Atomic imports
Make manifest imports all-or-nothing, including when a guard returns early.
Run history
Flash-Next IQ3_SRails Lemans · pilot · 2026-10-12
- Agent time
- 3m 20s
- Official checks
- 10 / 10
Run settings & provenance
- Run ID
- rails-pilot-20261012/ar-atomic-import
- Requested output allowance
- 196,608
- Min p
- 0.0
- Model
- openai/qwen3.8-flash-next-iq3_s#medium
- Presence penalty
- 0.0
- Reasoning effort
- medium
- Repetition penalty
- 1.0
- Temperature
- 1.0
- Top k
- 20
- Top p
- 0.95
- Agent
- miniswen-installed
- Lemans version
- 1.5.1
Timing, delivery & token accounting
- Timing seconds · Agent execution (seconds)
- 200.228
- Timing seconds · Wall clock (seconds)
- 225.9
- Usage · Input tokens
- 130,386
- Usage · Cached tokens
- 123,450
- Usage · Output tokens
- 7,205
- Usage · Steps
- 18
- Patch · Bytes
- 6,090
- Patch · SHA256
- 3e07084de5db125819415470337b4ad941d141ff43e513d42aadb0c751674674
- Verification · Status
- official-scored
- Verification · Verifier
- Rails Lemans official verifier
- Verification · Evidence
- verified campaign report and native verifier checks
Published comparisons
Published comparison sources
Rails-published successful attempts out of three runs per task, not individual checks. Provider models and efforts vary by benchmark; these are not matched local configurations.
Claude Opus 5 · high3/3 passed
Claude Fable 5.1 · high3/3 passed
Kimi K3 · max3/3 passed
Claude Fable 5 · high3/3 passed
GPT-5.6 Sol · medium3/3 passed
Grok 4.6 · high3/3 passed
GLM 5.3 · max3/3 passed
Claude Opus 4.8 · high3/3 passed
GPT-5.6 Terra · medium3/3 passed
Qwen3.8-27B · xhigh3/3 passed
Muse Spark 1.2 · medium3/3 passed
GPT-5.6 Luna · medium3/3 passed
Gemini 3.7 Flash · medium3/3 passed
Claude Sonnet 5 · high3/3 passed
GLM 5.2 · high3/3 passed
DeepSeek V4 Flash · high3/3 passed
GLM 5.3 Flash · max3/3 passed
Official assignment
Recorded assignment context
Books can be filled from an export manifest through `Book#import_manifest`, which presses each entry as a section. Imports are supposed to be all-or-nothing — and support tickets say they are not: when a manifest has a bad entry partway through, the import reports failure, but every entry *before* the bad one is left sitting in the book. People re-run the fixed manifest and end up with the first half duplicated.