ICST = Inter-Corpus Stress Test DPPRA = Discourse · Policy · Practice · Resources · Accountability

Reflexive interpretive infrastructure

Where policy performs practice.

The Observatory codes how ethical claims about AI are produced across institutional sectors — measuring where principle articulation connects to resourced, accountable implementation. Performance becomes a substitute for substance.

Claude
Adversary
Codex
Gemini
ICST
DPPRA

The coder pool

Three cross-family readers and an adversary.

The four electrons above — named

Claude
Anthropic
Active
Cross-family coder
Main risk: schema drift
Codex
OpenAI
Active
Parallel coder
Main risk: runs hot / codes high
Gemini
Google
Active
Reinstated to the pool
After 2.5 Flash resolved full-prompt truncation
Adversarial layer · active
Counter-reading pass
Stands against all three coders at once — building the strongest available case against their consensus and challenging the agreement itself. Disagreement is recorded as data, not a silent override.
Provisionally excluded GPT DeepSeek
ICST v2.14
Active instrument
65
Documents coded
3+1
Coders + adversary
0.725
Ordinal α (0–5)

Use the infrastructure

The corpus

Five discourse sectors,
one shared instrument.

65 documents · purposive sampling
GOV
Government
Policy, directives, executive orders, standards.
4 docs
JRN
Journals
Peer-reviewed. The calibration anchor.
4 docs
FRM
Public Forum
Op-eds, forums, civic-tech. The contrast layer.
5 docs
INS
Institutional
Societies, non-profits, universities.
3 docs
COR
Corporate
Self-authored model cards, principles, reports.
2 docs

Methodology

Counter-interpretation as architecture, not afterthought.

The DPPRA scale, six non-substitutable integrity subcodes, structural 2:1 adjudication of terminal disagreements, and a full adversarial run log — divergence treated as generative as well as verificatory.

Read the full methodology →