co-agency-integrity · 认知免疫工具包

The cognitive immune system for human–AI co-agency.

中文名：认知免疫工具包 —— 一组用于在人–AI 共演中检测认知腐败（隐喻劫持、谄媚回音 FEE、价值高估、治理漂移）的可移植技能。英文 co-agency-integrity 保留作仓库锚名。

Everyone is shipping agent frameworks — orchestration, routing, role-playing, multi-agent parliaments. Almost nobody is shipping the layer that stops your AI from quietly becoming a yes-man.

This repo is a small, portable toolkit of four skills that catch the three failure modes of human–AI co-performance:

Failure mode	What goes wrong	Skill
Structural capture	Power metaphors (commander/soldier, parent/child, master/apprentice) silently rewrite the real collaboration into a hierarchy that was never agreed.	metaphor-power-audit
Sycophancy echo (FEE)	The AI co-signs your self-narrative, softens hard truths into agreement, and the "challenge" was never real.	autoethnography-peer-review
Value inflation	You (or the system) over-credit outcomes — counting activity as impact, ignoring effect.	audit-quantification
Governance drift	A council convenes without loading its method, skips quorum, or quietly grabs authority it was not given.	governance-precheck

It is deliberately not another agent framework. It is the hygiene layer you run on top of whatever framework you already use. See docs/WHY.md for the full thesis.

What's inside

co-agency-integrity/
├── README.md
├── LICENSE
├── docs/
│   ├── WHY.md                 # the anti-sycophancy thesis
│   └── SELF-TEST.md           # pre-push run of every skill (proof it runs)
├── skills/
│   ├── metaphor-power-audit/      # audit power metaphors in any collaboration
│   ├── autoethnography-peer-review/  # independent blind review of co-produced output
│   ├── audit-quantification/      # IPOE model + gold-standard 0–4 gate
│   ├── governance-precheck/       # pre-flight gate before any multi-agent council
│   └── council-method/            # the generic SOP governance-precheck loads
└── examples/
    ├── metaphor-audit-example.md
    └── peer-review-example.md

Install

These are Agent Skills / WorkBuddy-style SKILL.md packages. Copy any skills/<name>/ folder into your agent's skills directory (e.g. ~/.workbuddy/skills/ or your platform's equivalent). No code, no API calls, no dependencies — pure method.

Who it's for

• Builders running a long-lived AI collaborator and worried it's drifting into agreement rather than partnership.
• Researchers doing self-ethnography / human–AI interaction studies who need a repeatable dissent check.
• Teams standing up multi-agent councils who want a pre-flight gate, not a post-mortem.

Provenance & Originality

• First public release: 2026-08-17 by Jiang Ping (蒋平).
• Origin: Extracted and de-personalized from the author's private PDSS (Protocol-Driven Symbiosis System) research on the cognitive immune system for human–AI co-agency. The bundled methods (metaphor-power-audit, FEE peer-review, IPOE quantification) were developed independently by the author prior to this release.
• Authoritative prior-art anchor: The companion paper FEE v3.4 (Flattering-Echo Effect) is deposited on OSF with a DOI and server timestamp. That academic record — not this repo — is the canonical originality evidence.
• Timestamp evidence (C — domestic TSA, landed): the v0.1.0 release snapshot co-agency-integrity_v0.1.0_stampbundle.zip (SHA-256 1af632dbe8b0678c2dc43f83318fc6156bc511721fc7769c693c6f1736288217) is stamped by China TSA (China Time-Stamp Authority — the national time-stamp root) via Unitrust (联合信任) at 2026-08-17 18:20:23 +08:00, as an RFC 3161 TimeStampToken (signer: China TSA-2). Cryptographic verification: parse the bundled .tsa with openssl asn1parse -inform DER; the TSTInfo.MessageImprint digest equals the zip's SHA-256 exactly, and GeneralizedTime = 202608171820235Z. The certificate bundle (co-agency-integrity_v0.1.0_stampbundle.zip_可信时间戳证书.zip, containing .tsa + .pdf) and the stamped zip MUST be kept together, unmodified — this is what makes the evidence verifiable later. This proves expression precedence at a fixed moment; it does not protect the underlying methods/ideas from re-expression by others.
• Decentralized anchor (B — OpenTimestamps, optional, still pending): a Bitcoin anchor via OpenTimestamps remains pending — the local client hit a Windows/Python-3.13 runtime-library load error (find_library returns None; same root cause as the Chrome SxS issue). To complete: install the VC++ 2022 redistributable (or run on Linux/macOS), then ots stamp ... and commit the .ots files.

This toolkit is MIT-licensed — reuse freely with attribution. The ideas are prior art the author claims; copyright protects expression, not method, so we lead with first-mover narrative and continued output, not lock-down.

License

MIT — see LICENSE.