Skip to content

FLARE-FIX-001 closed diagnostic record

The six-arm cache-row allocation experiment is closed. It does not qualify the candidate, change the default, promote a serve, or submit an upstream patch. The incumbent remains the recommendation. The fixed-work causal assessment was independently reviewed and accepted as diagnostic causal support only. It remains outside qualification, default, promotion, and general-performance decisions.

The candidate raises startup continuation-row allocation from the preserved 25% default to 75% within the same physical 32 GiB host-cache pool. Startup verified A at 52 rows / 7,314 pages and B at 158 rows / 2,434 pages. The startup option accepts 1–99; omission remains 25%, and changing it requires a restart. The physical pool is 34,359,738,368 bytes in both arms; row/page counts are allocation geometry, not simultaneous full-context capacity.

The formal AB, BA, AB matrix retained 2,550 formal requests (1,152 prime + 1,398 measured), plus 162 warmups across 54 fresh incarnations. There were no formal request failures. All derived cells remain performance_eligible=false.

Decision

Fixed-work, a serial one-token reverse-replay workload, observed a 95.038% to 96.065% lower median client-generated TTFT for B over three complete 52/52 output/work-matched populations. The accepted independent review rehashed 1,182 ledger inputs and verified 312 requests, 52 work matches per pair, and A=10/B=36 eligible completed fetches for this exact workload. It also links the workload to 112→0 row-pressure increments across prime plus replay and 199,468→2 measured prefill tokens. It does not establish a general Flare or incumbent benefit.

Qualification remains blocked: the all-case sampled-output numerical gate failed; protected repetition-3 small12 and pageheavy populations regressed by 8.197% and 13.105%; and decode C4 output/work populations are unmatched. Natural-output lanes have no retained output token IDs, so their equality is unverified, not proven unequal. The initial independent-oracle, restore-fidelity, and causal proof receipts remain false.

Matrix

Medians and p95 are derived from the complete measured populations. Latencies are milliseconds; decode is tokens/second. p95 uses nearest rank ceil(0.95*n). Positive delta favors B for the lane metric. These are raw diagnostic observations, not accepted performance results.

Lane Rep Unit A native N A median / p95 B native N B median / p95 Observed delta Output/work state
original-reverse 1 ms 52/52 2232.611 / 3155.010 52/52 1679.467 / 3092.464 +24.776% equality unverified (IDs absent)
original-reverse 2 ms 52/52 2229.204 / 3156.222 52/52 1710.350 / 3082.410 +23.275% equality unverified (IDs absent)
original-reverse 3 ms 52/52 2229.920 / 3154.062 52/52 1638.094 / 3068.251 +26.540% equality unverified (IDs absent)
original-forward 1 ms 52/52 2406.638 / 3108.717 52/52 1418.142 / 2918.061 +41.074% equality unverified (IDs absent)
original-forward 2 ms 52/52 2399.688 / 3093.179 52/52 1573.695 / 2344.282 +34.421% equality unverified (IDs absent)
original-forward 3 ms 52/52 2403.207 / 3152.255 52/52 1521.556 / 2922.577 +36.686% equality unverified (IDs absent)
small12 1 ms 24/24 16.030 / 20.827 24/24 15.659 / 25.949 +2.314% 24/24 output/work matched
small12 2 ms 24/24 19.340 / 23.079 24/24 18.206 / 23.281 +5.863% 24/24 output/work matched
small12 3 ms 24/24 19.167 / 23.571 24/24 20.738 / 25.165 -8.197% 24/24 output/work matched
fixed-work 1 ms 52/52 469.982 / 899.879 52/52 18.494 / 26.306 +96.065% 52/52 output/work matched
fixed-work 2 ms 52/52 464.987 / 899.623 52/52 23.075 / 30.644 +95.038% 52/52 output/work matched
fixed-work 3 ms 52/52 467.585 / 901.555 52/52 20.236 / 32.879 +95.672% 52/52 output/work matched
pageheavy 1 ms 12/12 33.330 / 53.303 12/12 33.061 / 76.012 +0.809% 12/12 output/work matched
pageheavy 2 ms 12/12 33.199 / 62.964 12/12 33.385 / 58.161 -0.560% 12/12 output/work matched
pageheavy 3 ms 12/12 33.724 / 63.550 12/12 38.144 / 64.631 -13.105% 12/12 output/work matched
cold 1 ms 12/12 61.882 / 62.572 12/12 62.427 / 63.246 -0.881% 12/12 output/work matched
cold 2 ms 12/12 62.312 / 63.159 12/12 62.342 / 63.032 -0.049% 12/12 output/work matched
cold 3 ms 12/12 62.012 / 63.052 12/12 62.366 / 63.400 -0.571% 12/12 output/work matched
decode-c1 1 tok/s 1/1 192.015 / 192.015 1/1 192.089 / 192.089 +0.039% 1/1 output/work; batch equal
decode-c1 2 tok/s 1/1 192.038 / 192.038 1/1 192.174 / 192.174 +0.071% 1/1 output/work; batch equal
decode-c1 3 tok/s 1/1 192.048 / 192.048 1/1 192.379 / 192.379 +0.172% 1/1 output/work; batch equal
decode-c4 1 tok/s 4/4 106.517 / 109.560 4/4 108.085 / 109.419 +1.472% unmatched 0/4; batch unequal
decode-c4 2 tok/s 4/4 109.196 / 110.215 4/4 109.456 / 109.887 +0.238% unmatched 0/4; batch unequal
decode-c4 3 tok/s 4/4 107.171 / 109.342 4/4 107.030 / 108.499 -0.132% unmatched 0/4; batch unequal
small12-original 1 ms 24/24 1341.452 / 1735.142 24/24 1339.276 / 1742.159 +0.162% equality unverified (IDs absent)
small12-original 2 ms 24/24 1344.510 / 1743.917 24/24 1344.733 / 1737.556 -0.017% equality unverified (IDs absent)
small12-original 3 ms 24/24 1344.097 / 1735.787 24/24 1344.726 / 1967.618 -0.047% equality unverified (IDs absent)

The frozen strict analyzer marks all 27 strict lane/repetition comparisons incomplete. Twelve retained decode parser issues arise from the declared batch-before.prom artifact and are not request failures. The reviewed closure assembler supplies scoped functional evidence without changing the frozen normalizer or qualifying the matrix.

Scope and limits

The source pin is 47581ccac82660953a5b59cb04b2750d64edc97d; v2 patch SHA-256 is b132065b27532028bf2a02e754e86553ea846281f616f22ce74e89442ff0ec76 and image digest is sha256:d8fb208298fd3dd3f6b2b631148b01467a9449a9dcd1fef7dca304714bcf9411. Two TP2 GPUs ran at 300 W. The opt-in byte auditor passed its exercised A/B copies but needs 32 MiB host scratch per rank (64 MiB total) outside the pool and was disabled during formal timing.

Validation retains 44 focused tests, a 120-pass/0-fail/8-ignored CPU offload receipt, generic and actual-model transfer audits, and a later CPU shared-fetch lifetime fixture. That fixture is outside frozen v2. Missing evidence includes a numerical component/teacher-forced oracle, Restore/re-Snapshot state oracle, and live DMA cancellation overlap. The causal extractor used assert checks; do not run it under python -O. This campaign used ordinary Python and the independent rehash closed the campaign-specific risk; general reuse needs explicit exceptions and bounded regular-file reads.

Nine restoration windows are retained. r1-B keeps its unknown outer-runner exit and monitoring limitations. r2-B keeps its false default restoration proof and separately accepted 54.9174-minute supplemental receipt. r3-B was historically verified at 04:32:41Z after 49.7004166683 minutes; this is not a claim of perpetual health.

Upstream was observed 132 commits ahead of the frozen source pin; no submission was made. Before any default decision, isolate the protected regressions and build an independent numerical component oracle. The current recommendation is the incumbent.

Sanitized records: summary, report, coverage and gaps, restoration, normalized matrix, strict analysis, descriptive analysis, and causal assessment.

Retained evidence and reconstruction

This bundle retains all six formal arms and failed attempts. Archives preserve original evidence-relative member names; decompress each archive into the same root for offline inspection. The three large JSON analyses use gzip.

Artifact manifest binds public bytes. Provenance maps native source hashes to sanitized payloads and archive members; native hashes are not hashes of redacted public bytes. Source aliases preserve twelve legacy plain-stdout files and three long filenames under safe publication names. Recipes are reconstruction records, not executable private operator settings. Use the existing managed recipe load/status/logs/unload commands and refresh current source/host state before another run. No new plugin is required.

Reviewable contribution retains split patches, prerequisites and replay provenance; no upstream submission occurred.