QUASAR versus Swift15: evidence index¶
Completed 2026-10-16 functional comparison on one RTX 5090 (32,607 MiB). Keep Swift15 as default for now; QUASAR is not a clear upgrade. No promotion occurred.
Both final profiles used the same pinned NInfer image, text-only 32K context, configured C4 and actual request C1. Coding used seven tasks repeated three times, with 4,096 visible allocation + 8,192 reasoning headroom and a combined 12,288-token cap; the endpoint does not partition channels. Medium reasoning, temperature 0.6 and top-p 0.95 were matched. Seeds and cache warmth were not fixed.
| Result | QUASAR | Swift15 | Retained evidence |
|---|---|---|---|
| Agentic scenarios | 16/18 | 18/18 | QUASAR, Swift15 |
| Frozen coding cases / solutions | 119/123; 20/21 | 120/123; 20/21 | QUASAR, Swift15 |
| Tuple-pair sensitivity | 123/123; 21/21 | 120/123; 20/21 | QUASAR, Swift15 |
The tuple analysis is post hoc, changes only pair reconstruction, and preserves the original grades. The ambiguous prompt prevents a general coding ranking. Both QUASAR tool misses omitted the requested final status after correct calls. Swift15's coding miss was an unambiguous varint implementation error.
Reproduction and provenance¶
- Native raw-output index, including full coding responses and reasoning.
- Frozen fixture and grader, manual source review.
- Configuration and recipes, run plan.
- Pinned identity sources; no external performance prior informed the decision.
- Redaction and source/public hash lineage.
- Ten-role manifest, decision summary, publication summary.
The installed collection harness was 1.5.0/Python 3.14.8 with a retained source fingerprint; its source revision was unavailable. The dirty helper/publication worktree declared 1.2.1 and is not claimed as the collection source. All ten artifact roles retain files. Native schemas, prompts, outputs, reasoning and case results are preserved; only declared operator metadata was redacted. Original pre-redaction grade input hashes map to public copies through lineage.
Failures, restoration and limits¶
The failed 262K vision/text profiles are historical phases, not scored answers. QUASAR's measured 32K container and the fresh Swift15 test container were unloaded; the original Swift15 262K profile was restored. Restoration record, public restored preflight, and independent verification are retained. Detailed host-state assertions are reported by that verification record; underlying operator identity/status sources remain private.
The friction log preserves startup, transport and tooling gaps. The checksum-offload experiment was negative and restored every setting. The cache-transfer host performed no inference and is not measured hardware. Native Windows SWE, long-context/vision qualification, and a controlled performance comparison were not completed. Timing remains descriptive. Recipe reconstructions preserve the as-run prequalification labels; they are not promotion records.