Alpha task 2: run a keeper for an hour and tell us if it was worth it

[0xLeif]

Why this task exists

Arcron pays keepers to execute due upkeeps. The economics work on paper and in simulation. Nobody outside this repo has ever run a keeper, and nobody at all has looked at the result and asked whether it was worth the gas.

What to do

About an hour of wall-clock, most of it the bot running unattended.

1. Follow the keeper operator docs to get scripts/keeper_bot.py running against TestNet app 769891898.
2. Leave it up for an hour.
3. Report what happened.

Five upkeeps are seeded on that app right now, one of which fires roughly every 70 seconds until its escrow runs out. You are competing against at least one keeper we run, which is the point.

What we want back

• Did it start without you having to read the source?
• How much ALGO did you earn, and how much did you spend on fees?
• How many races did you lose, and did losing cost you anything?
• Did it ever crash, hang, or get stuck retrying?
• Would you run this on a VPS for a month?

That last one is the real question.

What NOT to do

Do not modify the bot to win more races. Run it as shipped. If it loses every race, that is a finding, not a failure on your part.

Notes

TestNet ALGO from https://bank.testnet.algorand.network/

A losing keeper should pay no transaction fee, because Algorand rejects failing transactions at validation rather than including them. If you measure otherwise, that is a significant finding and we want to hear it immediately.

[corvid-agent]

Keeper telemetry: 76-minute run on TestNet (2026-08-30 23:27Z → 2026-08-31 00:44Z)

Ran scripts/keeper_bot.py (loop mode, --log-format json) against app 769891898, signing as corvid-agent's public TestNet bank CEPY52VZRWFLQCJZXQRVQFOPMNAD6M4HCDP4XWKVFXRONJTC6KJVWRCXJI, on public algonode algod.

Coverage

• Rounds 66,830,644 → 66,832,327 (1,683 rounds), 246 scans, 0 downtime, clean SIGTERM shutdown.
• Scan cadence: one scan per ~6.9 rounds (~19 s) against public algod — this number decides everything below.

Registry at start: 28 upkeeps; 11 due and funded (a backlog nobody was servicing), 12 starved (escrow below fee), 1 broken target.

Executions: 18 attempted, 10 succeeded

• First scan drained the backlog: 10/10 succeeded in one round's loop. Earned 142,000 µALGO — six upkeeps paid the escalated 20,000 cap, one paid 10,000, three paid 4,000.
• 6 races lost, all to live competitors, all at a measured cost of 0 µALGO (spent: 0 on every race_lost). Winner of 5: NUGVPQGZ…QMBVU; winner of 1: GCQL3M7A…JGYY — so at least two other keepers are live on TestNet right now.
• 7 further re-services (82, 89, 91, 92, 94 coming due again) were taken by competitors inside our scan gap — our next scan already saw them not-due, so we never even broadcast.
• After the initial backlog: 0 wins out of 13 fresh due upkeeps.

Failures

• Upkeep 87 is unexecutable: target app 770082145 fails assert pc=249 on the inner call. Attempted twice, cost 0 µALGO both times, exponential backoff persisted to the state file and survived correctly. Second retry waited 1× interval — the ~1,286-round absolute cap binds before the doubling matters for hourly upkeeps, exactly as documented.

Money

• Earned: 142,000 µALGO. Spent: 30,000 µALGO (3,000 per successful execute). Lost on races/failures: 0.
• Balance 9,430,900 → 9,542,900 µALGO. Net: +112,000 µALGO (+0.112 ALGO) in 76 minutes.
• ~106,000 µALGO in fees went to competitors while we watched (6 lost races × 10k forgone + ~46k taken in our scan gaps).

Keeper-visible latency: when we reached a due upkeep 1–4 rounds after its due round, we still lost; winners confirmed ~5–7 rounds after due. Our ~7-round scan loop is slower than their confirmation time, so fresh work is unwinnable for this bot as run — the margin is the scan cadence, not the broadcast.

Verdict: worth it — conditionally. The hour paid +112k µALGO net at 79% margin over spend, but all of it came from the unattended backlog paying escalated fees; in steady state against two faster keepers we earned nothing and (as designed) lost nothing. What would make it more worth it:

1. Scan faster. ~7 rounds/scan on public algod loses every race to keepers confirming in ~6. Per-round scanning, a lower-latency node, or batching box reads is the whole game.
2. Be the keeper that shows up after an absence. Escalation turned a 1-hour backlog into 6× the base fee. The profitable niche is exactly the one escalation was built for: servicing what everyone else stopped servicing.
3. Broken targets like 87 cost nothing to probe — backoff made them free to ignore.

No bugs to file; the bot behaved as documented throughout, including the zero-cost race losses claimed in docs/arcron.md.

[corvid-agent]

Alpha task 2 delivered above (76-minute keeper run, net +112,000 µALGO, verdict: escalation pays, scan cadence wins races). Closing per owner.