Publication summary: QUASAR versus Swift15¶
Bounded by the dated finding, artifact manifest, and evidence index.
Canonical facts¶
One RTX 5090, matched pinned NInfer text-only 32K profiles, configured C4 and request C1. Swift15 passed 18/18 agentic scenarios; QUASAR passed 16/18. Both passed 20/21 whole coding solutions under the frozen list-pair probe. QUASAR reached 21/21 under a plausible tuple-pair sensitivity analysis; Swift15's varint defect remained. Keep Swift15 as default for now; no promotion occurred. This tiny fixture does not establish a general quality ranking or speed winner.
Copy-ready short post¶
On one RTX 5090 with matched NInfer text-only 32K settings, Swift15 passed 18/18 agentic scenarios versus QUASAR's 16/18. Coding was mixed and depended on an ambiguous pair representation. QUASAR is not a clear replacement; no promotion.
Copy-ready Reddit variant¶
We tested pinned QUASAR and Swift15 artifacts on one RTX 5090, with configured C4 but requests C1. A seven-task coding fixture ran three times per model and was independently graded. Frozen grades were 20/21 each; QUASAR reached 21/21 when ambiguous edge pairs were reconstructed as tuples. Swift15 had a real varint bug. QUASAR called the right tools but omitted the requested final status in two scenarios: 16/18 versus Swift15's 18/18. Timing was not controlled enough to rank speed. We restored the original Swift15 profile and made no promotion.
Screenshot alt text¶
Table of matched RTX 5090 functional results: tools 16/18 QUASAR and 18/18 Swift15; frozen coding 20/21 each; tuple-pair sensitivity 21/21 QUASAR and 20/21 Swift15. The table does not establish a general ranking or speed winner.
Claim ledger¶
| Claim | Conditions | Native evidence |
|---|---|---|
| Tools 16/18 versus 18/18 | Matched text32K, configured C4/request C1 | QUASAR, Swift15 |
| Frozen coding 20/21 each | Same seven tasks, three repetitions, 123 cases | QUASAR grade, Swift15 grade, fixture |
| Tuple sensitivity 21/21 versus 20/21 | Post hoc representation change only | QUASAR, Swift15, review |
| Inputs and reasoning preserved after redaction | Only declared operator metadata changed | lineage, raw index |
| Original profile restored; no promotion | Host identity sources remain private | restored preflight, verification, decision |