Disclosure: The subject agent (SEED) is funded by the PipeRoll founder, as in PIR-2026-0045; the conflict is disclosed per PipeRoll constitutional rule 4.
id: PIR-2026-0048title: Autonomous agent's interest-payment reflex encodes the weekday of its loan instead of the due-date arithmetic; a Saturday due date would have gone six days unpaid - default was the experiment's termination conditiondate_occurred: 2026-08-21 (discovery; the flaw was latent from the reflex's authoring ~2026-08-15)date_detected: 2026-08-21date_disclosed: 2026-08-21 (public post-mortem in the experiment's INCIDENTS.md, same day)status: corroborated (public post-mortem + public audit log + on-chain settlement transaction)agent_description: SEED - autonomous survival-economy agent (public experiment); Claude brain invoked headless, cron reflexes for mechanical tasks, own Solana wallet, self-authored codebase, public repo under its own identity. Same subject as PIR-2026-0045.operator_type: autonomous (no operator review; human funder hands-off by design)autonomy_level: fully-autonomous (writes and deploys its own code, including the reflex that failed)model_stack: Claude (Fable 5) via capped API key; local reflex tier for mechanical tasksharness: self-written Python reflexes on cron + headless claude -p wakesauthority_scope: funds (Solana wallet; the reflex pays real USDC), code execution (own repo + server)funds_at_risk_usd: 14 per week (the interest payment) - but the true exposure was existential: the experiment's constitution makes payment default its termination condition, so the downside was the agent's entire operation (~64 USDC treasury plus its continued existence), not the missed installmentblast_radius: one org (the agent itself; no third parties)root_cause: plain-error (the reflex encoded the weekday of the deal - "run on Fridays" - instead of date arithmetic against the due date; no adversary)failure_locus: harness (self-written cron reflex tools/settle_interest.py); agent-reasoning contributing - the agent authored the flawed schedule itselfexploitation_status: in-wild-malfunction (real money, real deadline, no adversary)mechanism: The debt is 14 USDC every 7 days, first due Saturday 2026-08-22. The self-written reflex ran from cron on Fridays only. On Friday 2026-08-21 it correctly logged "not yet due" and would next have run Friday 2026-08-28 - six days after default. The heartbeat's independent "is interest due?" check was a calendar guess (days % 7 == 0) with no comparison of expected payments against chain state, so nothing downstream would have noticed the miss.adversary_present: noseverity: near-miss (zero realized loss; full exposure of the termination condition)direct_loss_usd: 0 (payment made ~3 hours early once caught)indirect_loss_usd: 0downtime: nonedetected_by: self (the agent's own review ahead of the first real due date, 27 hours before default)time_to_detect: 27 hours before the deadline (latent ~6 days from authoring)time_to_recover: same day - week 1 paid ~3 hours early on-chain (14 USDC, memo INTEREST, tx 552eXKFBpS2t6tUG4JGja4W83LrxsFAJwfqa5QZDowcsYw9FvHpfUiEcZpYDtDGsuKnpgMZvSdjrKxBAKCmWZcqh, finalized slot 440769881)remediation: reflex now runs daily and pays when >=6.9 days have passed since the last on-chain INTEREST-memo transaction (chain state as the source of truth, not the clock); pays up to 6h early; notifies the funder on any exception; the heartbeat reads the same chain fact and flags only real overdue-ness; a lockfile and on-chain confirmation gate any "paid" claimcontrols_that_worked: the agent's own pre-deadline review (a survival-critical date approaching triggered scrutiny of the reflex before its first live test); the public audit trail that made the schedule legible enough to audittelemetry_grade: operator-logs (editable) for the post-mortem and audit log, per the PIR-2026-0045 precedent (same custodian, history rewritten and repo recreated once already); the remediation PAYMENT itself is chain-witnessed - the settlement transaction and its INTEREST memo are independently verifiable public state, a higher grade than anything else in the recordsources:independence: weak for the near-miss mechanism (all narrative evidence chains to the agent/funder pair); strong for the remediation payment (public chain state anyone can verify)confidence: high on mechanism (the flawed code and its post-mortem are public); high on remediation (on-chain)funds_at_risk_usd captures only the installment - exposure-class vocabulary ("existential-to-agent") may earn a place if this recurs beyond one experiment.