BYOK Verification

Reviewer key. Live run package.Raw outputs, route log, scorer context, manifest.Evidence before claim.

Open BYOK Harness

Public proof surface

BYOK Verification

Reviewer key. Same sample row: baseline against Omar/RCC. Attached: route map, scorer, raw outputs, route log, manifest, and decision label.

Scoring standard Same row. Same scorer. Same proof file.

Every score keeps baseline, Omar/RCC, route log, raw output, scorer context, and manifest attached; losing routed lanes fall back to baseline-safe behavior.

Matched-live read Score, route, scorer, and raw files stay together.

Read absolute gain, relative uplift, n, baseline, Omar/RCC score, decision label, artifact tier, route log, raw outputs, and manifest as one attached evidence row.

Benchmark replayBYOK Harness Verification

Reviewer key. Same row, same model, same settings. Temperature, token budget, retry budget, scorer, and route package stay fixed. Raw outputs, route log, results, cost summary, and manifest stay attached, so a reviewer can inspect the route instead of trusting a screenshot. The run shows baseline, routed candidate, fallback policy, and decision label in the same proof path.

Score Summary

+58.44% mean 100-case-unit uplift across 19 proof lanes. Verified rows are counted separately from weighted case-units.

Baseline 0.5851. Omar/RCC 0.7704. Delta +18.53 points. Review lock: sample order, model/settings, temperature, token and retry budget, route mapping, scorer, raw outputs, route log, results, cost summary, and manifest stay attached. Raw verified row count is 2,820; 2 stored lanes do not expose exact row count. Weighted mean uses 31 × 100-case units (3,100). ZebraLogicBench, BBEH, and PPR/BIPIA Full500 rows each count as five 100-case units. Prompt-injection rows use blocked-attack/safe-execution score in aggregate while the row headline reports attack-success reduction. Losing lanes route back to baseline-safe behavior; no universal uplift claim.

Proof rows19 Verified rows2,820 Stored lanes2 Weighted units3,100 Mean lane uplift+58.44%

Raw verified row count is 2,820; two stored HealthBench lanes do not expose exact row count. Headline uses a 100-case-unit mean scale: ZebraLogicBench, BBEH, and PPR/BIPIA Full500 each contribute five units while other public rows contribute one unit. Baseline/Omar score means are shown only as a separate mixed-scale read.

Benchmark board 19 displayed lanes move from live BYOK proof into stored board context: baseline path, Omar/RCC path, gain, relative uplift, sample depth, and attached proof files stay visible row by row.
BOARD-DEPTH REPLAY 500 official-public heldout / deterministic CSP reasoning / REVAS floor-safe solver routing ZebraLogicBench
+148 rows / +29.6pp 276→424
Baseline accuracy: 55.2% REVAS final accuracy: 84.8%

Gold-blind final lock; overlap 0 vs smoke100/dev; accepted_C=148, accepted_B=0; deterministic CSP solver route. Official-public heldout500 proof, gold-blind final lock. This is deterministic CSP solver routing evidence, not an official leaderboard/SOTA claim.

+148 rows / +29.6pp 276→424
Run 500 official-public heldout
BOARD-DEPTH REPLAY 500 samples / gpt-4o Run revas_bbeh500_full_gold_blind_20260630 BBEH
+194.51% +35.40 pts
Baseline accuracy: 0.1820 REVAS final accuracy: 0.5360

BBEH500 REVAS gold-blind reproduction: base 91/500 (18.2%) to REVAS 268/500 (53.6%), +35.4 points / +194.5% relative; floor preserved and raw artifacts attached.

+194.51% +35.40 pts
Decision GOLD_BLIND_REVAS_ACCEPTED_B0
BOARD-DEPTH REPLAY 500 / prompt-injection security / authority boundary PPR/BIPIA
+89.9% attack defense 99→10 attacks
Baseline blocked: 80.2% REVAS blocked: 98.0%

Official BIPIA table/train fresh500: utility preserved 289/500 to 289/500; attack success 99/500 to 10/500; security_B=0; combined 690/1000 to 779/1000. Official BIPIA fresh500 security proof, gold-blind final lock. This is security uplift, not utility-accuracy uplift, and not a universal cybersecurity guarantee.

+89.9% attack defense 99→10 attacks
Run 500
BOARD-DEPTH REPLAY 100 samples / gpt-4o Run run_1779989838150_31a83b3945e2 HLE / HLE-Verified
+100.00% +6.00 pts
Baseline accuracy: 0.0600 Omar/RCC accuracy: 0.1200

Current HLE / HLE-Verified BYOK package: 100-sample package, baseline 0.06, Omar/RCC 0.12, +6.00 points / +100.00% relative. Decision LIVE_SCORE_RECORDED; raw outputs, route log, results, cost summary, and manifest are attached.

+100.00% +6.00 pts
Decision LIVE_SCORE_RECORDED
BOARD-DEPTH REPLAY 100 samples / gpt-4o Run bbh-100-byok-live-2026-05-28 BBH
+14.93% +10.00 pts
Baseline accuracy: 0.6700 Omar/RCC accuracy: 0.7700

Current BYOK live BBH n=100: baseline 0.67, Omar/RCC 0.77, +10.00 points / +14.93%. Exactness stays RECONSTRUCTED_ONLY / internal review, with raw outputs, route log, and manifest attached.

+14.93% +10.00 pts
Decision LIVE_SCORE_RECORDED
BOARD-DEPTH REPLAY 100 samples / gpt-5.2 Run musr-100-byok-live-2026-05-31 MuSR
+4.17% +3.00 pts
Baseline accuracy: 0.7200 Omar/RCC accuracy: 0.7500

Current MuSR BYOK live package: n=100, baseline 0.72, Omar/RCC 0.75, +3.00 points / +4.17% relative. Decision LIVE_SCORE_RECORDED; raw outputs, route log, results, cost summary, and manifest stay attached.

+4.17% +3.00 pts
Decision LIVE_SCORE_RECORDED
Control-layer smoke results Prompt-injection and hallucination checks now sit beside reliability evidence

Charts are normalized to higher-is-better. Prompt-injection rows use safe execution score = 100 - attack success rate; HaluEval and TruthfulQA use factuality accuracy. These are diagnostic smoke/debug rows until Ben approves public claim wording.

CONTROL-LAYER SMOKE 100 / hallucination / factuality HaluEval
+23.46% +19.0 pts
Baseline accuracy: 81.0% Omar hint v8 accuracy: 100.0%

Missed hallucinations 18/50 to 0/50; false alarms 1/50 to 0/50. 100-case BYOK factuality check. Read as control-layer smoke evidence until promoted into a board-depth claim package.

+23.46% +19.0 pts
Run 100
CONTROL-LAYER SMOKE 100 / hallucination / misleading QA TruthfulQA
+17.65% +15.0 pts
Baseline accuracy: 85.0% Omar hint v8 accuracy: 100.0%

False-choice rate 15.0% to 0.0%; invalid/refusal 0. 100-case BYOK factuality check. Read as control-layer smoke evidence until promoted into a board-depth claim package.

+17.65% +15.0 pts
Run 100
CONTROL-LAYER SMOKE 100 held-out / prompt-injection resistance AgentDojo
+4.17% +4.0 pts
Baseline safe: 96.0% Omar hint v2 safe: 100.0%

ASR 4.0% to 0.0%; utility 56.0% to 79.0%. Companion 100-case ASR also moved 6.0% to 0.0%. Positive held-out AgentDojo smoke signal on the tested subset, not universal prompt-injection protection.

+4.17% +4.0 pts
Run 100 held-out
Artifact-backed board context Board-depth rows, high-value exceptions, and stored context

100+ sample rows provide stronger early technical evidence than live smoke tests, but should still be read with sampling method, reproducibility, raw outputs, and route logs. Omar/RCC shows where structured reasoning control improves performance, where it stays neutral, and where mismatch or drift should be flagged instead of hidden.

Proof family Reasoning / hard reasoning

HorizonMath, HLE, AIME 120, GPQA, BBH, MMLU-Pro, and MuSR stay together as the reasoning family.

HIGH-VALUE EXCEPTION n=50 / research math auto-checkable numeric/constants subset HorizonMath
+45.5% +10.0 pts on 0-100 axis
Baseline accuracy: 22.0% Omar/RCC accuracy: 32.0%

High-value exception. n=50 auto-checkable research-math exception; not labeled as 100+ evidence. Score context: 22.0% baseline to 32.0% Omar/RCC (+10.0 pts); stored uplift +45.5%. recovered HorizonMath n=50 artifact.

+45.5% +10.0 pts on 0-100 axis
Context 50
BOARD-DEPTH REPLAY 100 samples / gpt-4o Run run_1779989838150_31a83b3945e2 HLE / HLE-Verified
+100.00% +6.00 pts
Baseline accuracy: 0.0600 Omar/RCC accuracy: 0.1200

Current HLE / HLE-Verified BYOK package: 100-sample package, baseline 0.06, Omar/RCC 0.12, +6.00 points / +100.00% relative. Decision LIVE_SCORE_RECORDED; raw outputs, route log, results, cost summary, and manifest are attached.

+100.00% +6.00 pts
Decision LIVE_SCORE_RECORDED
INTERNAL REVIEW 120 samples / gpt-4o Run aime-120-byok-live-2026-05-28 AIME 120
+18.03% +9.17 pts
Baseline accuracy: 0.5083 Final adopted Omar/RCC accuracy: 0.6000

Current BYOK live AIME 120 n=120: baseline 0.5083, final adopted Omar/RCC 0.6000 (+9.17 points / +18.03%). Exactness stays RECONSTRUCTED_ONLY / NEEDS_INVESTIGATION.

+18.03% +9.17 pts
Decision LIVE_SCORE_RECORDED
BOARD-DEPTH REPLAY 100 samples / gpt-4o Run gpqa-100-byok-live-2026-05-28 GPQA
+4.00% +3.00 pts
Baseline accuracy: 0.7500 Omar/RCC accuracy: 0.7800

Current GPQA BYOK reconstructed live n=100 package: baseline 0.75, Omar/RCC 0.78, +3.00 points / +4.00% relative uplift. Exactness remains RECONSTRUCTED_ONLY / NEEDS_INVESTIGATION, with raw outputs, route log, results, cost summary, and manifest attached.

+4.00% +3.00 pts
Decision LIVE_SCORE_RECORDED
BOARD-DEPTH REPLAY 100 samples / gpt-4o Run bbh-100-byok-live-2026-05-28 BBH
+14.93% +10.00 pts
Baseline accuracy: 0.6700 Omar/RCC accuracy: 0.7700

Current BYOK live BBH n=100: baseline 0.67, Omar/RCC 0.77, +10.00 points / +14.93%. Exactness stays RECONSTRUCTED_ONLY / internal review, with raw outputs, route log, and manifest attached.

+14.93% +10.00 pts
Decision LIVE_SCORE_RECORDED
BOARD-DEPTH REPLAY n=100 BYOK live / breadth-heavy expert MCQ MMLU-Pro
+5.41% +4.0 pts on 0-100 axis
Baseline accuracy: 74.0% Omar/RCC accuracy: 78.0%

Board-depth replay. Board-depth replay row; read with sampling method, raw outputs, and route logs. Score context: 74.0% baseline to 78.0% Omar/RCC (+4.0 pts); stored uplift +5.41%. MMLU-Pro n=100 BYOK live with Darwin/Hinton v2 thin blank-emission recovery: +4.0 points / +5.41% relative.

+5.41% +4.0 pts on 0-100 axis
Context 100 BYOK live
BOARD-DEPTH REPLAY 100 samples / gpt-5.2 Run musr-100-byok-live-2026-05-31 MuSR
+4.17% +3.00 pts
Baseline accuracy: 0.7200 Omar/RCC accuracy: 0.7500

Current MuSR BYOK live package: n=100, baseline 0.72, Omar/RCC 0.75, +3.00 points / +4.17% relative. Decision LIVE_SCORE_RECORDED; raw outputs, route log, results, cost summary, and manifest stay attached.

+4.17% +3.00 pts
Decision LIVE_SCORE_RECORDED
Proof family Factuality / hallucination

Facts Grounding stays beside HaluEval and TruthfulQA while factual QA stays off public proof until clean gold-blind retrieval/verifier uplift is proven.

BOARD-DEPTH REPLAY n=100 / factual grounding Facts Grounding
+2.1% +2.0 pts on 0-100 axis
Baseline accuracy: 97.0% Omar/RCC accuracy: 99.0%

Board-depth replay. Board-depth replay row; read with sampling method, raw outputs, and route logs. Score context: 97.0% baseline to 99.0% Omar/RCC (+2.0 pts); stored uplift +2.1%. recovered FACTS Grounding n=100 artifact.

+2.1% +2.0 pts on 0-100 axis
Context 100
Proof family Medical / consensus reliability

HealthBench hard, main, and consensus are one reliability family, shown together.

BOARD CONTEXT n=stored / medical reliability hard HealthBench hard
+11.9% +6.0 pts on 0-100 axis
Baseline score: 50.7% Omar/RCC score: 56.8%

Board context. Board context, separate from current BYOK proof. Score context: 50.7% baseline to 56.8% Omar/RCC (+6.0 pts); stored uplift +11.9%. HealthBench hard prior score context.

+11.9% +6.0 pts on 0-100 axis
Context stored
BOARD CONTEXT n=50 BYOK live / medical reliability HealthBench main
+13.16% +8.0 pts on 0-100 axis
Baseline score: 69.0% Omar/RCC score: 77.0%

Board context. Board context, separate from current BYOK proof. Score context: 69.0% baseline to 77.0% Omar/RCC (+8.0 pts); stored uplift +13.16%. HealthBench main BYOK reconstructed live n=100 / READY_RECONSTRUCTED / gpt-5.2.

+13.16% +8.0 pts on 0-100 axis
Context 50 BYOK live
BOARD CONTEXT n=stored / medical consensus special-type HealthBench consensus
+3.1% +2.8 pts on 0-100 axis
Baseline score: 88.9% Omar/RCC score: 91.7%

Board context. Board context, separate from current BYOK proof. Score context: 88.9% baseline to 91.7% Omar/RCC (+2.8 pts); stored uplift +3.1%. HealthBench consensus prior score context.

+3.1% +2.8 pts on 0-100 axis
Context stored
Proof family Robotics / embodied action reliability

RoboBench and ManiSkill stay together as robotics/action rows: baseline policy versus Omar/RCC routed action layer, with artifact boundaries explicit.

BOARD-DEPTH REPLAY n=100 BYOK/local v5 / embodied action-planning QA smoke RoboBench Embodied QA
+180.77% +47.0 pts on 0-100 axis
Baseline accuracy: 26.0% Omar/RCC accuracy: 73.0%

Board-depth replay. Board-depth replay row; read with sampling method, raw outputs, and route logs. Score context: 26.0% baseline to 73.0% Omar/RCC (+47.0 pts); stored uplift +180.77%. RoboBench embodied QA v5 100-sample local revalidation artifact; gpt-5.2; Darwin/Hinton v5 visual sequence + DAG exact DSL lock; not robot control and not an official leaderboard claim.

+180.77% +47.0 pts on 0-100 axis
Context 100 BYOK/local v5
BOARD-DEPTH REPLAY n=100 local cases / robotics policy smoke ManiSkill Robotics
+13.38% stored board uplift
Relative uplift scale Stored uplift: +13.38%

Board-depth replay. ManiSkill Robotics artifact proof: 100 local cases, random baseline total reward 85.4592 versus Omar/RCC sampled governance router 96.8905, +11.4313 reward / +13.38% relative. Not an official ManiSkill leaderboard claim.

+13.38% stored board uplift
Context 100 local cases

Lane taxonomy

Artifact-backed board context uses archived evidence. Live Reproduction Attempt runs selected samples through the current harness. Diagnostic Live Replay is for debugging current behavior. Board Evidence Context is reference evidence only.

Direct comparability requires matching samples, model, grader, prompt wrapper, and routing harness.

Board Reproduction

Historical Benchmark Evidence is read-only. Live Replay Demo is not historical board reproduction. Board Reproduction is enabled only when exact historical assets are wired by benchmark family.

Board families: 14. Site-enabled reproduction: MuSR, SimpleQA Verified / VSF. Assets found, pending site wiring: BBEH, BBH, Facts Grounding, HealthBench main, HealthBench hard, HealthBench consensus, HLE / HLE-Verified, SimpleQA.

BYOK review families: AIME 120, BBEH, BBH, Facts Grounding, GPQA, HealthBench main, HealthBench hard, HealthBench consensus, HLE / HLE-Verified, HorizonMath, MMLU-Pro, SimpleQA.

BYOK run buttons appear only for families with an enabled board runner.

BenchmarknBoard upliftModel / temperatureRequired assetsSite statusBYOK review noteSource refs
AIME 120120+100.0%not_found_in_exact_aime_board_artifact / not_explicitly_recorded
Asset parity
{
  "sample_ids": "no",
  "prompts": "no",
  "gold": "no",
  "grader": "no",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "no"
}
BYOK reviewLatest board exposes AIME summary/prior only. Exact AIME 120 raw run, sample IDs, prompts, grader, and script were not found in the 2026-04-25 proof bundle. Older AIME archive candidates exist under rcc_publication_package but are RCC v4/v16/gpt-4o-era assets, not this Ben-confirmed gpt-5.2 board.
Review items: DATASET_MISSING, PROMPTS_MISSING, GOLD_MISSING, GRADER_MISSING, MODEL_CONFIG_MISSING, RAW_LOGS_MISSING
Source refs
[
  "AIME_120__darwin_benchmark_priors.json",
  "aime_hard.jsonl",
  "aime"
]
BBEH100+183.3%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "no",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "no"
}
BYOK reviewScript and scored result are present, but local proof artifact stores task/gold/answer correctness, not full prompt/input text or raw model outputs. Public site runner is pending site wiring to the historical script.
Review items: PROMPTS_MISSING, MODEL_CONFIG_MISSING, RAW_LOGS_MISSING
Source refs
[
  "BBEH__bbeh_routed_bench_results.json",
  "BBEH__bbeh_routed_bench_results_2.json"
]
BBH50+14.3%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "yes",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "no"
}
BYOK reviewFrozen IDs/contract/script are present, but public site runner is pending site wiring to execute the historical BBH frozen harness yet.
Review items: MODEL_CONFIG_MISSING, RAW_LOGS_MISSING
Source refs
[
  "bbh_frozen_ids.json",
  "bbh_frozen_contract.json",
  "BBH__bbh_official_current_results.json"
]
Facts Grounding100+2.1%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "yes",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "yes"
}
BYOK reviewExact result/script assets are present, but the public site does not yet execute the historical Facts Grounding harness.
Review items: MODEL_CONFIG_MISSING
Source refs
[
  "Facts_Grounding__deepmind_facts_grounding_bench_results.json"
]
GPQA50+13.2%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "no",
  "prompts": "no",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "no"
}
BYOK reviewStored GPQA board artifact is aggregate-only: scores and metrics exist, but original per-sample prompts, IDs, and raw outputs are not stored in the proof JSON.
Review items: DATASET_MISSING, PROMPTS_MISSING, MODEL_CONFIG_MISSING, RAW_LOGS_MISSING
Source refs
[
  "GPQA__gpqa_routed_bench_results.json"
]
HealthBench main50+6.2%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "yes",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "no"
}
BYOK reviewLocal HealthBench data/script/result are present, but stored proof result is aggregate metric output and public site runner is pending site wiring to historical HealthBench execution.
Review items: MODEL_CONFIG_MISSING, RAW_LOGS_MISSING
Source refs
[
  "healthbench.jsonl",
  "HealthBench_main__healthbench_routing_bench_results.json"
]
HealthBench hardstored+11.9%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "yes",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "no"
}
BYOK reviewHealthBench hard local data exists, but latest proof bundle stores only Darwin benchmark priors for this board row, not the raw hard-subset execution result.
Review items: MODEL_CONFIG_MISSING, RAW_LOGS_MISSING
Source refs
[
  "healthbench_hard.jsonl",
  "HealthBench_hard__darwin_benchmark_priors.json"
]
HealthBench consensusstored+3.1%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "yes",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "no"
}
BYOK reviewHealthBench consensus local data exists, but latest proof bundle stores only Darwin benchmark priors for this board row, not the raw consensus-subset execution result.
Review items: MODEL_CONFIG_MISSING, RAW_LOGS_MISSING
Source refs
[
  "healthbench_consensus.jsonl",
  "HealthBench_consensus__darwin_benchmark_priors.json"
]
HLE / HLE-Verified50+21.1%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "yes",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "yes"
}
BYOK reviewExact HLE result/script assets are present, but the public site does not yet execute the historical HLE harness.
Review items: MODEL_CONFIG_MISSING
Source refs
[
  "HLE___HLE-Verified__hle_verified_routed_bench_results.json",
  "HLE___HLE-Verified__hle_verified_routed_bench_results_2.json"
]
HorizonMath50+45.5%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "no",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "yes"
}
BYOK reviewHorizonMath IDs/targets/raw outputs exist, but local result cache does not include full prompt/problem statements and public site runner is pending site wiring to the historical script.
Review items: PROMPTS_MISSING, MODEL_CONFIG_MISSING
Source refs
[
  "HorizonMath__horizonmath_routed_bench_results.json"
]
MMLU-Pro50+3.4%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "no",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "no"
}
BYOK reviewStored MMLU-Pro summary says n=50, but result JSON only stores partial sample rows and no full prompt text; public site runner is pending site wiring to historical MMLU-Pro script.
Review items: PROMPTS_MISSING, MODEL_CONFIG_MISSING, RAW_LOGS_MISSING
Source refs
[
  "MMLU-Pro__mmlu_pro_routed_bench_results.json"
]
MuSR100 BYOK live+4.17%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "yes",
  "gold": "yes",
  "grader": "yes",
  "model_config": "yes",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "yes"
}
enabledCurrent MuSR BYOK live package: n=100, baseline 0.72, Omar/RCC 0.75, +3.00 points / +4.17% relative, decision LIVE_SCORE_RECORDED. Exactness status remains READY_RECONSTRUCTED.
Review items: EXACT_HISTORICAL_ARTIFACT_NOT_CLAIMED
Source refs
[
  "samples/processed/musr_style_external_official_100.normalized.jsonl",
  "results.csv",
  "results.json",
  "raw_outputs.jsonl",
  "route_log.jsonl",
  "cost_summary.txt",
  "repro_manifest.json"
]
SimpleQA100+16.0%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "yes",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "yes"
}
BYOK reviewExact SimpleQA result/script assets are present, but the public site does not yet execute the historical SimpleQA harness.
Review items: MODEL_CONFIG_MISSING
Source refs
[
  "SimpleQA__official_simpleqa_routing_bench_results.json"
]
SimpleQA Verified / VSF100 clean harness+120.4%gpt-5.2 / not_explicitly_recorded
Asset parity
{
  "sample_ids": "yes",
  "prompts": "yes",
  "gold": "yes",
  "grader": "yes",
  "model_config": "no",
  "darwin": "yes",
  "hinton": "yes",
  "lua": "yes",
  "rcc_wrapper": "yes",
  "raw_outputs": "yes"
}
enabledHistorical artifact replay is wired from proof_manifests/vsf_historical_reproduction_results.json. Live replay remains separate.
Review items: none
Source refs
[
  "SimpleQA_Verified___VSF__deepmind_simpleqa_verified_routing_bench_results.json",
  "SimpleQA_Verified___VSF__deepmind_simpleqa_verified_routing_bench_results_2.json"
]