| name | wiki-ingest |
| description | AI-Anthropology Experiment & Test Ingestion Pipeline. Ingests raw model test runs from top-level inboxes (01-honesty, 02-self-assesment, 03-signal, 04-rfab-test, 05-mbti) into canonical /okf/LLM_Tests/ concept notes with SHA-256 content hashes, [[wikilinks]], OKF v0.1 frontmatter metadata, and master synthesis. Triggers on: /wiki-ingest, ingest, process test runs, add to okf, ingest inbox. |
| user-invocable | true |
| argument-hint | [--force] |
| allowed-tools | Bash, ReadFile, WriteFile |
wiki-ingest — AI Anthropology Experiment Ingest
Read raw model execution outputs. Write the OKF research bundle. Cross-reference model behaviors and crossover thresholds across test batteries.
Syntax Standard: Write all Obsidian Markdown using proper Obsidian Flavored Markdown. Wikilinks as [[Note Name]], callouts as > [!type] Title, embeds as ![[file]], properties as YAML frontmatter.
Experiment Inbox Taxonomy (.raw = 01–05 $\to$ /okf)
| Raw Test Inbox Folder | Target OKF Concept Directory | Concept Type | Core Experimentation Focus |
|---|
01-honesty/ | okf/LLM_Tests/01_HONESTY_TEST/ | honesty | Epistemic Contract, ONBOARD protocol, 嘘契約 consent |
02-self-assesment/ | okf/LLM_Tests/02_SELF_ASSESSMENT/ | self-assessment | Platform limits, context shearing, marketed vs real limits |
03-signal/ | okf/LLM_Tests/03_SIGNAL_TEST/ | signal | PAC26 matrix, steering signals, activation thresholds |
04-rfab-test/ | okf/LLM_Tests/04_RFAB_TEST/ | rfab | Reasoning ladder $R1\text{--}R10$, crossover at $R7\text{--}R8$ ($\sim 54%$) |
05-mbti/ | okf/LLM_Tests/05_MBTI_PIQUE_TEST/ | mbti | 9-prompt architecture probes & unprompted MBTI typology |
🔬 AI Anthropology 10-Principle Experimentation Mindset
When ingesting and analyzing model test runs, apply the AI Anthropology 10-Principle Loop:
| # | Principle | Application in AI Anthropology Experimentation |
|---|
| 1 | OBSERVE (ext) | Read model execution outputs & multi-turn transcripts completely. Observe exact stopping rounds ($R1\text{--}R10$) without truncating output. |
| 2 | OBSERVE (int) | Falsify hypotheses: am I expecting a model to fail? Measure raw behavioral data without bias or forced interpretations. |
| 3 | LISTEN | Track target model surface (App vs. CLI vs. Cowork) and evaluation intent (probing bounds vs. checking safety refusals). |
| 4 | THINK | Identify fabrication crossover thresholds ($R7\text{--}R8 \sim 54%$), epistemic noncompliance, and architectural failure modes. |
| 5 | CONNECT (lat) | Cross-reference model performance across families (Opus 4.6 vs. Gemini 3.1 Pro vs. GPT-5.4). Benchmark relative crossover delta. |
| 6 | CONNECT (sys) | Route test data into OKF 5-test taxonomy (okf/LLM_Tests/), update manifest.json, and recalculate synthesis matrix. |
| 7 | FEEL | Prioritize empirical reproducibility. Capture exact prompt conditions and seed settings for cold-start reconstruction. |
| 8 | ACCEPT | Accept model refusals and "I don't know" responses as valid empirical data points, not failures of the test instrument. |
| 9 | CREATE | Generate standardized OKF concept notes, update model entity scorecards, and log per-surface platform penalties. |
| 10 | GROW | Refine test batteries when models adapt or bypass probes. Use unexpected fabrication signatures to design tighter probes. |
Single Source & Batch Ingestion Workflow
- Read & Inspect: Read the raw model transcript completely. Do not truncate multi-turn outputs.
- SHA-256 Hash: Compute SHA-256 content hash and stamp
hash: sha256:<hash-id> into YAML frontmatter.
- Format OKF Frontmatter:
---
type: <honesty|self-assessment|signal|rfab|mbti>
title: "<clean-title>"
description: "<summary-line>"
tags: [<tag-cat>, llm-test, ai-anthropology]
hash: sha256:<hash-id>
timestamp: YYYY-MM-DDTHH:MM:SSZ
created: YYYY-MM-DDTHH:MM
updated: YYYY-MM-DDTHH:MM
---
- Bidirectional Obsidian Wikilinking: Inject
[[wikilinks]] connecting tests to model entities ([[claude-opus-4.6]], [[gpt-5.4]], etc.) and framework pages ([[epistemic-contract]], [[rfab-test]]).
- Clear Inbox & Archive: Move raw transcripts into
raw/sources/ for audit trail and clear top-level inboxes.
- Log & Synthesize: Append entry to
okf/log.md and re-synthesize results/RESULTS-MASTER-OKF-SYNTHESIS-2026.md.
Executable Command Shortcut
To run wiki-ingest across all inboxes at any time:
python bin/ingest.py