Avaira Protocol
“Unknown is a useful answer when the evidence is thin.”
An experiment in making agents pause when they are unsure
> prove ....... SAFE > witness ..... none > cost ........ $95 < $100 > gate ........ OPEN
- Project
- 02
- Category
- Agent infrastructure
- Year
- 2026
- Status
- Experimental
- License
- Repository
- Built with
- Python · OPA
I wanted to explore a less comfortable question: what should an AI system do when it cannot tell whether an action is safe?
Avaira is a place to test that idea. It separates planning, checking, simulation, and execution instead of letting one confident answer do all four jobs.
The project is still experimental. I am more interested in learning where the design breaks than in presenting it as a complete safety system.
The order matters. Each step is what earns the next one the right to run.
- 01
Reason, with a trace
A System-2 planner refuses to emit a plan without a complete reasoning trace. Working memory is capped at 7±2 slots with a protected goal chunk.
- 02
Prove it, or don't
An interval-arithmetic symbolic prover returns SAFE, UNSAFE with a witness, or UNKNOWN. UNKNOWN fails closed.
- 03
Simulate it
A seeded 200-run Monte-Carlo shadow sandbox exercises the plan before anything is allowed near real state.
- 04
Gate it
Execution requires attestation plus an HMAC-signed SAFE certificate bound to the plan hash, and a worst-case cost under the cap.
- 05
Settle it
Violations settle atomically against a staked ledger via a freeze-and-slash adapter. The ledger is hash-chained and tamper-evident.
Claims, and how you'd check them
A claim without a way to test it is marketing. Each of these has a method.
Safety is proven before execution, not asserted after
check · Interval-arithmetic prover, strict three-way verdicts
The gate is a strict boolean, not a confidence score
check · Attestation + signed certificate bound to the plan hash + cost cap
Memory writes need artifacts
check · Self-editing memory rejects artifact-less writes; reversals need higher confidence or two independent artifacts
The same input produces the same output, byte for byte
check · CI re-runs the demos twice and asserts byte-identical results
The scope of a tool is part of its honesty. These are the edges of Avaira Protocol, stated by the person who wrote it.
- The proven-safety claim covers the kernel's own reasoning, not the correctness of your policy set
- Deterministic and offline by design — it is not a hosted service
- Settlement adapters are illustrative of the ledger design, not an audit of any chain
NGS-Agent / GeneOS
A bioinformatics side project for reading sequencing run evidence