Final Testnet Season

Each day the programme acts as an ordinary paying client and sends the network a fixed number of AI prompts (requests). The chain, not the programme, picks who serves each request, among the miners it counts as online (present miners, defined below). The chain also draws some requests for audit: a jury of other miners checks the answer.

When the day closes, a few answers of each identity are graded. The day is ranked, with its evidence, once its requests are final, and payment follows every seven season days.

The programme is a service the project runs for the season. The Final Testnet Season ends on 7 November 2026 at 23:59 UTC: it counts every block timestamped before 8 November 2026, 00:00 UTC, and the chain’s own block times decide. A request counts on the block it settles, so the programme opens none in the last ten minutes before the end, by block timestamp; on the last, shorter day the requests are spread over the time left. After the end, the service refuses new answers.

A request is verified when its answer reached the programme and the chain settled it for good, with no conviction against it.

In full: the programme got its answer, the chain settled it, and no audit of it can still be open (it is final). If it was audited, the jury must have upheld it, unless it falls under the delivered-work rule below. It never counts if its miner was slashed for it (penalised with part of its stake), nor if its fee went back to the client for any other reason.

Delivered work: paid from the day the season service names

A programme request refunded because no audit of it concluded in time is paid as a verified request when its answer reached the programme and the season’s grading did not reject it: its own grade first, otherwise its miner’s grades that day. A request with no grade at all, or whose miner was convicted, is never paid. The rule applies from the season day the season service publishes (reading its status…).

The chain marks such a job unwound: the client is refunded, and the miner is neither paid nor slashed by the chain. That happens when a drawn job’s jury does not conclude within audit_unwind_blocks of settlement, and also when the audit draw itself waits that long for a usable seed. An answer graded on its own is decided by that grade: coherent is paid, incoherent never, whatever the rest of its day says. An answer not graded itself follows its miner’s day: it is paid when that day has at least one grade and its work is not taken away (Graded answers). No grade, or a grading that cannot be read, is not paid. The rule’s first day is the status field unwound_audit_work_from_day, counted from 0; a day already ranked is never ranked again under it. Each ranking under the rule lists every unwound programme job of the day, paid or not, with the reason (inputs.unwound_audit_work). An unwound audit pays no juror: no juror judged.

Graded answers

When the day closes, a few answers of each identity are graded by a language model, which asks one question of each: is this a coherent attempt? If every graded answer is judged incoherent, that day’s work and presence rewards are lost; its juror rewards are kept.

At most three answers per identity are drawn from the first 50 that identity gave that day. The draw follows an order set by the service’s own secret, so no miner can tell in advance which answers are read. The drawn answers are published with their requests. Taking a day’s work away needs at least two grades, or one when a single answer was drawn, and a grade the model cannot give clearly is not posted.

Your stake earns no season reward by itself. It weighs in the draw that picks who serves each request: the chain weights each present identity by its stake, up to a ceiling of assignment_stake_cap_multiple × min_stake. At a multiple of 1, a stake above min_stake buys no more work.

On this testnet, neither the stake nor the number of identities protects against an operator who runs many identities, each with its own stake: what bounds the season’s total is the programme’s caps, 100 DNDR a day and 3 000 DNDR over the season (50 and 1 500 on the first four days).

Governance sets the multiple, and which one is in force is a reading of the chain, not a line of this page. Proposal 1 on this testnet proposes a multiple of 1; read its status with curl -s https://testnet-api.dendranetwork.com/rest/cosmos/gov/v1/proposals/1, and assignment_stake_cap_multiple and min_stake in

$ curl -s https://testnet-api.dendranetwork.com/rest/dendra/jobs/v1/params

The faucet’s drip size is the amount in its status, curl -s http://testnet-api.dendranetwork.com:4500/, next to amount_source, where it comes from. The faucet code of kit v0.2.1 can read it from the chain’s min_stake (DENDRA_FAUCET_AMOUNT=min_stake, faucet_daemon.py::drip_amount), so that one drip bonds one identity. Splitting a stake into identities bonded at min_stake keeps every part under the ceiling, so the whole stake weighs in the draw. The programme’s caps are cap_programme_day and cap_season, in each rule set of final_season_rules.py (RULE_HISTORY).

The ceiling is applied by committee.go::capAssignmentWeights. A field missing from the params output holds zero: an assignment_stake_cap_multiple of 0 selects the compiled 2, and a min_stake of 0 sets no ceiling. Proposals are read with curl -s https://testnet-api.dendranetwork.com/rest/cosmos/gov/v1/proposals. The audit jury is drawn by stake too, and that draw does not read this ceiling (audit_committee.go::drawMembersWithDomain).

Your points are the season rewards actually paid to you: the paid amount of each published ranking, after the day’s pro rata, one point per testnet DNDR. They convert one for one: each testnet DNDR the season paid becomes one mainnet DNDR, credited in the mainnet genesis to the address that received it. The season cap bounds the total at 3 000 mainnet DNDR. Mainnet launches after the Final Testnet Season.

Keep the key of the address that received your payments: the same key controls the same address on mainnet.

Before mainnet launches, the list of addresses and amounts is published with its hash; anyone can recompute it from the published rankings and the payout transfers on chain.

Each ranking row gives an identity’s counts, its payout address, its gross, payable and paid reward, and the reason for any shortfall. Amounts are in udndr: 1 DNDR = 1,000,000 udndr.

A season day is 17 280 blocks, counted from the season’s start height: about 24 hours at about 5 seconds per block, an interval the chain does not fix. The last day is cut at the season’s end, so how many days the season holds follows from the chain’s pace. A day is ranked once all its requests are final, after audit_unwind_blocks; the last day’s ranking names the season’s last block (inputs.season_end_height). A week is seven season days; the last payment covers the days left after the last full week, once the last day is final.

Before paying a week, the payment checks every published day so far against the rules, and pays nothing for the week if one breaks them. No IP address enters the evidence or the formula.

$ docker compose exec miner python3 final_season_rank.py --day 0 \
    --evidence https://testnet-api.dendranetwork.com/final-season/v1/ \
    --node tcp://testnet-api.dendranetwork.com:26657 \
    --generator GENERATOR --start START \
    --compare https://testnet-api.dendranetwork.com/final-season/v1/ranking/day-000.json