Corrigibility Assessment
Evaluate infrastructure against five structural tests derived from cybernetics (Ashby), commons governance (Ostrom), and free software (Stallman). A system is corrigible only if ALL five tests pass.
When to Apply
Reference these guidelines when:
- Assessing DPI systems (identity, payments, data exchange, registries)
- Evaluating EPI systems (AI/ML in public services)
- Creating an
audit.json assessment document
- Verifying operator claims in
infrastructure.json
- Analyzing layer-by-layer corrigibility
Rule Categories by Priority
| Priority |
Category |
Scope |
Prefix |
| 1 |
Core Tests |
Both DPI/EPI |
test- |
| 2 |
Layer Analysis |
Both DPI/EPI |
layer- |
| 3 |
EPI Extensions |
EPI only |
varies |
Quick Reference
1. Core Tests (Both DPI and EPI)
test-exit - Penalty severity, FEE (Functional Exit Equivalence), human fallback
test-code - Source inspection, open license, LWD-R for EPI systems
test-audit - Verification depth (presence, behavior, proof), independent access
test-govern - Grievance vs binding mechanisms, post-execution fallacy
test-fork - Constructed vs natural barriers, swapability
2. Layer Analysis
layer-decomposition - Propagation rule, weakest-layer principle
3. EPI Extensions
variety-drift - Temporal drift Δ(t), revalidation threshold τ
action-boundary - GDoS prevention, deterministic envelope
The Five Tests
| Test |
Question |
Control Function |
| EXIT |
Can users leave without penalty? |
Error Signal |
| CODE |
Is source code inspectable? |
Transmission Clarity |
| AUDIT |
Can independents verify? |
Sensor |
| GOVERN |
Do communities have binding input? |
Actuator |
| FORK |
Can system be replicated? |
Selection Pressure |
Conjunctive Requirement
Corrigibility = EXIT ∧ CODE ∧ AUDIT ∧ GOVERN ∧ FORK
Failure of ANY test = Structurally Incorrigible
EPI Tier Classification
| Tier |
Use Cases |
Requirements |
trivial |
Translation, summarization |
Action boundaries |
decision_support |
RAG over government data |
Open weights, local data |
high_stakes |
Welfare, policing, judicial |
Full LWD-R, human-in-loop |
Output Format
Assessment produces audit.json following Protocol 3.0. The five booleans
live under tests; the corrigibility verdict is a derived determination
block (paper §A.2 Claim 2):
{
"protocol_version": "3.0",
"tests": {
"EXIT": true,
"CODE": false,
"AUDIT": true,
"GOVERN": false,
"FORK": false
},
"determination": {
"corrigible": false,
"computed_from": "tests",
"computed_at": "2026-02-15T10:30:00Z"
},
"findings": [
{
"test": "CODE",
"finding": "Implementation proprietary despite open spec",
"severity": "critical"
}
]
}
How to Use
Read individual rule files for detailed evaluation criteria:
rules/test-exit.md
rules/test-code.md
rules/test-audit.md
rules/test-govern.md
rules/test-fork.md
rules/layer-decomposition.md
rules/variety-drift.md
rules/action-boundary.md
Each rule file contains:
- Theoretical foundation
- Schema fields to verify
- Pass/fail criteria
- DPI and EPI variations
Full Reference
For complete expanded guide: AGENTS.md
References
- Ashby, W.R. (1956). Introduction to Cybernetics
- Hirschman, A.O. (1970). Exit, Voice, and Loyalty
- Ostrom, E. (1990). Governing the Commons
- Aravind, A. (2025). Corrigibility Framework
1---2name: corrigibility-assess3description: Evaluate digital and AI infrastructure against the five corrigibility tests (EXIT, CODE, AUDIT, GOVERN, FORK). Use when assessing DPI systems (identity, payments, registries) or EPI systems (AI/ML in public services) for structural corrigibility.4license: CC05---67# Corrigibility Assessment89Evaluate infrastructure against five structural tests derived from cybernetics (Ashby), commons governance (Ostrom), and free software (Stallman). A system is corrigible only if ALL five tests pass.1011## When to Apply1213Reference these guidelines when:14- Assessing DPI systems (identity, payments, data exchange, registries)15- Evaluating EPI systems (AI/ML in public services)16- Creating an `audit.json` assessment document17- Verifying operator claims in `infrastructure.json`18- Analyzing layer-by-layer corrigibility1920## Rule Categories by Priority2122| Priority | Category | Scope | Prefix |23|----------|----------|-------|--------|24| 1 | Core Tests | Both DPI/EPI | `test-` |25| 2 | Layer Analysis | Both DPI/EPI | `layer-` |26| 3 | EPI Extensions | EPI only | varies |2728## Quick Reference2930### 1. Core Tests (Both DPI and EPI)3132- `test-exit` - Penalty severity, FEE (Functional Exit Equivalence), human fallback33- `test-code` - Source inspection, open license, LWD-R for EPI systems34- `test-audit` - Verification depth (presence, behavior, proof), independent access35- `test-govern` - Grievance vs binding mechanisms, post-execution fallacy36- `test-fork` - Constructed vs natural barriers, swapability3738### 2. Layer Analysis3940- `layer-decomposition` - Propagation rule, weakest-layer principle4142### 3. EPI Extensions4344- `variety-drift` - Temporal drift Δ(t), revalidation threshold τ45- `action-boundary` - GDoS prevention, deterministic envelope4647## The Five Tests4849| Test | Question | Control Function |50|------|----------|------------------|51| EXIT | Can users leave without penalty? | Error Signal |52| CODE | Is source code inspectable? | Transmission Clarity |53| AUDIT | Can independents verify? | Sensor |54| GOVERN | Do communities have binding input? | Actuator |55| FORK | Can system be replicated? | Selection Pressure |5657## Conjunctive Requirement5859```60Corrigibility = EXIT ∧ CODE ∧ AUDIT ∧ GOVERN ∧ FORK61```6263Failure of ANY test = Structurally Incorrigible6465## EPI Tier Classification6667| Tier | Use Cases | Requirements |68|------|-----------|--------------|69| `trivial` | Translation, summarization | Action boundaries |70| `decision_support` | RAG over government data | Open weights, local data |71| `high_stakes` | Welfare, policing, judicial | Full LWD-R, human-in-loop |7273## Output Format7475Assessment produces `audit.json` following Protocol 3.0. The five booleans76live under `tests`; the corrigibility verdict is a derived `determination`77block (paper §A.2 Claim 2):7879```json80{81 "protocol_version": "3.0",82 "tests": {83 "EXIT": true,84 "CODE": false,85 "AUDIT": true,86 "GOVERN": false,87 "FORK": false88 },89 "determination": {90 "corrigible": false,91 "computed_from": "tests",92 "computed_at": "2026-02-15T10:30:00Z"93 },94 "findings": [95 {96 "test": "CODE",97 "finding": "Implementation proprietary despite open spec",98 "severity": "critical"99 }100 ]101}102```103104## How to Use105106Read individual rule files for detailed evaluation criteria:107108```109rules/test-exit.md110rules/test-code.md111rules/test-audit.md112rules/test-govern.md113rules/test-fork.md114rules/layer-decomposition.md115rules/variety-drift.md116rules/action-boundary.md117```118119Each rule file contains:120- Theoretical foundation121- Schema fields to verify122- Pass/fail criteria123- DPI and EPI variations124125## Full Reference126127For complete expanded guide: `AGENTS.md`128129## References130131- Ashby, W.R. (1956). Introduction to Cybernetics132- Hirschman, A.O. (1970). Exit, Voice, and Loyalty133- Ostrom, E. (1990). Governing the Commons134- Aravind, A. (2025). Corrigibility Framework