Skip to content

IBM Phase 2 Preregistered Run Manifest — 2026-05-05

This manifest is the pre-submission contract for the next IBM hardware run. It is intentionally narrow: it permits a Phase 2 DLA parity expansion only, not a broad quantum-advantage claim and not any frontier or live-loop claim.

Submission status

Field Value
Status Preregistered; not submitted
Prepared date 2026-05-05
Target script scripts/phase2_full_campaign_ibm.py
Primary backend ibm_kingston
Replication backend ibm_marrakesh only if credit window permits
Available QPU budget assumption Approximately 120 minutes remaining
Required account state IBM promotional or credit window active
Submission guard --confirm-promo-active plus explicit human approval
Live-ready reduced command PYTHONDONTWRITEBYTECODE=1 /home/anulum/.local/bin/python scripts/phase2_full_campaign_ibm.py --confirm-promo-active --backend ibm_kingston --skip B C D E F --max-live-depth 1100
Dry-run command PYTHONDONTWRITEBYTECODE=1 /home/anulum/.local/bin/python scripts/phase2_full_campaign_ibm.py --dry-run --backend ibm_kingston --skip B C D E F
Dry-run result Passed on 2026-05-05
2026-05-05 live attempt Aborted and cancelled after hardware transpilation drift
Cancelled job ibm-run-ca8b9612732b84dc, IBM status CANCELLED, reported usage 0 quantum seconds

No IBM job may be submitted from this manifest until the promotional or credit window is live and the exact command is approved immediately before execution.

Scientific objective

Primary objective: test whether the Phase 1 DLA parity asymmetry observed at n=4 on ibm_kingston reproduces with higher statistics and extends across larger system sizes under the same raw-count evidence standard.

The promotable claim, if the run succeeds, is limited to:

  • DLA parity leakage asymmetry as a function of Trotter depth and system size.
  • Backend-specific Heron r2 hardware evidence, not device-independent proof.
  • Raw-count-backed statistics only.

The manifest does not permit claims about broad quantum advantage, frontier scaling, multi-QPU validation, or live-loop IBM execution.

Minimum live circuit plan

The first live run should spend only the QPU time needed to validate the central claim: high-statistics n=4 replication plus a same-run readout baseline. Scaling and mitigation sweeps are deferred until the primary raw-count dataset is retrieved and reviewed.

Block Experiment n Depths / preps Sectors Reps Shots Circuits Live status
A High-statistics DLA parity 4 2,4,6,8,10,14,20,30,40,50 even, odd 30 4096 600 approved for the first live command
G Readout baseline 4, 6, 8 000.., 111.., alternating states baseline 1 8192 12 approved for the first live command
Total Reduced primary run 4, 6, 8 preregistered above mixed mixed mixed 612 QPU-minimised validation run

The reduced live script estimate is approximately 5.6 minutes at 0.55 seconds per circuit. The estimate is not a guarantee; it is a pre-run budget guard. Wall-clock, queue, and backend scheduling behaviour are not promoted evidence.

Deferred circuit plan

The full campaign inventory was dry-run once and remains documented, but blocks B-F are not approved for the first live run. They require separate approval after block A+G counts are retrieved and reviewed.

Block Experiment n Depths / preps Sectors Reps Shots Circuits
B Scaling 6 4,8,14,20 even, odd 20 4096 160
C Scaling 8 4,8,14,20 even, odd 15 4096 120
D Scaling 10 4,8,14 even, odd 12 4096 72
E Scaling 12 4,8,14 even, odd 8 4096 48
F GUESS calibration 4 4,8,14; noise scales 1,3,5 even, odd 10 4096 180
Deferred total Follow-up only 4–12 preregistered above mixed mixed mixed 580

The full campaign inventory is 1192 circuits. It must not be submitted as the first live command under the current 120-minute remaining-budget constraint.

Dry-run transpilation gate

Reduced dry-run timestamp: 2026-05-05T120231Z.

Metric Value
Circuit count 612
Estimated QPU cost at 0.55 s/circuit 5.6 minutes
ISA depth minimum 1
ISA depth maximum 605
ISA depth mean 221.4
Gate-count minimum 4
Gate-count maximum 1503
Gate-count mean 544.3

Decision: the dry-run transpilation budget gate passed. It authorises readiness only; it does not authorise QPU submission.

Aborted live attempt — 2026-05-05

A first live reduced A+G attempt was started on 2026-05-05 after IBM readiness approval. The live hardware transpilation path produced deeper circuits than the reduced simulator dry-run manifest:

Source Max depth
Reduced dry-run manifest 605
Live hardware transpilation log 1014

This triggered the manifest drift abort rule. The submitted main-batch job ibm-run-ca8b9612732b84dc was immediately cancelled.

IBM job metadata after cancellation:

Field Value
Status CANCELLED
Created UTC 2026-05-05T12:06:37.041668Z
Running UTC 2026-05-05T12:06:38.800236Z
Finished UTC 2026-05-05T12:07:19.703948Z
Reported quantum seconds 0
Reported usage seconds 0

Decision: the cancelled job is quarantined execution evidence only. It is not raw-count evidence, not validation, and not promotable.

Abort criteria

Abort before submission if any condition below holds:

Gate Abort condition
Credit window Promotional or credit allocation is not live.
Backend Target backend is unavailable, paused, or replaced by an unreviewed backend.
Script drift scripts/phase2_full_campaign_ibm.py differs from the committed manifest assumptions.
Dry-run Dry-run fails, or max ISA depth or circuit count changes without manifest update.
Evidence path Raw-count output directory, job log, and post-run review path are not writable.
Approval Human approval is absent in the current turn.

Abort after submission if any condition below holds:

Gate Abort condition
Empty counts Any circuit returns zero total counts.
Missing job identifier Any retrieved count set cannot be bound to an IBM job identifier.
Partial run A batch fails and the partial output cannot be isolated from promoted evidence.
Backend mismatch Result metadata names a backend other than the approved backend.
Readout baseline Readout baseline is missing or cannot be matched to the run.
Hardware transpilation drift Live hardware transpilation exceeds the manifest budget before usable counts are retrieved.

Evidence capture path

The submission script writes live results under:

  • <private-local-record>
  • <private-local-record>
  • results/ibm_runs/ through HardwareRunner

Before any public promotion, the raw counts must be copied into a reviewed, tracked data directory with a manifest equivalent to the Phase 1 dataset:

  • data/phase2_dla_parity/{run_files}.json
  • data/phase2_dla_parity/README.md
  • a SHA-256 integrity table
  • a reproducer that recomputes every promoted statistic from raw counts
  • a hardware-ledger row naming the exact artefacts

private local workspace/ output is execution evidence for review, not public proof by itself.

Statistical tests

The post-run reproducer must compute, from raw counts only:

Statistic Rule
Parity leakage Opposite-parity counts divided by total counts per circuit.
Per-point mean and SEM Grouped by (n, depth, sector, noise_scale) as applicable.
Per-depth asymmetry (leak_even - leak_odd) / leak_odd with uncertainty propagation.
Welch tests Even versus odd leakage for every preregistered depth and system size.
Combined evidence Fisher combined statistic per system size and for the preregistered primary block.
Readout baseline State-retention table for every baseline prep; used as a systematic bound.
GUESS calibration Noise-scale trend for block F only; never mixed silently with unmitigated rows.

Promotion thresholds:

  • Primary n=4 block must reproduce the Phase 1 sign at the preregistered middle-depth region.
  • The combined n=4 Fisher test must be significant at p < 0.001.
  • Larger-n blocks are exploratory unless their raw-count reproducer and multiplicity handling are reviewed separately.
  • Negative or null results must remain in the ledger.

Submission commands

Reduced dry-run, already passed:

PYTHONDONTWRITEBYTECODE=1 /home/anulum/.local/bin/python scripts/phase2_full_campaign_ibm.py --dry-run --backend ibm_kingston --skip B C D E F

Reduced live submission now includes a pre-submit hardware-backend transpilation check. The accepted live depth budget is 1100, above the observed 1014 depth from the cancelled attempt and still below a manifest-controlled guard:

PYTHONDONTWRITEBYTECODE=1 /home/anulum/.local/bin/python scripts/phase2_full_campaign_ibm.py --confirm-promo-active --backend ibm_kingston --skip B C D E F --max-live-depth 1100

Deferred scaling/GUESS follow-up, blocked until the reduced primary run is retrieved and reviewed:

PYTHONDONTWRITEBYTECODE=1 /home/anulum/.local/bin/python scripts/phase2_full_campaign_ibm.py --confirm-promo-active --backend ibm_kingston --skip A G

Current decision

Prepared for IBM run approval gate: yes.

Prepared to submit IBM jobs without a live credit window and explicit current approval: no.

Prepared to promote new hardware validation immediately after submission: no. Promotion requires raw-count retrieval, integrity review, a committed reproducer, and an updated hardware ledger row.