reasoning-analysis-results

Private dataset repo for reasoning-analysis result bundles.

Latest corrected run

• Subdir: reasoning-analysis-20260321-092638
• Ground truth: imo_answerbench_clean.csv
• Validation: PASS
• Manifest verification: PASS

Recommended defaults from the corrected exact analysis

• cheap_fast: high_match[d=5,n=2]
• fast_pure_high: high_match[d=7,n=3]
• normal_default: high_match[d=12,n=4]
• accuracy_first: high_match[d=15,n=5]

Start here inside the latest run folder

• FINAL_CORRECTED_INDEX.md
• FINAL_PLAIN_ENGLISH_GUIDE.md
• EXACT_DEPLOYMENT_DEFAULTS.md
• EXACT_OPERATING_POINTS.md
• EXACT_CAPPED_HIGH_SWEEP_MATCH_FRONTIER_SUMMARY.md
• EXACT_CAPPED_HIGH_MATCH_FRONTIER_PRESETS.md
• EXACT_PURE_HIGH_CAPPED_GRADING.md
• EXACT_ADAPTIVE_MEDIUM_PROBE_ACCEPT_THEN_CAPPED_HIGH_VS_PURE_HIGH.md
• EXACT_ADAPTIVE_MEDIUM_PROBE_ACCEPT_OR_BUCKET_ROUTE_VS_PURE_HIGH.md
• EXACT_ADAPTIVE_MEDIUM_PROBE_ROUTE_VS_PURE_HIGH.md
• EXACT_ADAPTIVE_MEDIUM_PROBE_BUCKET_ROUTE_VS_PURE_HIGH.md
• EXACT_ADAPTIVE_ROUTE_FAMILY_COMPARISON.md
• EXACT_ADAPTIVE_FAMILY_SUMMARY.md
• EXACT_ACCEPT_FAMILY_COMPARISON.md
• EXACT_PARALLEL_MEDIUM_PROBE_CHOOSE_HIGH_RULE_VS_PURE_HIGH.md
• EXACT_PARALLEL_MEDIUM_ACCEPT_OR_CHOOSE_HIGH_RULE_VS_PURE_HIGH.md
• EXACT_PARALLEL_MEDIUM_ACCEPT_OR_CAPPED_HIGH_VS_PURE_HIGH.md
• EXACT_MEDIUM_CONSENSUS_EXPLAINED.md
• EXACT_MAJ8_VS_4MATCH.md
• FOCUSED_POLICY_COMPARISON.md
• ANY4MATCH_8H4M_ANATOMY.md
• BRANCH_PROFILE_COMPARISON.md
• PURE_HIGH_TRIGGER_DEPTH.md
• PURE_HIGH_PROBLEM_DELTAS.md
• PURE_HIGH_UNLOCKS.md
• PURE_HIGH_STABILITY.md
• PURE_HIGH_HARD_SLICE.md
• PURE_HIGH_PLURALITY_TRAPS.md
• PURE_HIGH_ORDER_TRAPS.md
• PURE_HIGH_ERROR_ATTRIBUTION.md
• TIME_MODEL_AUDIT.md
• TIMING_PATCH_IMPACT.md
• FILTERED_ALL_STRATEGIES_FRONTIER.md
• FRONTIER_VS_EXACT_DEFAULTS.md
• FINAL_BUNDLE_VALIDATION.md
• FINAL_BUNDLE_MANIFEST.md

This repo is now kept intentionally small:

• only the latest corrected run is retained
• old run folders are deleted
• superseded approximate files inside the latest run are deleted
• the kept files are the validated exact outputs, the cleaned filtered frontier, and a minimal raw CSV set needed to reproduce or inspect them

Important split:

• the exact pure-high and exact mixed subset files are the authoritative deployment analysis
• EXACT_MAJ8_VS_4MATCH.* explains the exact maj@8 versus 8 high -> 4-match tie behavior
• EXACT_CAPPED_HIGH_SWEEP_MATCH_FRONTIER.* gives the broader exact capped-token high_match sweep over the frontier-relevant (d,n) pairs
• EXACT_CAPPED_HIGH_MATCH_FRONTIER_PRESETS.* gives the named presets for that broader capped high_match sweep, keyed by the exact (d,n) pair
• EXACT_PURE_HIGH_CAPPED_GRADING.* recomputes the pure-high family comparison under the same boxed-digit extraction rule as the capped sweeps
• EXACT_ADAPTIVE_MEDIUM_PROBE_ACCEPT_THEN_CAPPED_HIGH.* sweeps a fully causal sequential family where medium may answer directly if it reaches a threshold, otherwise a capped high policy is run afterward
• EXACT_ADAPTIVE_MEDIUM_PROBE_ACCEPT_THEN_CAPPED_HIGH_VS_PURE_HIGH.* shows that this family owns the extreme low-latency end, but loses again by about 150s and does not win on actual-token budgets
• EXACT_ADAPTIVE_MEDIUM_PROBE_ACCEPT_OR_BUCKET_ROUTE_THEN_CAPPED_HIGH.* sweeps a fully causal sequential hybrid where medium may answer directly, but reject cases route to easy, mid, or hard capped high
• EXACT_ADAPTIVE_MEDIUM_PROBE_ACCEPT_OR_BUCKET_ROUTE_VS_PURE_HIGH.* shows that this hybrid improves over plain sequential accept-then-high on some fast wall-time budgets, but still loses to pure-high once budgets are practical
• EXACT_ADAPTIVE_MEDIUM_PROBE_ROUTE_THEN_CAPPED_HIGH.* sweeps a fully causal family where medium traces only choose which capped high policy to run
• EXACT_ADAPTIVE_MEDIUM_PROBE_ROUTE_VS_PURE_HIGH.* shows that this routing family has no wall-time wins over pure-high, but does have a small actual-token-budget niche
• EXACT_ADAPTIVE_MEDIUM_PROBE_BUCKET_ROUTE_THEN_CAPPED_HIGH.* extends that idea to easy/mid/hard routing buckets
• EXACT_ADAPTIVE_MEDIUM_PROBE_BUCKET_ROUTE_VS_PURE_HIGH.* shows that the 3-way bucketed family still has no wall-time wins, but slightly sharpens the same actual-token-budget niche
• EXACT_ADAPTIVE_ROUTE_FAMILY_COMPARISON.* compares the 2-way and 3-way adaptive routing families directly
• EXACT_ADAPTIVE_FAMILY_SUMMARY.* compares the sequential-accept, sequential accept-or-route, 2-way route, 3-way route, parallel accept-or-high, parallel accept-or-choose-rule, and parallel choose-rule adaptive families on one page
• EXACT_ACCEPT_FAMILY_COMPARISON.* compares the four families where medium directly answers: sequential accept-then-high, sequential accept-or-route, shared-batch parallel accept-or-high, and shared-batch parallel accept-or-choose-rule
• EXACT_PARALLEL_MEDIUM_PROBE_CHOOSE_HIGH_RULE.* tests a fully parallel family where medium and capped 15 high launch together and medium only chooses whether to use an easier or stricter stopping rule on those same high traces
• EXACT_PARALLEL_MEDIUM_PROBE_CHOOSE_HIGH_RULE_VS_PURE_HIGH.* shows that this parallel family is effectively tied with 15->5 on accuracy, but slower and more expensive
• EXACT_PARALLEL_MEDIUM_ACCEPT_OR_CHOOSE_HIGH_RULE.* tests the stronger shared-batch family where medium and capped 15 high launch together, medium may answer directly, and reject cases choose between easier and stricter stopping rules on those same high traces
• EXACT_PARALLEL_MEDIUM_ACCEPT_OR_CHOOSE_HIGH_RULE_VS_PURE_HIGH.* shows that this stronger shared-batch family is the closest adaptive challenger in the 150-180s wall-time region, but it still trails pure-high on practical budgets and on actual generated-token budgets
• EXACT_PARALLEL_MEDIUM_ACCEPT_OR_CAPPED_HIGH.* tests a fully parallel family where medium and capped 15 high launch together and the answer comes from whichever policy decides first
• EXACT_PARALLEL_MEDIUM_ACCEPT_OR_CAPPED_HIGH_VS_PURE_HIGH.* shows that this family gets onto the combined wall-time frontier at the very fast end, but still loses on actual-token budgets and does not replace pure-high at moderate budgets
• FOCUSED_POLICY_COMPARISON.* gives the discussed branch/token anatomy with actual generated-token accounting
• ANY4MATCH_8H4M_ANATOMY.* explains why 8 high + 4 medium -> any 4 match is fast but low-accuracy
• BRANCH_PROFILE_COMPARISON.* compares the branch-level quality/cost tradeoffs across the key policies
• PURE_HIGH_TRIGGER_DEPTH.* explains exactly why 12 high -> 4-match beats 8 high -> 4-match
• PURE_HIGH_PROBLEM_DELTAS.* shows which actual problems move as you step up the pure-high ladder
• PURE_HIGH_UNLOCKS.* shows which rung first gets each problem over 50%, 75%, and 90% reliability
• PURE_HIGH_STABILITY.* shows that most problems are nearly flat across the pure-high ladder, while a small slice has large, high-variance rung effects
• PURE_HIGH_HARD_SLICE.* explains what that small hard slice looks like in the raw 16-high answer sets
• PURE_HIGH_PLURALITY_TRAPS.* shows whether the correct answer is top, tied, runner-up, lower, or absent in the full 16-high pool
• PURE_HIGH_ORDER_TRAPS.* shows that later pure-high wins are mostly order traps even when the full 16-high pool already prefers the correct answer
• PURE_HIGH_ERROR_ATTRIBUTION.* shows, at the exact subset level, how much each pure-high rung's error is early wrong stopping versus a genuinely bad selected rollout pool
• TIME_MODEL_AUDIT.* proves that the lookup path matches explicit token-by-token batch simulation, and documents the batch-32 continuation derived from the CSV
• TIMING_PATCH_IMPACT.* shows which approximate sweep rows are still stale after that timing-model patch and provides the unaffected <=16 frontier slice
• FILTERED_ALL_STRATEGIES_FRONTIER.* is the cleaned approximate sweep frontier
• FRONTIER_VS_EXACT_DEFAULTS.* is the one-page bridge between those two views
• the requested 8m+8h, stop at 8m+1h or 4h row is included there as mh_m8_h1_or_h4 (8,8,0)