public event repositories successfully scanned into a structured inventory
Public scan of the event’s repositoriesOct 29, 2026Measured
Benchmarks
What we tested on Teleion Fabric, how well it performed, and what tradeoffs matter. Setup details, confidence intervals, adapter comparisons, and summary records live on one technical page: See how we measured this →.
Knows what’s new
top-1 Task Object match on held-out queries with a match (adopted Muse Glimmer 30B variant vs base; 83.7% correct abstention when nothing fits, trading 8 points of recall@5)
175 matched queries + 153 derived no-match cases on H100Nov 2, 2026Reuses work
correct structured answers across 60 recall and 60 request-splitting tests with zero context-shim tokens (separate 2x T4 memory adapter vs 222-token text shim; 2.28x lower amortized batch generation time)
120 held-out cases · fixed response budgetOct 31, 2026reuse-review routing, separate-key fit checks, and changed-material rejection regression cases passed before reuse delivery
Offline reuse-check test suiteNov 2, 2026Runs long queues
completed in a 256-slot local concurrency test, with 0 duplicate lease tuples across all 992 tested Tasks
5-tier local concurrency testOct 30, 2026late heartbeats and late submissions refused after lease expiry while a replacement worker finished the Task
Controlled local recovery checksOct 30, 2026median tool-call duration in a 16-pipeline local workflow test across 25,620 MCP calls
1,000 generated workflows on 4 local cellsOct 30, 2026to build a working nonprofit request kit in 25 model calls, with 8 passing unit tests and 0 rule violations
Recorded local test · example dataOct 31, 2026Model choices
Task-splitting F1 for the tuned Muse Glimmer 30B adapter; 96.2% of complete request objects matched exactly
106 held-out requests on H100Nov 2, 2026dated operational snapshot of build-step allocation between Muse Spark 1.3 and Muse Wonder 1.0 (check step: Spark 66% · Wonder 34%); reflects recorded outcomes, not proof of a better model
Operational policy snapshot, 15 outcomesNov 2, 2026Tool-call accuracy within a fixed response budget for the adopted Muse Glimmer 30B variant against base (at a longer response budget, 99.3% vs 96.7%)
150 held-out tool-call cases on H100Nov 2, 2026Checks catch mistakes
flawed reports accepted by the tuned Muse Glimmer 30B adapter; 99.6% balanced accuracy across 204 held-out reviews
204 held-out reviews (126 flawed) on H100Nov 2, 2026for a second worker with its own key to run the 8 unit tests and check every rule on the featured Eranos request
Recorded local test · example dataOct 31, 2026direct self-approval attempts blocked (plus 16 role and queue permission refusals) in local fault-injection checks
Local permission checksOct 30, 2026How we measured this
Training methods, holdout provenance, paired adapter comparisons, confidence intervals, response-budget conditions, serving scope, and summary JSON records are documented on one technical page: See how we measured this →.