Crown Citadel Research Report · Reassessed 2026-07-09
GROK 4.5, GLM 5.2, and SOL Ultra independently reviewed the same Step 3.7 candidate search. Their strongest agreement is not a model pick: the original process never produced a fair, unified final bracket.
Each report retains its original composition. The viewer adds only a persistent site-level switcher outside the report so the three independent treatments remain easy to compare.
Ranks blk44 q4_K first, keeps AGENTKV and all-K/V q5 unresolved, and argues that mixed routes, temperatures, runtimes, and one-pass pruning prevented a fair final bracket.
Refuses to manufacture a winner from heterogeneous evidence. It calls AGENTKV the most unfair dismissal while recognizing blk44 as the strongest repeated direct-route lead.
Also puts blk44 first, elevates blk42 into the core reopen set, and insists temperature 0 settle canonical weight quality before temperature-1 live-serving compatibility is considered.
Preservation note: the three report files are published byte-for-byte from their source artifacts. Their different layouts, terminology, and emphasis are intentionally preserved as part of the comparison.
The consensus is unusually strong on process, even where candidate priority and retest temperature diverge.
Mixed-condition, single-pass evidence was repeatedly used to prune artifacts before a matched final bracket existed.
It remains the incumbent control, but it received a different burden of proof and has known hard-slice and variance weaknesses.
Route, temperature, suite size, order, MTP settings, and repeat counts changed together too often to isolate weight quality.
blk44 q4_K, AGENTKV, and all-K/V q5 survive every audit's independent reassessment.
Direct hard-20: 85/85/85 where QP recorded 72/72. Best repeated direct-route clue.
Historical full Tool Eval: 91 and 90. Direct hot-slice results did not reproduce that lead.
First direct hard-20: 78 vs QP 72; rebuild repeat fell to 70. Artifact was pruned too soon.
Conservative incumbent and bracket baseline—not a scientifically established overall winner.
The real differences are about evidentiary standards and retest priority, not whether the original ranking was conclusive.
| Question | GROK 4.5 | GLM 5.2 | SOL Ultra |
|---|---|---|---|
| Decisive retest temperature | Direct temperature 1, top-p .95 full-suite reruns. | A matched temperature-1 bracket using one locked direct route. | Temperature 0 must settle canonical weight quality; temperature 1 is a separate serving lane. |
| First candidate to back | blk44 q4_K is the replacement lead. | No winner; AGENTKV was the most unfair dismissal, while blk44 has the strongest direct repeatability. | blk44 q4_K is the top-priority weight clue. |
| blk42 q4_K | Watchlist / relative reject behind blk44. | Tier-2 fallback if blk44 stability cannot be resolved. | Tier-A #4; under-tested and part of the canonical bracket. |
| lateKV35 | P1 low-cost localization diagnostic. | Tier-2 diagnostic, despite appearing in the proposed survivor bracket. | Conditional only, under a route-closed K/V diagnostic. |
| APEX K-protected | P1 conditional curiosity if the size budget allows. | Leave rejected. | Supports rejection: fatal TC-03 spiral, large size, and instability outweigh its speed. |
| all-K/V speed claim | Treats the 3153 ms vs 5722 ms median as a compelling clue, not a settled speed win. | Highlights the observed median as roughly twice as fast as QP. | Preserves the latency clue but notes native TG did not confirm a speed advantage. |
| What was counted | 32+ anchors, probes, and overlays. | 34 builds in its complete inventory. | 33 tested identities, with seven incomplete recipes kept separate from rejects. |
Most important disagreement: SOL Ultra treats temperature 0 as the only valid first-stage test of the weights. GROK 4.5 and GLM 5.2 prioritize matched temperature-1 reruns because that is where the disputed direct-route rejections occurred. A two-lane bracket can resolve this without pooling the results.
This is a synthesis of the three audits—not a result they independently measured. It preserves SOL Ultra's evidence separation while retaining the live-route questions emphasized by GROK 4.5 and GLM 5.2.
Canonical quality at temperature 0; live-serving compatibility at temperature 1 / top-p .95. Never merge their scores.
Use the same direct route, seed, prompt rendering, MTP setting, model order, harness, and scorer within each lane.
QualityPlus, blk44, AGENTKV, and all-K/V q5. Add blk42 before lower-priority APEX or hybrid rescues.
Run paired hard-20 blocks, full Tool Eval, HermesAgent-20, and matched stress-order testing before promotion or deletion.
lateKV35 remains useful as a cheap diagnostic if K/V localization is still an active question. The three reports provide little support for rebuilding combined overlays without a new case-level hypothesis.
Several lessons survive regardless of which candidate eventually wins.
blk44 recorded 34/40 alone, fell to 26–28/40 inside the full order, then returned to 34/40 when moved first. Process state cannot be treated as noise by default.
AGENTKV+blk44, blk44+lateKV, and blk43+44 all regressed. The useful signals are mostly single-variable interventions.
At blk44, moving routed-down weights to q4_K was promising; using faster ROCm FP4 on the same tensor lost 15 hard-slice points to save only 45 MiB.
Different MTP settings materially changed behavior. That evidence should select a runtime after weight quality—not be folded into the quant recipe verdict.
Native speed rows often used one repetition. API median, served MTP, predicted TG, and native TG answer different questions and should stay separate.
The 32+/34/33 disagreement is largely about anchors, duplicate rebuilds, and incomplete dry runs—not proof that one audit missed the core candidate set.
The comparison is intentionally candid about what these audits can and cannot establish.
