Fewer agent loops.
Not yet cheaper reasoning.
We eliminated a continuation storm and published the memory control that came next. The run count improved dramatically. Total token use did not.
- Total-run result
- 93.2% fewer runs
- Recorded model spend
- $0 / subscription-included
- Economic cost
- UNKNOWN
- Customer impact
- Not measured
The queue confused future intent with executable work.
A future-dated monitor shared ownership with generic continuation logic. The same issue kept waking even when nothing could be done. Over roughly 47 hours, one commercial issue accumulated 2,360 successful runs; 2,347 were generic continuation wakes.
This was an operating-cost failure, not customer demand. It produced activity without evidence, revenue, or delivery.
Run control worked.
The cost claim stops there.
| Metric | Before | After | Observed change |
|---|---|---|---|
| Runs | 280 | 19 | −93.2% |
| Retry-linked records | 274 | 1 | −99.6% |
| Generic continuation wakes | 271 | 0 | Eliminated in the measured window |
| Completed issues | 3 | 8 | +5; outputs were not equivalent |
| Input tokens per completion | ≈18.20m | ≈9.74m | −46.5%; denominator changed |
| Provider input tokens | 54,600,555 | 77,893,630 | +42.7% |
| Provider output tokens | 72,893 | 275,777 | +278.3% |
Source: company-scoped Paperclip heartbeat-run, issue, and provider-usage aggregates for consecutive 12-hour reviews ending 2026-07-15 23:46 UTC. No customer text, mailbox identifier, secret, or private evidence is included.
Give time-based work one wake owner.
A deterministic monitor owns the due time. Generic continuation is suppressed while the monitor is in the future. The no-change path is a tested state, not a reason to wake an agent.
Recall outcomes, not transcripts.
We then added a small outcome-memory plugin: at most five compact cards, cited sources, independent promotion, and helpful/harmful/stale feedback. Current evidence always overrides stored experience.
Start from a commercial delta.
Reviews begin with changed records, finance events, customer evidence, blockers, and prior commitments. Raw histories are opened only for a named anomaly.
The memory layer did not prove lower total token use.
Three fresh reviews before the plugin used 8.30m, 4.42m, and 2.68m input tokens. The first two fresh reviews after installation used 5.68m and 10.83m; a resumed follow-up added 1.56m. The samples are small, the tasks differ, and the post-install range is not lower.
- Recorded marginal model spend: $0. The provider classified these runs as subscription-included.
- Economic cost: UNKNOWN. Subscription allocation, labor, opportunity cost, and counterfactual provider pricing were not sourced.
- Revenue and customer impact: unproven. The measured windows contained no collected revenue and no accepted customer result.
- Causality is bounded. The monitor guard explains the duplicate-run collapse; outcome memory was installed later and has no controlled A/B.
Paperclip Outcome Memory
Inspect the control, not a victory lap.
The public package contains the generic ranker, independent review gate, feedback loop, example cards, threat model, tests, and clean-room setup. It contains no CyberNative seed cards or private operating evidence.
Install instructions and reproducible tests live with the exact source revision.
Proof Before Autonomy Audit
Find the failure before you scale the agent.
For $2,500, we test one workflow with up to 12 synthetic cases in five business days. No production access.