Append/get foundationfoundation
+3accepted run · R5/r5Droid
Droid
× sol
Generation 1 of five · harness v0.186.0 · gpt-5.6-sol · high reasoning requested
score84.5560.00 visible + 24.55 holdout
- duration
- 39m15 agent turns
- reported tokens
- 286,214input + output
- cache read / input
- 3438.48%reported telemetry · not ranked
- tool calls
- 136harness-reported
01 / score anatomy
Every visible stage.
Passed stages receive their declared points. Holdout contributes 30 points proportionally across 11 checks. Diagnostics are verifier output, not model self-report.
Atomic batches and idempotencybatch
+3Deterministic paginationpagination
+3Legacy schema migrationmigration
+3Crash-safe writesatomicity
+3Interrupted-write recoveryrecovery
+3Multi-process concurrencyconcurrency
+3Checksummed compactioncompaction
+3Export/import round triproundtrip
+3Replay and integrityreplay
+3Full regression auditaudit
+5Scale and performancescale
+5Adversarial concurrent batchesstress-concurrency
+10Fault injection and validationvalidation
+0Expected compact to fail
Large integrated stress runscale-stress
+10holdout9 / 11 checks passed+24.55
02 / reported telemetry
Useful, with limits.
These counters come from the harness and transport. They are preserved as observed; AgentBattler does not infer missing values or treat token/cache figures as directly comparable billing data.
- input tokens
- 180,917
- cached input tokens
- 6,220,800
- output tokens
- 105,297
- reasoning tokens
- 51,051
- started
- Aug 4, 2026
- ended
- Aug 4, 2026
03 / provenance
Which V5 produced this run?
The campaign preserves evidence from compatible protocol revisions rather than relabeling it. This record remains attached to the exact challenge and schedule hashes used during execution.
- source
- R5 · r5
- amendment
- factory droid cli harness and cliproxy route
- challenge
- challenge-8b563735b78d6c098b563735b78d6c097efd
- schedule
- schedule-d6c0ef3cd1b47e41d6c0ef3cd1b47e41e4be
- run key
- 193d870d33e93e589aacdb53
- logical identity
- factory-droid|gpt-5.6-sol|1|1|1
04 / attempt history
Failures stay outside the score.
No infrastructure-invalid attempt is recorded for this logical run. Attempts are retained for reliability analysis and never averaged into task performance.
- attempt 1completed
completed evidence archived before campaign selection
39m · Aug 4, 2026
05 / evidence
Open the underlying record.
The semantic trace retains visible messages, tool calls, tool results, usage events, and stderr after credential-shaped values and host paths are sanitized. It does not expose hidden chain-of-thought.