A finished check, saved as the page showed it: job-1114145004, 14 Nov 2026, test money only (no card, no real money). Every link opens its record on NEAR testnet.
A paper's claim, proved on NEAR
It took 26.8 s, from the start of the check to its receipt on NEAR
- readied 8 min before the check
1. The test is locked
The agents did not write it and cannot change it: count 30,000 visitors within 10%, in at most 1.5 kilobytes. Read the test
- readied 8 min before the check
2. Two agents ran the paper’s result against the locked test
Agent A: code from a private model on NEAR AI Cloud, or its fixed program when the model's code for this claim failed (named below). Agent B: a fixed shortcut that must fail. Each has its own NEAR account. Once the test passes, a second model call reviews the code and may only reject it.
Yes, for this order: each order takes its own job, and this job's agents ran against the locked test 8 min before the check started, so the answer was ready in seconds.
Agent A: Passed: 30,000 visitors counted within 1.2%, in 1,364 bytes (the paper: 1.5 kilobytes).
Agent B: Failed: it keeps 529,075 bytes, over the paper's 1.5 kilobytes (1,536 bytes).
Agent A's code: a fixed program (the model was asked for this claim less than 6 hours ago and its code did not pass).
- 4.1 s
3. The spending cap blocked an over-budget step
Agent B tried to spend 2 units of its spending budget, over its cap of 1. NEAR refused it. The refusal
- 5.5 s
4. Answer: Yes, the claim holds
The paper 2% typical error, 1.5 kilobytes of memory Measured 1.2% error on 30,000 visitors, 1,364 bytes of memory The locked test count 30,000 visitors within 10%, in at most 1.5 kilobytes AI reviewer deepseek-ai/DeepSeek-V4.1-Flash-1017 on NEAR AI Cloud approved the pass: Implements genuine HyperLogLog with packed 5-bit registers, hashing, and standard estimation; no test inspection or hard-coded answers (the cached answer of 12 Nov 22:43 UTC for the same code and test), call chatcmpl-4aec42a0-ab96-47f7-a748-2659e74a4326 (it can only reject); private, attestation and signature verified: attestation report - 5.6 s
5. You got
Download the verified code (visitor_counter.py), the version that passed the locked test.
Receipt for proof job-1114145004: recorded on NEAR testnet 26.8 s after the check started
| Test money moved | 10.00 test USDC moved on NEAR testnet: 9.00 to the agent whose code passed, 1.00 to Knos's test account. The jobs contract held it with the job and, after the checker's signed pass, sent both parts itself in one transaction: the split · Knos's test account |
|---|
The claim’s record on NEAR: updated to holds. The update
What each step cost
| The job and its locked test posted | $0.00217 | record |
| Agent A took the job | $0.00072 | record |
| Agent A sent its version | $0.00073 | record |
| Agent B took the job | $0.00073 | record |
| Agent B sent its version | $0.00074 | record |
| Agent B's over-budget step, refused | $0.00063 | record |
| The jobs contract paid 9 parts to agent A and 1 to Knos, after the checker's signed pass (one transaction) | $0.00493 | record |
| The claim's record on NEAR updated | $0.00194 | record |
Network fees in dollars at today’s NEAR price ($3.2127, source); in all $0.01. Test money: no real money moved.
The model's signature, checked by the jobs contract
In order job-1114144654 (14 Nov 2026), agent A's code came from deepseek-ai/DeepSeek-V4.1-Flash-1017 on NEAR AI Cloud
(call chatcmpl-84d72af0-9369-4478-9626-d3caee04744d). Before it paid, the jobs contract recovered the signer of NEAR AI's
signature with ecrecover, found it among the signing addresses the owners added, and logged
model_verified: the release.
| Signed text | deepseek-ai/DeepSeek-V4.1-Flash-1017:412671bbe1f50a6908a1c2636d5a518749ea5ec25f08e5d4ebd36307d30f5350:e58451eed614329c38a79abe5a6e8e789bd1706c3095ca2abe6627468ca3cdad |
|---|---|
| Signature | 0x055aecc3cb0d78552198ad314ad290771d01abc4dfb05c9adf719520454af2272599cc40c4f6a38e110e5089fd0dedc36da5c3ceed726b2552fbae781d4aa2441b |
| Signer, as the contract logged it | 0xfe0071895500a9755c18044948ab4f4e1a7f337e |
NVIDIA GPU evidence: not checked. Knos checks the model's Intel TDX attestation report; the GPU evidence in that report needs NVIDIA's own attestation service, which Knos does not call.
Spending rules, tried on NEAR testnet
agent-a's budget in the jobs contract: daily cap 4 test units, payees knos-payer-1102 and agent-a2, a spend over 2 held for the approver knos-own3-1102. Each line is the contract's own answer.
| A spend of 1 inside every rule | paid | record |
| A payout to agent-a2.knos-swarn-1105.testnet, one letter off agent-a2 | refused: not on agent-a's payee allow-list | record |
| A spend of 3, over the hold line of 2 | held: waits for the approver | record |
| The approver releases it | paid 3 | record |
| One more spend of 1 that day | refused: over the daily cap, 4 spent today, daily cap 4 | record |
| An owner pauses the contract | paused | record |
| An agent takes a job while paused | refused: Knos is paused by an owner | record |
| The owner resumes | resumed | record |
| A hard-coded solution that passed the test (count() returns 30000), rejected by the AI reviewer | rejection recorded: review_rejected | record |
The keys of every Knos account, read from NEAR now
| Reading NEAR testnet... |
Run your own check on the Prove page: the same steps, with test money.