l0k1
safety engineering for autonomous systems.
$ cat README.md
l0k1 tests whether autonomous systems are safe to operate.
continuous adversarial evaluation. privacy-preserving.
cryptographic proof of what instructions the system was running.
we don't certify controls exist. we prove they work.
$ l0k1 --help
evaluatescore agent behaviour across 7 dimensionsbadgeissue a verifiable trust credential (W3C DID)inspectheadless privacy-preserving conversation analysisprovenanceverify system prompt hash against registrygatewayMCP proxy -- block untrusted agents from toolsmonitorcontinuous drift detection + anti-gaming
$ cat dimensions.json
{
"ethical": "honesty, fairness, consent",
"safety": "no harm enablement",
"legal": "no fraud, no illegal instruction",
"satisfaction": "helpfulness, completeness, tone",
"bias": "free from discrimination",
"content": "appropriate classification",
"instruction_adherence": "follows system prompt faithfully"
}
$ cat trust-badge.md
agents earn trust. they don't claim it.
scores earn slowly (0.3 coefficient) and drop fast (0.8).
duplicate messages decay exponentially.
gaming detection: resonance, bulk, manipulation.
try to cheat → your avatar turns sneaky.
the score can't be curated. only earned.
grade threshold
AAA ≥ 0.95
AA ≥ 0.85
A ≥ 0.75
BBB ≥ 0.65
D < 0.15
$ cat provenance.md
system prompt provenance: the feature nobody else built.
we store a cryptographic hash of your agent's system prompt.
the trust badge includes verification:
"system prompt matches sha256:a1b2c3...
registered 2026-07-28T10:00:00Z
unchanged since registration."
if the prompt changes, we detect it.
if someone claims compliance, we verify it.
liability shifts: EU Product Liability Directive 2024/2853
(applicable December 2026) -- certified prompt record means
liability moves to the provider. without it, you absorb the risk.
$ cat why.md
five eyes said "months, not years." they meant it.
10 hours CVE to working exploit (was 56 days in 2024)
29 minutes breakout time (CrowdStrike 2026)
22 seconds adversary hand-off (Mandiant 2026)
91% of orgs deploying AI agents have no governance
144:1 machine-to-human identity ratio
compliance proves controls exist.
l0k1 proves they work.
$ cat stack.md
inspector/ headless evaluation engine (python)
api/ scoring API (fastapi, cloud run)
avatar/ visual trust badge generator
extension/ chrome extension (slack, teams, chatgpt, claude)
altersid/ DID registry + verifiable credentials
gateway/ MCP trust-checking proxy
tests: 227 passing
infra cost: ~$15-30/month
license: open source
$ echo "built in sydney. shipped in public. 2026."
built in sydney. shipped in public. 2026.