schema string | source dict | destination dict | mirror dict | files list |
|---|---|---|---|---|
kaymyg-hf-mirror/v1 | {
"platform": "github",
"owner": "kaymyg",
"repository": "kaymyg",
"url": "https://github.com/kaymyg/kaymyg",
"commit": "84b16d07f9148f8e9073066791754d570962eec8"
} | {
"platform": "huggingface",
"namespace": "Sahek",
"repository": "kaymyg",
"type": "dataset"
} | {
"timestamp_utc": "2026-08-30T12:08:17.229130+00:00",
"delete_remote_files": false,
"file_count": 1
} | [
{
"path": "README.md",
"sha256": "ad40933cb063a2733413bb61e90574bb9264f666f755a60c2b3404f5af8dfe2f",
"size": 3824
}
] |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
kaymyg
Independent researcher working on verifiable execution — making systems produce evidence of what they did that holds up when the system itself is the thing you don't trust.
Two specs so far, each with a reference implementation and a conformance suite:
- EMES-V1 — a structural event stream for EVM block execution, with a validation gate that rejects malformed or partially suppressed streams. Hiding an action requires hiding a consistent set of events; the balance invariants make partial omission visible.
- AMES-V1 — the same idea generalised to AI agent runtimes. Call frames become processes, frame balance becomes process lineage. An agent that suppresses its own spawn event leaves an orphan behind.
Specs and conformance
| kaysentinel | EMES-V1 reference implementation. Go live tracer against go-ethereum's core/tracing, plus an eight-crate Rust runtime for SSZ encoding, account certificates and verification. |
| kaysentinel-fixtures | Conformance fixtures for the awkward cases — EIP-6780 selfdestruct semantics, EIP-1153 transient storage under rollback. |
Agent runtime safety
| ames | AMES-V1 reference implementation: append-only ledger, capability brokering, environment fingerprinting, kill system. |
| attest | Design stage. Signed attestation of AMES ledger state, extending RFC 9421 HTTP Message Signatures. Threat model first, code later. |
| lean-agent-governance | Pre-runtime prompt-injection scanning and PR blast-radius limits for coding agents. |
| multi-agent-mcl | Meta-cognitive calibration layer — risk scoring and attribution across cooperating agents. |
| hydra-cage | Five-plane containment architecture — independent roots of trust per plane, eight-stage breach-to-rebuild transition, ten-attack adversary suite (10/10 blocked). |
Prototypes and reference implementations
Built out from a spec or draft I supplied as a starting point, not independent research in the same sense as the work above — but real, tested, honestly-documented code, not vaporware.
| ubda-engine | Capability-based data access prototype — six-state trust ladder, signed short-lived capabilities, hardware-authority/software-broker separation. Mock authorizer, documented as such. |
| caio | Synthesis prototype pulling capability leasing, graduated containment, and multi-perspective divergence detection into one request path. 100% line coverage on the core module. |
Systems and protocols
| jay-squad-sec | Adaptive zero-trust endpoint agent in Rust. Zero dependencies on purpose — builds anywhere, no supply chain. Alert-only. |
| ethan | AEP-1: an immutable USDC treasury protocol on Base. No owner, no upgrade path, no admin key over funds. Anyone may propose; the contracts decide. |
| CIMF | Cosmological Information Management Framework, with an accompanying paper. |
How I work
Every repo builds and tests from a clean checkout in CI. Fixes ship with a test that would have failed before them. Where a limitation is a calibration decision rather than a bug, it gets a test that documents the ceiling instead of a quiet change to the threshold.
Datasets and demos also mirrored on Hugging Face.
- Downloads last month
- 11