retag: drop stale pages/

pages/claim-1/page.md

@@ -1,74 +0,0 @@
-# Certificate-Guided Pruning (CGP) maintains an explicit active set A_t of candidate optima using confidence-adjusted Lipschitz envelopes, certifying with high probability that any point outside A_t is suboptimal (Section 3, Algorithm 1).
-
-## Verdict
-
-**COMPARABLE SUPPORT**
-
-## Method and evidence
-
-The CPU verifier implements the paper-specific construction for this claim and
-records the following independently computed result:
-
-```json
-{
-  "method": "Algorithm-1 confidence-adjusted Lipschitz envelopes on a 2001-point domain",
-  "runs": 12,
-  "false_optimum_eliminations": 0,
-  "mean_final_active_fraction": 0.21809928369148757,
-  "example_trajectory": [
-    {
-      "samples": 8,
-      "active_fraction": 0.27386306846576713,
-      "active_points_grid": 548,
-      "outside_points_certified": 1453,
-      "false_eliminated_optimum": false,
-      "simple_regret": 0.0050000000000000044
-    },
-    {
-      "samples": 16,
-      "active_fraction": 0.23588205897051473,
-      "active_points_grid": 472,
-      "outside_points_certified": 1529,
-      "false_eliminated_optimum": false,
-      "simple_regret": 0.0050000000000000044
-    },
-    {
-      "samples": 32,
-      "active_fraction": 0.23588205897051473,
-      "active_points_grid": 472,
-      "outside_points_certified": 1529,
-      "false_eliminated_optimum": false,
-      "simple_regret": 0.0030000000000000027
-    },
-    {
-      "samples": 64,
-      "active_fraction": 0.16591704147926037,
-      "active_points_grid": 332,
-      "outside_points_certified": 1669,
-      "false_eliminated_optimum": false,
-      "simple_regret": 0.0010000000000000009
-    },
-    {
-      "samples": 96,
-      "active_fraction": 0.15292353823088456,
-      "active_points_grid": 306,
-      "outside_points_certified": 1695,
-      "false_eliminated_optimum": false,
-      "simple_regret": 0.0010000000000000009
-    }
-  ],
-  "verdict": "comparable finite-grid support; not a proof of the high-probability certificate"
-}
-```
-
-## Scope boundary
-
-comparable finite-grid support; not a proof of the high-probability certificate
-
-## Provenance
-
-- Command: `python -u scripts/verify.py`
-- Result: `results/certificate_guided_pruning_results.json`
-- Result SHA-256: `4f75212b07ab5334b186cc4a667527bdb941ad522d226ca26f4ac61e6a27e7e6`
-- PDF SHA-256: `dac38c63ac690606fa90c74c5bb0b6d708af33a7b49777eaf7bb3f60f4be4465`
-- Seed `20260729`; CPU only; cost `$0.00`

pages/claim-2/page.md

@@ -1,72 +0,0 @@
-# Under a margin condition with near-optimality dimension α (Assumption 2.3), the Shrinkage Theorem bounds the active set volume as Vol(A_t) ≤ C·(2(β_t + Lη_t) + γ_t)^(d−α) (Theorem 4.6).
-
-## Verdict
-
-**SCALED SUPPORT**
-
-## Method and evidence
-
-The CPU verifier implements the paper-specific construction for this claim and
-records the following independently computed result:
-
-```json
-{
-  "trajectory": [
-    {
-      "samples": 8,
-      "active_fraction": 0.2644338915271182,
-      "active_points_grid": 1058,
-      "outside_points_certified": 2943,
-      "false_eliminated_optimum": false,
-      "simple_regret": 0.0050000000000000044
-    },
-    {
-      "samples": 16,
-      "active_fraction": 0.2644338915271182,
-      "active_points_grid": 1058,
-      "outside_points_certified": 2943,
-      "false_eliminated_optimum": false,
-      "simple_regret": 0.0050000000000000044
-    },
-    {
-      "samples": 32,
-      "active_fraction": 0.2644338915271182,
-      "active_points_grid": 1058,
-      "outside_points_certified": 2943,
-      "false_eliminated_optimum": false,
-      "simple_regret": 0.0050000000000000044
-    },
-    {
-      "samples": 64,
-      "active_fraction": 0.255936015996001,
-      "active_points_grid": 1024,
-      "outside_points_certified": 2977,
-      "false_eliminated_optimum": false,
-      "simple_regret": 0.00275000000000003
-    },
-    {
-      "samples": 256,
-      "active_fraction": 0.1937015746063484,
-      "active_points_grid": 775,
-      "outside_points_certified": 3226,
-      "false_eliminated_optimum": false,
-      "simple_regret": 0.0002500000000000835
-    }
-  ],
-  "active_volume_loglog_slope_vs_samples": -0.11682190341959578,
-  "monotone_shrinkage": true,
-  "verdict": "scaled support for shrinkage; the theorem's universal volume bound remains analytic"
-}
-```
-
-## Scope boundary
-
-scaled support for shrinkage; the theorem's universal volume bound remains analytic
-
-## Provenance
-
-- Command: `python -u scripts/verify.py`
-- Result: `results/certificate_guided_pruning_results.json`
-- Result SHA-256: `4f75212b07ab5334b186cc4a667527bdb941ad522d226ca26f4ac61e6a27e7e6`
-- PDF SHA-256: `dac38c63ac690606fa90c74c5bb0b6d708af33a7b49777eaf7bb3f60f4be4465`
-- Seed `20260729`; CPU only; cost `$0.00`

pages/claim-3/page.md

@@ -1,67 +0,0 @@
-# CGP achieves ε-optimality with probability at least 1−δ using T = Õ(L^d ε^{-(2+α)} log(1/δ)) samples, improving on the worst-case Õ(ε^{-(2+d)}) rate whenever α < d (Theorem 4.8).
-
-## Verdict
-
-**COMPARABLE SCALING**
-
-## Method and evidence
-
-The CPU verifier implements the paper-specific construction for this claim and
-records the following independently computed result:
-
-```json
-{
-  "rows": [
-    {
-      "epsilon": 0.2,
-      "grid_points": 11,
-      "repeats_per_point": 877,
-      "samples": 9647,
-      "simple_regret": 0.030000000000000027
-    },
-    {
-      "epsilon": 0.14,
-      "grid_points": 16,
-      "repeats_per_point": 1789,
-      "samples": 28624,
-      "simple_regret": 0.030000000000000027
-    },
-    {
-      "epsilon": 0.1,
-      "grid_points": 21,
-      "repeats_per_point": 3506,
-      "samples": 73626,
-      "simple_regret": 0.020000000000000018
-    },
-    {
-      "epsilon": 0.07,
-      "grid_points": 30,
-      "repeats_per_point": 7155,
-      "samples": 214650,
-      "simple_regret": 0.009310344827586192
-    },
-    {
-      "epsilon": 0.05,
-      "grid_points": 41,
-      "repeats_per_point": 14023,
-      "samples": 574943,
-      "simple_regret": 0.0050000000000000044
-    }
-  ],
-  "sample_complexity_loglog_slope_vs_epsilon": -2.940470877127776,
-  "expected_cusp_alpha0_exponent": -2.0,
-  "verdict": "comparable constructive scaling; not an exact implementation of every CGP refinement rule"
-}
-```
-
-## Scope boundary
-
-comparable constructive scaling; not an exact implementation of every CGP refinement rule
-
-## Provenance
-
-- Command: `python -u scripts/verify.py`
-- Result: `results/certificate_guided_pruning_results.json`
-- Result SHA-256: `4f75212b07ab5334b186cc4a667527bdb941ad522d226ca26f4ac61e6a27e7e6`
-- PDF SHA-256: `dac38c63ac690606fa90c74c5bb0b6d708af33a7b49777eaf7bb3f60f4be4465`
-- Seed `20260729`; CPU only; cost `$0.00`

pages/claim-4/page.md

@@ -1,30 +0,0 @@
-# A matching lower bound shows any algorithm requires Ω(ε^{-(2+α)}) samples under the same margin condition, establishing CGP's minimax sample-complexity optimality (Theorem 4.9).
-
-## Verdict
-
-**NOT TESTED / INCONCLUSIVE**
-
-## Method and evidence
-
-The CPU verifier implements the paper-specific construction for this claim and
-records the following independently computed result:
-
-```json
-{
-  "tested": false,
-  "reason": "A finite simulation cannot establish an algorithm-independent minimax lower bound. The local claim-3 hard-family sweep is evidence about one executable procedure only.",
-  "verdict": "not tested / inconclusive"
-}
-```
-
-## Scope boundary
-
-not tested / inconclusive
-
-## Provenance
-
-- Command: `python -u scripts/verify.py`
-- Result: `results/certificate_guided_pruning_results.json`
-- Result SHA-256: `4f75212b07ab5334b186cc4a667527bdb941ad522d226ca26f4ac61e6a27e7e6`
-- PDF SHA-256: `dac38c63ac690606fa90c74c5bb0b6d708af33a7b49777eaf7bb3f60f4be4465`
-- Seed `20260729`; CPU only; cost `$0.00`

pages/claim-5/page.md

@@ -1,61 +0,0 @@
-# CGP-Adaptive learns the Lipschitz constant L online via a doubling scheme, adding only an O(log T) multiplicative overhead to the sample complexity (Theorem 5.1, Section 5).
-
-## Verdict
-
-**EXACT MECHANISM COUNT**
-
-## Method and evidence
-
-The CPU verifier implements the paper-specific construction for this claim and
-records the following independently computed result:
-
-```json
-{
-  "rows": [
-    {
-      "initial_L": 0.0078125,
-      "doublings": 7,
-      "final_L": 1.0,
-      "ceil_log2_ratio": 7
-    },
-    {
-      "initial_L": 0.03125,
-      "doublings": 5,
-      "final_L": 1.0,
-      "ceil_log2_ratio": 5
-    },
-    {
-      "initial_L": 0.125,
-      "doublings": 3,
-      "final_L": 1.0,
-      "ceil_log2_ratio": 3
-    },
-    {
-      "initial_L": 0.5,
-      "doublings": 1,
-      "final_L": 1.0,
-      "ceil_log2_ratio": 1
-    },
-    {
-      "initial_L": 1.0,
-      "doublings": 0,
-      "final_L": 1.0,
-      "ceil_log2_ratio": 0
-    }
-  ],
-  "all_doubling_counts_match": true,
-  "verdict": "exact doubling-scheme count; downstream stochastic sample overhead not fully replicated"
-}
-```
-
-## Scope boundary
-
-exact doubling-scheme count; downstream stochastic sample overhead not fully replicated
-
-## Provenance
-
-- Command: `python -u scripts/verify.py`
-- Result: `results/certificate_guided_pruning_results.json`
-- Result SHA-256: `4f75212b07ab5334b186cc4a667527bdb941ad522d226ca26f4ac61e6a27e7e6`
-- PDF SHA-256: `dac38c63ac690606fa90c74c5bb0b6d708af33a7b49777eaf7bb3f60f4be4465`
-- Seed `20260729`; CPU only; cost `$0.00`

pages/claim-6/page.md

@@ -1,33 +0,0 @@
-# CGP-TR, a trust-region variant, scales to dimension d > 50 via certified restarts that provably never falsely eliminate the region containing the true optimizer x* (Theorem 6.1, Section 6).
-
-## Verdict
-
-**CONSTRUCTED SAFETY CHECK**
-
-## Method and evidence
-
-The CPU verifier implements the paper-specific construction for this claim and
-records the following independently computed result:
-
-```json
-{
-  "dimension": 64,
-  "regions": 5,
-  "regions_containing_optimum": 0,
-  "containing_regions_eliminated": 0,
-  "suboptimal_regions_eliminated": 5,
-  "verdict": "constructed safety invariant at d=64; not the paper's full benchmark suite"
-}
-```
-
-## Scope boundary
-
-constructed safety invariant at d=64; not the paper's full benchmark suite
-
-## Provenance
-
-- Command: `python -u scripts/verify.py`
-- Result: `results/certificate_guided_pruning_results.json`
-- Result SHA-256: `4f75212b07ab5334b186cc4a667527bdb941ad522d226ca26f4ac61e6a27e7e6`
-- PDF SHA-256: `dac38c63ac690606fa90c74c5bb0b6d708af33a7b49777eaf7bb3f60f4be4465`
-- Seed `20260729`; CPU only; cost `$0.00`

pages/conclusion/page.md

@@ -1,5 +0,0 @@
-# Conclusion
-
-The executable checks support the algorithmic mechanisms on bounded finite
-instances. They do not replace universal high-probability or minimax proofs, and
-the named high-dimensional benchmark suite was not run. Result SHA-256: `4f75212b07ab5334b186cc4a667527bdb941ad522d226ca26f4ac61e6a27e7e6`.

pages/executive-summary/page.md

@@ -1,8 +0,0 @@
-# Executive summary
-
-This local-only, zero-cost CPU reproduction exercises CGP's envelope pruning,
-active-volume shrinkage, epsilon/sample scaling, adaptive-L doubling, and
-high-dimensional restart safety on constructed Lipschitz instances. The minimax
-lower bound is explicitly not claimed from simulation.
-
-Result SHA-256: `4f75212b07ab5334b186cc4a667527bdb941ad522d226ca26f4ac61e6a27e7e6`.

pages/index.md

@@ -1,14 +0,0 @@
-# Reproduction: Certificate-Guided Pruning for Stochastic Lipschitz Optimization
-
-## Pages
-
-| Page |
-|---|
-| [Executive summary](#/executive-summary) |
-| [Certificate-Guided Pruning (CGP) maintains an explicit active set A_t of candidate optima using confidence-adjusted Lipschitz envelopes, certifying with high probability that any point outside A_t is suboptimal (Section 3, Algorithm 1).](#/claim-1) |
-| [Under a margin condition with near-optimality dimension α (Assumption 2.3), the Shrinkage Theorem bounds the active set volume as Vol(A_t) ≤ C·(2(β_t + Lη_t) + γ_t)^(d−α) (Theorem 4.6).](#/claim-2) |
-| [CGP achieves ε-optimality with probability at least 1−δ using T = Õ(L^d ε^{-(2+α)} log(1/δ)) samples, improving on the worst-case Õ(ε^{-(2+d)}) rate whenever α < d (Theorem 4.8).](#/claim-3) |
-| [A matching lower bound shows any algorithm requires Ω(ε^{-(2+α)}) samples under the same margin condition, establishing CGP's minimax sample-complexity optimality (Theorem 4.9).](#/claim-4) |
-| [CGP-Adaptive learns the Lipschitz constant L online via a doubling scheme, adding only an O(log T) multiplicative overhead to the sample complexity (Theorem 5.1, Section 5).](#/claim-5) |
-| [CGP-TR, a trust-region variant, scales to dimension d > 50 via certified restarts that provably never falsely eliminate the region containing the true optimizer x* (Theorem 6.1, Section 6).](#/claim-6) |
-| [Conclusion](#/conclusion) |