Paper-to-code study, 2026-10-01. Supersedes the earlier rho-recovery recommendation. Main result The largest code-derived changes are (1) the actual three-snapshot power reference and (2) the arithmetic confirmation gate. Their effects interact. The strongest third change for the original continuously driven benchmark is the code's smaller upward lambda gain. Idle accumulation and smoothing are separate changes, not hidden inside the gate. No tested three-change combination closes all safety, idle and roughness gaps. Original-seed results Method Err W Err N U5 U6 U7 U8 Rough J Code .0134 .0230 1.6 0.0 2.8 3.9 3.15 .0203 Paper .0165 .0739 3.4 0.0 34.8 25.1 11.37 .0610 Reference only .0330 .0996 3.7 0.0 7.7 2.7 14.61 .0698 Reference + arithmetic gate .0155 .0265 3.7 0.0 8.2 2.2 10.83 .0245 ... + code upward gain (3) .0152 .0263 1.5 0.0 3.6 1.0 10.73 .0223 ... + code guard (3) .0155 .0274 1.2 0.0 5.4 1.5 10.84 .0235 ... + activity factor (3) .0155 .0268 3.7 0.0 8.2 2.7 10.83 .0248 ... + smoothing (3) .0135 .0266 3.5 0.0 7.9 3.4 3.54 .0237 Here each '...' starts from reference + gate, not from the preceding row. U5--U8 are individual active-case unsafe percentages. Rough is wideband. J is retained for comparison but was not used to bury the safety tradeoff. In fixed-lambda narrowband Case 3 during 45--75 s, reference + gate estimates 0.30844 against true 0.34385. The code estimates 0.35724. The new estimate is conservative on this seed. This is not a uniform accuracy improvement. What each change actually copies 1. c_ref: sol_osi's median of three RAW output-power snapshots, taken at 0.5, 1.0, 1.5 s. The old med3 switch sampled at 0.875, 1.375, 1.875 s, because it counted from the first full estimator window. It was not the same operation. The paper's running-average reference is different too. 2. c_gate: sol_osi's rho_target=1+shape_weight*(rho_best-1), THEN logarithm. Existing paper_osi gate blended log(g), so it was not an exact port. This port uses the basic code fit floor 0.1 and coherent ramp slope 2. The code's extra level-dependent allowance is a separate switch. 3a. c_activity: alpha_g=A*eta_g in the accumulation update. This is necessary to match the code's memory behavior while idle. It does not prohibit power-based recovery when the observed ratio falls below one. 3b. c_up: positive margin errors are multiplied by 2*b_ref^2. With target 0.4 and no guard, that is 0.32. Negative errors retain full gain. 3c. smooth: existing code output filter with time constant 0.25 s. Slow power averaging is an experimental alternative, not a current code detail: actual sol_osi uses memory_s=0.25 for all three spectra. Experiment design - 22 atomic code-detail switches were screened individually on original seeds 11,12,21,22. No source constants were tuned. - Reference was the best single safety change. Every other available detail was added to it, producing 21 pairs. - Retained both the lowest-worst-unsafe pair (reference + upward gain) and the lowest-error pair (reference + gate), then tested 39 distinct triples. This is a two-branch sequential study, not an exhaustive global search. - Single/safety-anchor ranking is worst unsafe among Cases 5--8, followed by mean error and roughness. The error-best branch prevents conservative bias being silently treated as a success. The full tables retain both tradeoffs. - A shortlist was fixed before testing offsets 7000,8000,9000 (24 new cases). Source constants stayed fixed. Four/five-change runs and all-port replay are subsequent diagnostics, explicitly labelled as such. Unseen results (mean across the three offsets) Method Mean err U5 U6 U7 U8 Rough J Code .0236 1.3 0.5 11.6 0.3 3.19 .0271 Paper .0385 2.1 2.9 31.9 22.4 11.19 .0533 Reference + gate .0310 2.8 2.8 12.3 1.7 11.16 .0359 ... + upward gain .0307 1.2 0.5 12.1 0.1 11.13 .0342 ... + code guard .0303 0.9 0.0 8.9 1.7 11.09 .0332 ... + activity .0309 2.8 2.8 12.3 1.7 11.16 .0358 ... + smoothing .0301 3.3 3.0 12.4 1.7 3.50 .0352 Rho recovery + slow average .0281 2.0 2.6 8.6 0.0 11.31 .0314 The slower-power alternative remains visible, but has the old idle defect. Its good safety results are not dismissed on J alone. Do not hide the bad seed in those means. For Case 7, seed 7021: Code: 32.17% unsafe. Reference + gate + upward gain: 33.04%. Reference + gate + guard: 23.80%. Rho recovery + slow average: 13.04%. Thus even the code benchmark is not uniformly near 2% unsafe on unseen seeds. Idle tests A. Isolated memory test: seed remembered g=2, zero input, eight updates (1 s). Keep output-power evidence ABOVE its reference, then repeat BELOW it. Percent of initial LOG-gain remaining: Above reference Below reference Code 100 8.21 Old rho recovery 8.06 8.06 Reference + gate 35.96 8.06 Reference + gate + upward 35.96 8.06 Reference + gate + activity 100 8.06 Activity is a separate, necessary port. This is a controlled state test, not a claim that power evidence always survives real idle transitions. B. Closed-loop drive-idle-drive: same eight cases, same plants and noise, set u=0 during 55--70 s while damage remains. Code active-case unsafe percentages over the usual scoring interval are 13.9,11.4,12.4,13.8. In Case 7 during 60--70 s specifically, code is unsafe 98.75% of the time. Reference + gate + activity is unsafe 100% during that interval. The code's power-based forgetting remains active in idle. Matching this implementation does not establish the manuscript's unconditional idle memory claim. That claim needs qualification or a separate method change. Cost of additional details (original seeds, explicitly not three changes) Reference + gate + activity + upward gain (4): mean err .0209, unsafe [1.5,0,3.6,1.0], rough 10.73. Add output smoothing (5): mean err .0196, unsafe [1.7,0,3.7,1.1], rough 3.55. These do not fix the actual drive-idle-drive failure above. Faithfulness and reproduction All 22 source ports together reproduce code per-case unsafe percentages exactly on the original and all three new seed sets. The maximum difference in mean error over a seed set is 0.00000840. Defaults reproduce the saved as-written baseline exactly. All switches default off. Run from this folder through MATLAB MCP after making each script active: paper_code_bridge_study('single',1:24) paper_code_bridge_study('two',1:21) paper_code_bridge_study('three',1:39) paper_code_bridge_study('unseen',1:13) paper_code_bridge_study('idle',[1 2 3 6 7 8 9 10 11 12]) paper_code_bridge_idle_memory paper_code_bridge_study('extended',1:3) Main evidence: paper_code_bridge_study.mat (one result struct), paper_code_bridge_study.csv (aggregate comparisons), paper_code_bridge_seed_metrics.csv (every seed, every case's unsafe time), paper_code_bridge_idle_memory.mat (controlled idle check). Proposed source-matching equations: paper_code_bridge_equations.tex. No changes to main checkout main/, the manuscript, builder or reference code. No physical workflow, commits or new parameter search.