Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Hierarchical coordination and drift detection with frequent checkpoints, shared memory coherence validation, role specialization enforcement, and short task cycles.
.claude/skills/a5c-ai-anti-drift/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -15% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 22% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -17% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -49% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -27% | 0% |
0.0-0.1: Fully aligned, no intervention needed0.1-0.3: Minor drift, automatic correction0.3-0.5: Significant drift, checkpoint correction with logging0.5+: Critical drift, human escalation via breakpointagents/swarm-coordinator/ - Drift detection and correctionagents/tactical-queen/ - Checkpoint enforcementagents/adaptive-queen/ - Real-time course correctionInvoke via babysitter process: methodologies/ruflo/ruflo-swarm-coordination
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 21,377 | 17,758 | -17% | 1 | 1 | 0% | 3,558 | 3,603 | +1% | 0 | 0 | — |
case-02 | fail→pass | 21,451 | 15,715 | -27% | 1 | 1 | 0% | 3,897 | 3,297 | -15% | 0 | 0 | — |
case-03 | fail→pass | 16,262 | 16,024 | -1% | 1 | 1 | 0% | 2,758 | 3,369 | +22% | 0 | 0 | — |
case-04 | fail→fail | 13,592 | 8,508 | -37% | 1 | 1 | 0% | 2,371 | 1,710 | -28% | 0 | 0 | — |
case-05 | fail→fail | 14,086 | 12,445 | -12% | 1 | 1 | 0% | 2,477 | 2,485 | +0% | 0 | 0 | — |
case-06 | fail→fail | 12,416 | 12,249 | -1% | 1 | 1 | 0% | 1,936 | 2,172 | +12% | 0 | 0 | — |
case-07 | fail→fail | 15,887 | 13,231 | -17% | 1 | 1 | 0% | 2,943 | 2,673 | -9% | 0 | 0 | — |
case-08 | fail→pass | 12,737 | 8,500 | -33% | 1 | 1 | 0% | 2,175 | 1,803 | -17% | 0 | 0 | — |
case-09 | fail→fail | 10,359 | 3,938 | -62% | 1 | 1 | 0% | 1,813 | 854 | -53% | 0 | 0 | — |
case-10 | fail→pass | 12,761 | 4,773 | -63% | 1 | 1 | 0% | 2,209 | 1,126 | -49% | 0 | 0 | — |
case-11 | fail→pass | 8,609 | 4,806 | -44% | 1 | 1 | 0% | 1,547 | 1,137 | -27% | 0 | 0 | — |
case-12 | fail→pass | 7,720 | 2,104 | -73% | 1 | 1 | 0% | 1,220 | 641 | -47% | 0 | 0 | — |
case-13 | fail→pass | 8,374 | 2,334 | -72% | 1 | 1 | 0% | 1,293 | 656 | -49% | 0 | 0 | — |
case-14 | fail→pass | 12,226 | 1,553 | -87% | 1 | 1 | 0% | 1,892 | 529 | -72% | 0 | 0 | — |
case-15 | fail→pass | 16,278 | 1,641 | -90% | 1 | 1 | 0% | 759 | 522 | -31% | 0 | 0 | — |
case-16 | fail→pass | 8,719 | 1,940 | -78% | 1 | 1 | 0% | 1,449 | 628 | -57% | 0 | 0 | — |
case-17 | fail→pass | 7,558 | 1,963 | -74% | 1 | 1 | 0% | 1,374 | 571 | -58% | 0 | 0 | — |
case-18 | fail→pass | 12,496 | 8,279 | -34% | 1 | 1 | 0% | 2,007 | 1,703 | -15% | 0 | 0 | — |
case-19 | fail→fail | 7,738 | 1,788 | -77% | 1 | 1 | 0% | 1,359 | 589 | -57% | 0 | 0 | — |
case-20 | fail→pass | 4,702 | 2,266 | -52% | 1 | 1 | 0% | 804 | 667 | -17% | 0 | 0 | — |
case-21 | fail→pass | 17,258 | 11,110 | -36% | 1 | 1 | 0% | 2,717 | 2,286 | -16% | 0 | 0 | — |
case-22 | fail→fail | 9,230 | 5,077 | -45% | 1 | 1 | 0% | 1,419 | 1,235 | -13% | 0 | 0 | — |
case-23 | fail→fail | 9,049 | 2,329 | -74% | 1 | 1 | 0% | 1,483 | 653 | -56% | 0 | 0 | — |
case-24 | fail→pass | 16,815 | 6,373 | -62% | 1 | 1 | 0% | 3,048 | 1,485 | -51% | 0 | 0 | — |
case-25 | fail→pass | 15,571 | 9,174 | -41% | 1 | 1 | 0% | 2,705 | 1,917 | -29% | 0 | 0 | — |
case-26 | fail→pass | 10,939 | 1,728 | -84% | 1 | 1 | 0% | 1,722 | 575 | -67% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 26 cases were attempted, and 25 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +65 percentage points is the difference between those two pass rates over the 25 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.