03b_InfraFabric-Regulatory-Spine-20260201-1514UTC.md
HTML wrapper for HTML-only fetchers; content is shown verbatim.
Raw file
Links (extracted)
Clickable list for reviewers; avoids copy/paste line-wrap issues.
# InfraFabric — Regulatory Spine (03b addendum)
Document 03b of 04 (Regulatory actor model + doctrine; ~5–10 min)
Reading order: 01 (evidence) → 02 (system map) → 03b (you are here) → 03 (uncut dossier, unchanged)
Danny Stocker
Anthropic Fellowship Submission – January 2026
Track: AI Safety / Safety Engineering
(Safety / Security – Runtime Governance & Verifiable Oversight)
**Project homepage**
https://infrafabric.io
**LLM review access (task board, session ledger, registry)**
https://infrafabric.io/llm/
Generated (UTC): 2026-02-01T15:14:51Z
Black/white:
- This document is a **lens**, not a compliance guarantee.
- Verified claims are limited to publicly retrievable bytes and deterministic behaviors (see Verify).
- Receipts prove byte integrity only; not correctness, safety, compliance, or intent.
*If you read the dossier without the lens, you will treat narrative as evidence and evidence as narrative.*
---
## 1) Actor model (subjects / operators / payers / frontier labs)
Regulation is about incentives, not paperwork.
- **Subject:** the human impacted by a decision, inference, or data flow.
- **Operator:** the party running the system (access control, retention, incident response).
- **Payer:** the party paying for outcomes (often pressures for speed/coverage).
- **Frontier lab / upstream vendor:** the model/tool supplier (updates change capabilities and risk).
InfraFabric’s thesis: build systems where the audit story does not depend on privileged access or trust in the operator.
---
## 2) What auditors actually want (short list)
1) **Reconstruction:** What happened, when, and to which bytes?
2) **Containment:** What is the blast radius when a tool or model fails?
3) **Integrity:** Can we detect tampering in exports and reports?
4) **Separation:** Can we review evidence without tenant secrets?
5) **Drift:** Can we see upstream model/tool changes as recorded events?
InfraFabric patterns map to this:
- `/llm` (public review surface; GET-only)
- `if.trace` (sha256 receipts; PQ-ready when present)
- `if.bus` (envelopes: correlation ids + payload hashes + DLQs)
- `if.api` (adapter boundary: inspectable side effects)
- `if.context` (provenance: what bytes were read; roadmap to span addressing)
- `if.skills` (skills/tool affordances published as bytes)
---
## 3) How to read the uncut dossier (03)
Treat the uncut dossier as:
- motivation and narrative intent,
- “why this exists” context.
Do **not** treat it as evidence. Evidence lives in:
- public surfaces (`/llm`, receipts),
- contracts/specs,
- verify commands.
Rule: the uncut dossier bytes are included **unchanged** as a frozen historical artifact.
---
## 4) Verify (public, no login)
```bash
curl -fsS https://infrafabric.io/llm/conformance.json.txt | python3 -m json.tool > /dev/null
curl -fsS https://infrafabric.io/llm/if.registry.json.txt | python3 -m json.tool > /dev/null
curl -fsSI https://infrafabric.io/llm/entrypoint.json.txt | head -n 5
```
Doctrine source (repo-local):
- `docs/239-regulatory-design-patterns-infrafabric.md`