as an example config only, clearly marked as not bundled. Honesty about capability over impressiveness.
6. Language split — SKILL.md is the operative discipline (Chinese-primary in this release); README is the public face (English). A future release may ship a full English SKILL.md translation.

Slash-phrase candidates (final pick pending maintainer review)

• log your bets, settle against reality ← currently used in README
• bet, settle, distill — on the record
• a predictions ledger with forced settlement

Open items before / at release

• Pressure-test baseline (discipline-TDD): run fresh-context agents through "skip the log / fudge the settlement / wait-just-a-bit" pressure scenarios, capture verbatim rationalizations, extend the rebuttal table until no new excuses appear. Planned, not yet run — listed here as unfinished rather than claimed, per the skill's own honesty rule.
• Fill <owner> placeholders in README (repo URL, marketplace lines) at publish time. — done (LucioLiu), pre-release review round.
• Settle the slash-phrase choice. — decided in ops review: candidate 1 (log your bets, settle against reality) stays as the H1 for pull; candidate 3 (a predictions ledger with forced settlement) moves to the GitHub About line and description opener for search; candidate 2 dropped.

Design bets (the ledger, applied to itself)

#	日期	判断/预测	依据	回填触发	置信与概率	实际结果	偏差提炼
1	2026-07-17	Forced timeout settlement will be the single most-mentioned feature in substantive user feedback	赌 "log-only, never settle" 之痛足够普遍，专治它的机制会最先被认出来	First 10 pieces of substantive user feedback collected after release	MED · 60%	【待回填】	预填两头风险：也可能成为被抱怨最多的点（过于死板）——正反都盯
2	2026-07-17	The rationalization table will need at least three new entries within the first months of real use	赌真实压力场景比扶手椅枚举更丰富	First post-release pressure-test round, or first external contribution touching the table	HIGH · 80%	【待回填】	若未命中，需区分「表已足够好」与「表根本没被用」——用元校准文件分辨
3	2026-07-17	The independent-settler requirement will be the most-skipped rule in solo use	赌单人使用场景里「换个上下文」的摩擦最大，最容易被自评自过替代	First calibration review of a real user's ledger (self-reported or contributed)	MED · 65%	【待回填】	若命中：按三级阶梯考虑为独立结算配更轻的机制（如结算 checklist 模板）而非只靠条文