S1d IBM Policy-Direction Sweep Result¶
Date: 2026-05-20
Backend: ibm_kingston
Experiment ID: s1d_policy_direction_sweep_2026-05-20
Parent experiment: s1_dynamic_feedback_preregistration_2026-05-06
Artefacts¶
- Readiness:
data/s1_feedback_loop/s1d_xy_observable_readiness_ibm_kingston_20260520T134614Z.json - Readiness SHA-256:
ecbfc54979c86159ca40a05da2249bf7fb807963893f2a812b3a07fcf36258d7 - Raw counts:
data/s1_feedback_loop/s1d_xy_observable_raw_counts_ibm_kingston_20260520T134614Z.json - Raw counts SHA-256:
f26f3c442a2e91a215e47ec6bf0a515170a70cede3ce0ee4639ba1e7b121129d - Analysis:
data/s1_feedback_loop/s1d_xy_observable_analysis_ibm_kingston_20260520T134614Z.json - Analysis SHA-256:
6e6bc2f513f23633c72756016392b58cbf140d89c6f9cc7fdd07880c66b61c76
Purpose¶
S1d is a same-paper continuation of S1/S1b/S1c. It tests whether the observed S1c weakness is best explained by an incorrect feedback correction direction, excess gain, or a more fundamental limitation of the monitored-feedback policy on the selected backend/calibration window.
The sweep keeps the S1c one-round shallow body and direct XY observable family, but compares three preregistered policy variants:
| Variant | Rounds | Correction angle | Base gain | Purpose |
|---|---|---|---|---|
current_shallow_positive |
1 | 0.06 |
0.4 |
Repeat the completed S1c shallow positive-correction policy. |
polarity_flipped |
1 | -0.06 |
0.4 |
Test whether S1c's negative shift was a correction-polarity effect. |
weak_positive |
1 | 0.03 |
0.2 |
Test whether positive correction needs a lower-gain stability boundary. |
All policy variants passed readiness with maximum transpiled depth 237 and a
total estimated QPU budget of 72.0 seconds under the 120.0 second ceiling.
Execution Notes¶
The first approved submission produced 14 completed IBM jobs before the local
approval scheduler stopped because provider wall-clock wait time was reported
as qpu_seconds. The accounting bug was fixed so QPU budget accounting uses the
preregistered arm estimate while preserving provider wait time as
wall_time_s. The run was resumed without duplicating completed arms: 14 jobs
were recovered from IBM job lookup and the 10 missing arms were submitted.
Recovered jobs:
d86rm9h789is7390302gd86rmbas46sc73f7cufgd86rmd5g7okc73elhcsgd86rmeh789is739030a0d86rmg9789is739030bgd86rmhop0eas73dlcipgd86rmjlg7okc73elhd40d86rml2s46sc73f7curgd86rmmis46sc73f7cut0d86rmo0p0eas73dlcj10d86rmpop0eas73dlcj2gd86rmr8p0eas73dlcj50d86rmsp789is739030u0d86rmulg7okc73elhdk0
Resume-submitted jobs:
d86roris46sc73f7d1b0d86rq18p0eas73dlcmj0d86rqkp789is7390358gd86rqmh789is739035c0d86rqo5g7okc73elhhs0d86rqpp789is739035fgd86rqras46sc73f7d3igd86rqt1789is739035kgd86rquop0eas73dlcnigd86rr0gp0eas73dlcnl0
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 |
-0.0078125 |
-0.0136718750 |
-0.0605468750 |
0.1848958333 |
0.0257161458 |
0.0667317708 |
polarity_flipped |
-0.0130208333 |
-0.0065104167 |
0.0156250000 |
-0.0123697917 |
-0.0040690104 |
0.0118815104 |
weak_positive |
-0.0026041667 |
-0.0039062500 |
0.0065104167 |
-0.0201822917 |
-0.0050455729 |
0.0083007813 |
By mean signed feedback-control delta, the best variant is
current_shallow_positive. This ranking is driven by the YYI channel; it does
not by itself establish a stable feedback improvement.
Interpretation¶
S1d does not close the paper with a simple positive-feedback claim. It shows
that the direct-XY response is policy- and calibration-window-sensitive. The
completed S1c run moved all four direct-XY channels negative, while the S1d
repeat of the same shallow positive policy produced one large positive YYI
shift and three negative shifts. The polarity-flipped and weak-positive variants
are closer to zero and do not rescue the policy.
The conservative same-paper conclusion is therefore diagnostic: the tested
monitored-feedback law is not robustly beneficial at N = 3 system qubits on this
backend class, but the direct-XY sector exposes structured policy sensitivity
that is hidden by the original binary proxy. Any stronger claim requires a
predeclared statistical repeat or a redesigned policy law, not post hoc selection
of the favourable YYI channel.