Published November 12, 2026 | Version v4.0

Reading Why: Born-Measured Attention with Exact Word Evidence on Light and Qubits

  • 1. School of Integrated Innovative Technology, King Mongkut's Institute of Technology Ladkrabang
  • 2. Faculty of Science and Technology, Suan Dusit University

Description

The code, models, data and IBM job records behind the fourth arXiv version of the paper (arXiv:2611.05355v4). The paper's PDF is attached.

What is new since v3.2.1:

  • Compute-uncompute on ibm_kingston. Four new jobs (pair 3765 and pair 3763 without "aliens"; 580 circuits, 316 quantum seconds) read each attention Born weight as the probability of 0000. src/cu_analysis.py gives the answers, calibrations and evidence from the recorded counts; src/hw_cu_compare.py compares the two methods on the same Born weights. The device gave the model's answers; the evidence followed the exact one on one run (0.971) and not on the other (0.48).
  • What each device check can show. src/device_power.py, device_blocks.py, device_rows.py, device_errors.py, device_floor.py, device_rowsd.py, device_block2_nulls.py: uniform and random attention, block by block; the error per circuit against shot noise; nulls for the evidence check.
  • The shots the evidence needs. src/shot_bound.py checks the paper's proposition on 100 pairs (results/shot_bound.json holds our run); src/device_shot_floor.py compares it with the device.
  • Baselines. An order-free bag of embeddings (src/bag_baseline.py, 77.79 +- 0.26 %; seed-0 weights in results/) and its own evidence (src/bag_evidence.py); rankings that need no model and the evidence across seeds (src/esnli_baselines.py, src/esnli_seed_stability.py).
  • Optics and the coherent circuit. src/photon_draws.py (shot noise over 20 draws, photons per detector); src/coherent23.py (the 23-qubit circuit as counted, from |0...0>).

What the numbers say is in the README: accuracy does not set the models apart; the top-word deletion gain is the model's, not the evidence method's; exactness buys the complete sum and a better low end of the ranking. Account details are removed from every job record.

Notes

If you use this code or data, please cite the paper.

Files

stein-research/reading-why-v4.0.zip

Files (85.8 MB)

Name Size
md5:22e5c413a9abbc8c40e217c726b28831
85.8 MB Preview Download

Additional details

Related works