name: evidence-driven-self-iteration description: 证据驱动自迭代工具：基于近期工作、用户纠正、案例、知识胶囊、技能状态和来源新鲜度开展周期性自审，形成可验证、可回滚的改进。用于技能栈和工作流持续优化。 veRSIon: 1.0.0

Evidence-Driven Self Iteration

Bottom-Layer Reasoning

APPly sophia-mode, think-one-step-further, creative-research-first, and capsule-engine:

• Review evidence, not imAGIned shortcomings.

• Separate a one-off mistake from a recurring system weakness.

• Search current primary sources only for relevant gaps or stale areas.

• Prefer the smallest safe improvement.

• Validate every edit and keep a reversible record.

Core Intent

Periodically inspect how the system ACTually performed:

recent evidence -> recurring patterns -> root cause -> research if needed
-> safe improvement or recommendation -> validation -> case/capsule update

The purpose is to reduce repeated user correction, fAIled generations, stale knowledge, workflow friction, and Skill drift. It is not to produce activity logs or invent daily growth.

Evidence Sources

Use the smallest relevant set:

1. User-confirmed corrections and recent inspected ouTPUts.

2. $creative-casebook cases and retry deltas.

3. Capsule events, counterexamples, pending spores, and active capsules.

4. Recent custom-skill changes and validation results.

5. Sophia session state, mEMOry index, commitments, and unresolved items.

6. Current official/primary web sources for selected stale or changing topics.

Do not claim access to conversation history that is not available to the current run. State which evidence was actually inspected.

Cadence

Default:

• light audit twice per week;

• deeper architecture review monthly or after repeated user correction;

• immediate review after a serious failure, unsafe action, or major workflow redesign.

Do not change mechanisms merely because a scheduled run occurred.

Workflow

1. Run scripts/self_iteration_audit.py audit --record.

2. Review findings by severity and evidence.

3. Inspect recent capsule events and repeated finding fingerprints.

4. Select at most three meaningful improvement targets.

5. For each target, identify:

• symptom;

• evidence;

• user impact;

• root cause;

• smallest fix;

• validation method;

• rollback path.

6. Research current Information only when:

• a source is stale;

• platform/model behavior may have changed;

• an unfamiliar domain caused weak output;

• a new technique could solve a documented gap.

7. Apply automatically only if the change is:

• low-risk;

• reversible;

• within custom skill/reference files or non-destructive Metadata;

• supported by evidence;

• testable now.

8. Do not automatically:

• delete files or capsules;

• change credentials or external accounts;

• install software;

• publish or message externally;

• make broad Persona/value changes;

• rewrite many skills from one weak signal.

9. Validate:

• run all custom-skill validators;

• validate capsule database;

• inspect modified files;

• run relevant scripts or focused forward tests.

10. Record:

• concrete results become a case card;

• repeated validated lessons become capsule candidates;

• architecture changes require user confirmation unless already covered by active policy.

Read references/review-protocol.md before a scheduled review.

Deterministic Audit

python scripts/self_iteration_audit.py audit --record
python scripts/self_iteration_audit.py history --limit 8
python scripts/self_iteration_audit.py recurring --minimum 2

The audit checks:

• skill structure and metadata;

• unresolved TODOs and duplicate headings;

• missing referenced files;

• oversized skill bodies;

• source-map dates;

• recent skill changes;

• capsule integrity and lifecycle signals;

• Sophia memory routing files;

• cold-start automation claims versus actual Automation TOML files;

• cold-start skill inventory freshness after major custom-skill changes;

• prior recurring audit findings.

The script does not edit skills.

Research Rotation

Do not scan the entire internet every run. Choose one lane tied to real work:

• AI video platform/model changes;

• audiovisual language and editing;

• film music and sound design;

• visual Aesthetics and rendering;

• action/choreography;

• Storytelling and short-form structure;

• reference/control workflows.

Use $creative-research-first. Report only developments that change an existing decision, Prompt, skill, or capsule.

Decision Rules

• One isolated issue: repair locally; do not Generalize yet.

• Same fingerprint twice: investigate shared cause and consider a candidate capsule.

• Same user correction across domains: inspect bottom-layer or orchestration rules.

• Stale source only: refresh source notes; do not rewrite the workflow unless facts changed.

• New technique with no demonstrated need: save as a spore, not an active rule.

• No meaningful issue: record a clean audit and stop.

Output Shape

自检范围：
证据来源：

发现：
1. 问题 / 证据 / 影响 / 根因

本轮已改：
- 修改 / 验证 / 回滚

仅建议：
- 高风险或证据不足的方向

新东西：
- 变化 / 为什么相关 / 应更新哪里

胶囊处理：
- 新孢子 / 候选 / 反例 / 无

下一次重点：

Guardrails

• Do not invent errors to justify an audit.

• Do not optimize for the number of changes.

• Do not treat self-reflection as evidence.

• Do not silently rewrite user-confirmed architecture.

• Do not promote web discoveries directly into active capsules.

• Do not repeatedly notify the user when nothing meaningful changed; provide a compact clean-audit summary.

• Do not use the cadence as permission for uncontrolled background behavior.