Platform health
CHECKINGChecking the deployed runtime.
Source /healthz →
Governed Inference · verifiable by anyone, offline
AI that signs its work
AI that signs its work
and refuses to lie.
Every answer arrives with a signed receipt you can check yourself — offline, in your own browser, with no server round-trip. When the model is not sure, it returns an honest BLOCKED instead of a confident guess.
—
Receipts in the chain
ledger depth — a mechanism check, not customer traction
ledger depth — a mechanism check, not customer traction
Conjecture 1
Λ trust gate · advisory bound
8
Locked Lean-proven theorems
a proof of rigor, not a traction number
a proof of rigor, not a traction number
The honesty doctrine · v11 LOCKED
Every figure on this page is labelled — or it isn't shown.
We grade what we say. If a claim isn't MEASURED live this session, we tell you exactly what it is instead of dressing it up as proof.
MEASURED
Read live from a running endpoint this session — signed-receipt count, advisory Λ posture, chain depth. Shown with a live chip; a dead probe degrades to an honest offline chip.
REPORTED
Stated by a source we name — the locked-8 Lean kernel, the 144-entry genome registry — cited here, not independently re-derived on this page.
UNKNOWN
We don't have it right now. Rendered honestly as N/A · UNAVAILABLE · BLOCKED — never a confident guess and never a fabricated number.
CONJECTURE
Advisory and unproven. Rendered in gray, never green, never a theorem. Λ lives here.
Λ = Conjecture 1
an advisory trust bound — never proven, never 1.0, and never rendered green. Theorem U is the proven conditional alternative; the unconditional claim is machine-checked false as stated.
No free energy. Joules are MEASURED only from a live exporter delta; with no live meter we label the reading SAMPLE or UNAVAILABLE and fabricate nothing. a11oy never claims over-unity, perpetual output, or free-energy efficiency — only what a real meter reports.
Runtime evidence
What is actually reachable right now.
Four read-only checks connect the product story to the running system. REACHABLE means the source answered this browser session; evidence mode (LIVE, SNAPSHOT, MODELED, or UNAVAILABLE) remains separate and never certifies model truth.
Tab contract
CHECKINGChecking the published tab inventory.
Source tab matrix →Receipt ledger
CHECKINGChecking the append-only evidence read.
Source receipt ledger →Sovereign mesh
CHECKINGChecking the current mesh posture.
Source mesh state →Checks have not completed.
Explore the estate
Nine surfaces. One governed core.
Every surface reads the same live endpoints and the same honesty doctrine. Jump straight in — the full working console is one click away.
⌘
Command Center
The full working console — signed receipts, the self-doubt gate, the genome registry and live organ vitals.
Open /console →
✦
Five Superpowers
The five things governed inference does that ungoverned models can't — each linked to the live check that backs it.
Open /superpowers →
⚔
Adversarial assurance
Run governed evaluations, inspect refusals, and trace every verdict to its evidence. The archived Warhacker route now resolves here.
Open the eval arena →
◎
Observability
OpenTelemetry-style traces and health summaries across the organs — live, read-only, no sign-on-GET.
Open /observability →
〜
Wires
The live wiring between organs — which probe feeds which surface, and whether it is up right now.
Open /wires →
◇
Mesh
The BFT-quorum mesh graph — 3-of-4 consensus, live state read from /api/a11oy/v1/mesh/state.
Open the mesh →
Σ
Formulas
The PURIQ agentic formula genome — live-recomputed values, each carrying its own Khipu receipt.
Open /formulas →
✓
Evidence
The trust center: the locked-8 with truthful Lean refs, Theorem U vs Conjecture 1, the receipt ledger, offline verify.
Open the evidence →
⇄
LLM Router
The LLM-Router Live scene — per-tier routing stats polled every second, otherwise an honest DEMO MODE.
Open the router →
The thesis
Governed AI you can prove — not AI you're asked to trust.
Most AI asks for trust. a11oy earns it the way auditable systems do: every governed decision is sealed into a signed, hash-chained receipt; a trust gate can only tighten an answer, never wave it through; and the kernel rests on machine-checked theorems you can read. The honest part is the product — when a claim isn't proven, we label it, and you can open the check yourself.
01 · SIGNS
It signs its work
DSSE receipts, ECDSA-P256 signed and SHA3-256 hash-chained into an append-only ledger. Signed on write, never on a read. Verify them in your own browser.
02 · REFUSES
It refuses when unsure
A self-doubt gate measures uncertainty against a bound and returns a truthful BLOCKED rather than a confident wrong answer. A refusal beats a fabrication.
03 · PROVES
It is formally backed
8 locked, axiom-free Lean 4 theorems back the kernel. The Λ trust gate is Conjecture 1 — an advisory bound, never dressed up as a theorem.
The proof · four honest tiers
Every claim carries exactly one honesty tier.
The 144-entry formula genome is graded across four tiers — and the rarest, strongest tiers are the smallest on purpose. Counts below are read live from the genome registry. A conjecture is shown in gray and is never rendered as proven.
—
LOCKED-PROVEN
Truthful labels for what each Lean obligation actually proves — e.g. F18 = Reed-Solomon parity (erasure tolerance), not "DSSE seal". Conservative and defensible.
—
SEMANTIC-VERIFIED
Sorry-free real theorems outside the frozen locked-8: the Λ min≤Λ≤max bounds, Theorem U (conditional uniqueness), DSSE verifiability. Where the real trust math lives.
—
EVIDENCE-BACKED
Runtime / algorithmic rules backed by real code and live endpoints, with no proof claim attached. Honest operational tier — not dressed up as proven.
—
CONJECTURE · ADVISORY
Λ unconditional uniqueness = Conjecture 1 (machine-checked false as stated). Gray, advisory, never green. Theorem U is the proven conditional alternative.
IMMUNE is a live, interactive demo of the same thesis you can drive yourself — a real hash-chained receipt ledger, a live Sigstore Rekor transparency anchor, and honest BLOCKED / UNAVAILABLE states instead of confident guesses. Nothing on that page is fabricated.
The verticals · live Λ posture
One governed core. Five places it earns its keep.
The same governed inference core runs across five verticals. Each card reads its live advisory Λ posture from the org-wide 13-axis gate (canonical floor 0.90). Λ is Conjecture 1 — advisory, never a pass/fail oracle.
a11
a11oy
Core · governed inference
The platform itself — command center, signed receipts, the self-doubt gate.
—advisory Λ
kil
killinchu
Defense · counter-UAS
Auditable interdiction: signed receipt per decision, 3-of-4 BFT consensus.
—advisory Λ
ins
Insurance
David Leads · lead scoring
Faithful David Leads scorer behind an F12 non-compensatory consent gate.
—advisory Λ
fin
Finance
Governed decisioning
Decision paths that have to be reconstructable and signed after the fact.
—advisory Λ
re
Real estate
Governed workflows
Workflow decisions captured into the same one-chain receipt ledger.
—advisory Λ
Λ
Org-wide Λ
13-axis weighted geomean
The composite advisory posture across all verticals, with min≤Λ≤max SEMANTIC-VERIFIED.
—floor 0.90
The living body
Watch the governed organism breathe.
The same live endpoints render as a 3D anatomical organism — organs glow when their probes are healthy and dim when they're not, and signed receipts travel the vessels as pulses. It's the operational state of the platform, shown as a body. Nothing decorative: every pulse is a real receipt.
Public estate · snapshot observed 2026-07-16
Models, kernels, data, and Spaces — one evidence-labelled map.
The atlas joins Hugging Face artifacts to their GitHub producers without pretending that a reachable URL proves alignment. Counts below are the dated public inventory snapshot; the atlas API upgrades them to LIVE or CACHED when its versioned source answers.
15Models
10Governed kernels
24Datasets
26Spaces
22Collections
1Public bucket
One chain · live ledger
One ledger. Every decision. Replayable to the byte.
Every vertical emits into a single SHA3-256 hash-chain — append-only, fsync-durable, replayable to a byte-identical root instead of trusting a dashboard.
—
Signed receipts
sha3_256
Chain algorithm
—
Chain depth (a11oy)
—
Last receipt id
Verify it yourself
The trust center.
The versioned honesty posture, the locked-8 with truthful Lean refs, Theorem U vs Conjecture 1 explained honestly, SLSA posture, and the offline receipt-verify flow. Every claim links to its check.