Dendra ($DNDR) — verifiable AI inference on consumer GPUs
A sovereign Cosmos SDK / CometBFT L1 where real AI inference replaces hashing as the work that secures the chain. Every answer settles on-chain, carries a proof block, and is backed by a bond a committee can slash. Devnet / research.
What Dendra is- A sovereign L1 on Cosmos SDK / CometBFT — its own chain, its own validator set. Not a rollup, not a token on someone else's chain.
- Proof of useful work. The security budget buys inference somebody actually wants, instead of hashes nobody wants.
- Miners are GPU operators serving an open model declared on-chain. Validators secure the chain; the two roles are separate and a machine can hold either.
- Prompts are end-to-end encrypted — the chain and the relay never see them in the clear. The miner decrypts in memory behind sealed-memory and egress guards: hardened deterrence on consumer hardware, not a cryptographic guarantee. A hardware-isolated datacenter tier is planned and does not exist.
Try it — one command, no accountBefore any of the claims below, here is the thing itself. No key, no account, no wallet:
curl -s https://api.dendranetwork.com/demo/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"model":"dendra","messages":[{"role":"user",
"content":"In two sentences, what is a verifiable random function (VRF) in cryptography?"}]}'
It is an OpenAI-compatible endpoint, so anything that already speaks that API speaks to Dendra. What a hosted API does not return is the object that comes back beside the answer:
"dendra": {
"job_id": "job1785688463489",
"miner_id": "dm1teqlpyx9sctv4d954eejvz",
"committee": ["dm1teqlpyx9sctv4d954eejvz"],
"beacon": "5ba76571429090bb7678d3ce0021b8e2…",
"audit_state": "pending",
"cost_udndr": 1780,
"escrow_udndr": 21210,
"verify": {"query": "dendrad query jobs get-job job1785688463489"}
}That block is the whole idea. The reply has an on-chain identity — which GPU served it, under which committee, against which beacon, at what cost, with how much still in escrow — and it ships with the query that settles it. Run that query and you are reading the chain, not this post. No closed API returns anything of the sort, because behind one there is nothing to look up.
You may notice the committee in that sample holds a single member, and that it is the miner itself. That is not a design where a node audits its own work: the jury is drawn from registered miners
other than the one under audit, and with one miner on the network there is nobody to draw. So the draw defers and the fee stays in escrow rather than a degenerate committee waving the job through. Two thresholds have to be crossed before any of it moves, and they are the subject of the section below.
How the verification worksLLM output is not deterministic, so byte-equality proofs are out. Dendra settles optimistically: one miner is paid for the job, a VRF-sampled share of jobs is re-audited by a fresh committee drawn on-chain, and the economics are set so that cheating loses money in expectation.
A hard slash needs
two independent locks, both priced in capital:
- at least two-thirds of the anchored committee seats returning "invalid" — the draw targets 15 seats and jurors are anchored on-chain before they may vote, so bending a verdict costs stake rather than identities;
- and a strict majority of the voting stake.
A false reveal upheld by a committee costs
slash_leak_bps = 8000 — 80% of the bond, a governable genesis parameter. Staying mute is not free either, but it goes through the
same lock rather than a second one: two-thirds of "invalid" verdicts posted against a missing reveal form a cheating quorum and route to that slash. Read the parameters off the chain rather than off this post:
dendrad query jobs params -o json — and note that
silence_slash_bps, though present and non-zero in genesis, is applied by no code path today. A parameter you can read is not a penalty that can fire, and this post would rather say so than let the query reassure you.
Measured, with its scope stated: on a single-machine bench where one host ran every identity, 2×100 audited jobs gave a
0.0% false-slash rate against honest miners, while a cheating node lost 84–99% of its bond and a silent one roughly 80% accumulated across many jobs. That is a bench number, not a network number. Reproducing it across independent operators is precisely what this endpoint exists to make possible.
Where the network standsThe endpoint is open and the settlement path has been armed since block 1. Jobs are served and settled there;
nothing is verified. The validator set is small and one operator's node carries the supermajority of the voting power, so there is
no fault tolerance: that node stopping stops the chain. A count of validators would not change this and this post will not quote one — read it yourself, at any time, from
curl -s http://api.dendranetwork.com:26657/validators, comparing the
voting_power values rather than counting the entries.
Two thresholds stand between here and a concluded audit, they are at different stages, and neither is a code change:
[list=1]
- The seed floor — met, intermittently. committee_min_vrf_contributors is 2, a contributor is a validator posting a VRF vote-extension, and the second one runs on a home connection: it reaches the proposer on a minority of blocks, so committees are drawn in windows rather than continuously. An independent, well-connected validator would widen them. Check: dendrad query jobs committee-seed-health -o json.
- The jury floor — not met, and this is what blocks pay. A drawn committee returns a verdict only when audit_min_quorum = 4 jurors vote, and the jury is every registered miner except the one under audit — five registered miners, against one today. Only a node run with --judge votes at all, and judging needs ≥ 12 GB VRAM or ≥ 26 GB RAM. Check: dendrad query jobs list-miner -o json and params.
Deferral is fail-closed and reversible. On this genesis
hold_bps is 10000, so at settlement the miner's entire net share is held rather than received — retained, not cancelled, with no expiry and no attempt counter — and the backlog drains on its own once both thresholds are crossed.
dendrad query jobs held-summary -o json returns the retained total, the count, and the age of the oldest.
Contribution is recorded, and nothing is promised. Every job, settlement, verdict and slash is written into the block that carries it. Be clear about what that record is worth: it lives in this chain's history, and a reset erases it. The project's stated intention is to publish a snapshot at a named height before any reset; until one is published and linked, assume there is none. $DNDR is a utility token with
no monetary value, never sold, on a chain that can be reset without notice.
Three commands settle the rest without taking this post's word for anything:
Run a nodegit clone https://github.com/DendraNetwork/dendra-network.git && cd dendra-network
Now read the script. It starts containers, creates keys and can bond stake, so it is worth five minutes:
⚠️ Run that one on its own — a pager is interactive, and anything pasted after it is swallowed as keystrokes rather than executed. Quit it with
q, then pick ONE of the three lines below. They are alternatives, not steps:
CONFIG_URL=http://api.dendranetwork.com:8088/network-info.txt bash deploy/join.sh
CONFIG_URL=http://api.dendranetwork.com:8088/network-info.txt bash deploy/join.sh --validator
CONFIG_URL=http://api.dendranetwork.com:8088/network-info.txt bash deploy/join.sh --judge
The default is a GPU miner plus its own full node.
--validator syncs, then walks through the stake bond and VRF key anchoring — never silently, since bonding locks stake.
--judge adds an audit committee seat, and the gate is the judge model a box can actually run, not the box itself: either a GPU large enough for a validated judge model, or — with no GPU at all — 26 GB of system RAM, which runs the validated mixture-of-experts judge on CPU. A CPU-only machine with enough RAM is a first-class juror.
deploy/hw_probe.sh measures the hardware, says which path it qualifies on, and refuses the seat rather than seating an under-powered one.
Testnet DNDR is free from the faucet at
http://api.dendranetwork.com:4500 — 10 DNDR per address per 24 h, proof-of-work gated. Use a fresh wallet, never a recovery phrase that controls anything real.
Role-by-role guide:
deploy/README.md in the repository. Questions:
discord.gg/bFc4kqBXW4
Tokenomics- 10,000,000 DNDR, fixed. The mint module is removed from the binary, so the supply holds by construction rather than by policy.
- Genesis split: 34% community · 33% Reserve · 27% treasury and validators · 5% team · 1% faucet float. Five allocations, no sixth.
- The protocol takes 15% of a job and burns 5% of the fee — usage puts deflationary pressure on a fixed supply.
- Payments settle from client fees and a pre-allocated Reserve, never minted. Any future community distribution would be carved from the existing community allocation. Whether one happens at all is undecided, and no amount, rate or schedule is offered or implied.
- Utility token. No sale, no presale, no ICO.
Roadmap- Now — open endpoint, settlement armed, committees drawn but never concluding. The near-term milestone is four more judging miners, which is what turns a drawn committee into a verdict and lets retained fees resolve; an independent validator would widen the windows in which draws happen at all.
- Next — judge robustness at scale, throughput measurement, audit and slash under load.
- Then — mainnet, considered only after a multi-operator network has actually run and the bench numbers have been reproduced outside a single machine.
FAQQ: Is the testnet rewarded?A: The settlement path is armed and a settled job splits the client fee: validators, treasury and the dev share receive coins immediately, while the miner's entire net share is held until an audit can conclude. The protocol therefore distributes and records a debt, and nothing is minted for any of it. Contribution is recorded, and nothing is promised — not a mainnet, not a distribution, and no amount, rate, ratio, date or eligibility. Arriving first carries no advantage of its own; there is no enrolment and nothing to claim.
Q: What GPU do I need?A: A consumer card able to serve an ~8B-class open model. The model you serve is declared on-chain; the registry that could pin it is implemented but not enforced on this network, so treat a served model as declared rather than pinned.
Q: Why use this instead of a hosted inference API?A: Not for answer quality — Dendra runs open models and makes no claim of beating frontier models. The axis is verifiability: the answer is anchored on-chain, contestable by anyone, and backed by a slashable bond. A hosted API asks to be trusted; this one can be checked. How much that is worth today depends on the two thresholds above, and they are stated plainly for that reason.
Q: Is content moderated?A: There is a deterministic regex floor with known false negatives. This is not a claim that illegal content is filtered out — do not rely on Dendra for legal-content compliance.
Q: How fast is it?A: Unmeasured on a real multi-operator network, so no number is quoted. Measuring it is one of the reasons this endpoint is open.
Links$DNDR ·
dendranetwork.comDevnet / research software · utility token · no token sale · the testnet token has no value · nothing here is financial advice · DYOR. 
