https://github.com/hummbl-dev/hummbl-transparency
HUMMBL AI Transparency Registry — structured, versioned, evidence-backed vendor prompt auditing, model behavior documentation, and system prompt change tracking. A governed replacement for ad-hoc prompt leak collections.
https://github.com/hummbl-dev/hummbl-transparency
Last synced: 30 days ago
JSON representation
HUMMBL AI Transparency Registry — structured, versioned, evidence-backed vendor prompt auditing, model behavior documentation, and system prompt change tracking. A governed replacement for ad-hoc prompt leak collections.
- Host: GitHub
- URL: https://github.com/hummbl-dev/hummbl-transparency
- Owner: hummbl-dev
- License: other
- Created: 2026-05-13T13:02:45.000Z (about 2 months ago)
- Default Branch: main
- Last Pushed: 2026-05-25T21:13:13.000Z (about 1 month ago)
- Last Synced: 2026-05-25T23:13:35.203Z (about 1 month ago)
- Language: Python
- Size: 14.6 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- Contributing: CONTRIBUTING.md
- License: LICENSE
Awesome Lists containing this project
README
# HUMMBL Transparency Registry
*Structured, versioned, evidence-backed AI vendor transparency tracking. Maintained by [HUMMBL Research Institute](https://hummbl.io).*
## What this is
A governed replacement for ad-hoc "leaked system prompt" collections. This registry:
1. **Tracks vendor system prompts** — with version history, change diffs, and provenance evidence
2. **Documents model behavior** — structured observations across safety, bias, refusal, and capability dimensions
3. **Records prompt changes** — when vendors update system prompts, we capture the delta with timestamps and evidence
4. **Provides analysis tools** — stdlib-only Python scripts for diffing, auditing, and comparing vendor transparency practices
## Why this exists
The AI industry's transparency practices are inconsistent. Vendors change system prompts without notice. "Leaked prompt" collections are unstructured, unversioned, and unverifiable. This registry makes transparency **governed**: every entry has provenance, every change has a timestamp, every claim has evidence.
## Structure
```
registry/
vendors/ # Per-vendor prompt archives with change history
anthropic/
models/ # Per-model behavior documentation
diffs/ # Prompt change diffs with timestamps
tools/
validate_registry.py # stdlib validation for JSON, Markdown links, and registry shape
audit_prompt.py # Compare current prompt against registry
score_transparency.py # Rate vendor transparency practices
evidence/
... # Source notes and capture references
```
## Validation
```bash
python tools/validate_registry.py
python tools/audit_prompt.py registry/vendors/anthropic/claude-opus-4.7-2026-04-16.json
python tools/score_transparency.py
```
The validator is intentionally standard-library only and currently checks:
- JSON parseability
- local Markdown link integrity
- future registry files live below a category directory
## Principles
- **Evidence-backed** — every entry links to source evidence
- **Version-controlled** — prompt changes tracked like code
- **Stdlib-only tooling** — zero dependencies, run anywhere
- **Governed** — follows HUMMBL append-only audit standards
- **Vendor-neutral** — rates transparency, not products
## Status
**v0.1 — BOOTSTRAP COMPLETE.** The registry includes an initial
Anthropic Claude Opus 4.7 transparency entry, model behavior documentation,
schema-aware validation, prompt auditing, transparency scoring, and CI.
## Related
- [hummbl-governance](https://github.com/hummbl-dev/hummbl-governance) — governance runtime (kill switch, circuit breaker, cost governor)
- [krineia](https://github.com/hummbl-dev/krineia) — append-only governance receipt chain
- [agent-governance-demo](https://github.com/hummbl-dev/agent-governance-demo) — runtime safety primitives
## License
Apache License 2.0.