Skip to content

S1e IBM Policy-Sweep Confirmatory Repeat

Date: 2026-05-20 Backend: ibm_kingston Experiment ID: s1e_policy_sweep_confirmatory_repeat_2026-05-20 Parent experiment: s1_dynamic_feedback_preregistration_2026-05-06

Artefacts

  • Readiness: data/s1_feedback_loop/s1e_xy_observable_readiness_ibm_kingston_20260520T140919Z.json
  • Readiness SHA-256: 80b78f3deafae9465e1f057be5fe20048a82a28330b69028cc481e829edb9fed
  • Raw counts: data/s1_feedback_loop/s1e_xy_observable_raw_counts_ibm_kingston_20260520T140919Z.json
  • Raw counts SHA-256: 99d97d60ce83949100ef52f6ead274776f49515990672d17fd91074b793a38ea
  • Analysis: data/s1_feedback_loop/s1e_xy_observable_analysis_ibm_kingston_20260520T140919Z.json
  • Analysis SHA-256: 4f7e106e2af8e42aa74c9c07d66032bc76a78a936c499370c56fcef8c32ad522

Purpose

S1e is the confirmatory statistical repeat for the S1d policy-direction sweep. It tests whether the large positive YYI response in the current_shallow_positive policy reproduces at higher repetitions, or whether it was a calibration-window fluctuation.

The repeat keeps the same one-round shallow dynamic body and the same three policy variants as S1d, but increases repetitions from 3 to 5.

Readiness

All three policy variants passed the readiness gate on ibm_kingston.

Variant Rounds Correction angle Base gain Max transpiled depth Estimated QPU seconds
current_shallow_positive 1 0.06 0.4 237 40.0
polarity_flipped 1 -0.06 0.4 237 40.0
weak_positive 1 0.03 0.2 237 40.0

Total estimated QPU seconds: 120.0 under the 130.0 second ceiling.

Execution

The 24 IBM jobs completed on ibm_kingston:

d86s140p0eas73dlcumg, d86s160p0eas73dlcuq0, d86s185g7okc73elhp30, d86s1cqs46sc73f7dam0, d86s1eh789is73903dtg, d86s1g9789is73903dvg, d86s1i2s46sc73f7dar0, d86s1jop0eas73dlcva0, d86s1n8p0eas73dlcveg, d86s1p8p0eas73dlcvh0, d86s1qp789is73903eg0, d86s1sqs46sc73f7db90, d86s1ugp0eas73dlcvqg, d86s20dg7okc73elhq0g, d86s220p0eas73dld020, d86s241789is73903es0, d86s25gp0eas73dld07g, d86s278p0eas73dld09g, d86s29h789is73903f3g, d86s2bas46sc73f7dbt0, d86s2d1789is73903fc0, d86s2ep789is73903fe0, d86s2ggp0eas73dld0og, d86s2i1789is73903fqg.

The first analysis file was generated before the s1e alias was routed through the grouped policy-sweep packager. The final raw-count and analysis artefacts regroup the same completed jobs without resubmission.

Results

Feedback minus matched open-loop control by sorted direct-XY observable row:

Variant IXX IYY XXI YYI Mean signed Mean absolute
current_shallow_positive -0.0035156250 0.0082031250 -0.0488281250 0.1824218750 0.0345703125 0.0607421875
polarity_flipped -0.0429687500 -0.0406250000 -0.0281250000 0.0167968750 -0.0237304688 0.0321289063
weak_positive -0.0074218750 0.0042968750 -0.0105468750 -0.0406250000 -0.0135742187 0.0157226563

Interpretation

S1e confirms that the S1d YYI response in current_shallow_positive is not a single-run artefact at the observed scale: S1d measured YYI = 0.1848958333, and S1e measured YYI = 0.1824218750.

The result still does not promote the full feedback law as robustly beneficial. The favourable response is channel-specific: current_shallow_positive remains negative in IXX and XXI, slightly positive in IYY, and strongly positive in YYI. The polarity-flipped and weak-positive policies do not reproduce the same mean-signed improvement. The safe claim is therefore a reproducible channel-specific policy sensitivity in the direct-XY sector, not backend-general feedback control.