PhaseWrap

PhaseWrap is a negative-results and methodology repository for a fixed phase-wrapped relative-position score. The original positive thesis is closed: this repository does not claim a RoPE replacement, production transformer superiority, quantum advantage, or broad cross-backend hardware robustness.

The public project name is PhaseWrap. The historical Python package and frozen artifact namespace remain qrope for reproducibility; phasewrap is provided as a compatibility import surface for new code.

from phasewrap import phase_margins, phase_residual, phasewrap_features, phasewrap_score

score = phasewrap_score(reference_delta=37, candidate_delta=13)
features = phasewrap_features(reference_delta=37, candidate_delta=13)

The same API is available from qrope because archived scripts, manifests, and packet IDs use that namespace.

What This Repo Shows

PhaseWrap is organized around three evidence lanes.

Lane	Decisive evidence	Supported conclusion	Not supported
Score theory	Stage 11	The fixed (8, 12) score has period 24, only 10 distinct residue values, mirror aliases, and Fourier support [1, 2, 3, 5].	A general-purpose positional encoding.
Benchmark methodology	Stages 5, 67, 74/75, 80/81, 93/94/96	Assistance pipelines can repair retrieval for no_position controls too, so benchmark wins must isolate whether the positional method did the work.	Novel discovery of shortcut learning, or proof that external papers are wrong.
Hardware-readout audit	Stages 216-218	One archived IBM Fez packet favors PhaseWrap on normalized readout-noise delta after known-state calibration.	Transformer-scale evidence, quantum advantage, or cross-backend robustness.

The strongest transferable result is the methodology warning: support-routed or assistance-based retrieval repairs can make a positional method look effective even when the same repair also solves the no-position control. This is a worked example of established shortcut-learning, control-task, and NoPE/NoPos concerns, not a claim that the general concern is new.

The score itself is exactly classically computable and does not need a quantum computer. The hardware lane exists to audit saved provider readout evidence, not to make a compute-advantage claim.

Claim Boundary

Supported:

• Exact characterization of the fixed PhaseWrap score, including SQR = m8 * m12.
• Negative-results evidence that the original phase-cued target is exposed to simple periodic baselines such as lookup_mod24; Stage 5 is a motivation constraint, not a positive result.
• A reproducible worked example showing assistance-pipeline confounds in positional-retrieval benchmarks.
• No-credential verification of saved artifacts, including archived hardware readout packets.
• Bounded hardware-readout audit evidence using known-state calibration and provider-aware bitstring handling.

Not supported:

• RoPE replacement.
• Production transformer superiority.
• Quantum advantage.
• Generalization to production long-context models.
• Broad cross-backend hardware robustness.
• Patentability, commercial defensibility, or implementation exclusivity as a scientific conclusion.

Hardware-Positive Means Bounded Readout Audit

Hardware-positive in this repository means a bounded two-qubit readout-audit result on archived packets. It does not imply RoPE replacement, transformer-scale improvement, quantum advantage, or general hardware robustness.

The public Stage 218 decision name is:

IBM_FEZ_FROZEN_PACKET_READOUT_NOISE_DELTA_FAVORS_PHASEWRAP

The older manifest identifier FULL_REPLACEMENT_HARDWARE_POSITIVE_PHASEWRAP_ADVANTAGE is preserved only as frozen historical evidence. New prose and review materials should use the bounded readout-noise wording.

Reviewer Start

Use the no-credential public verifier first:

python -m pip install -e ".[dev]"
python -m qrope.verify_publication --profile public

Expected marker:

PHASEWRAP_PUBLIC_VERIFY_PASS

Equivalent console entry point:

phasewrap-verify --profile public

Focused checks:

python scripts/run_stage11_phasewrap_theory.py
python scripts/run_stage80_support_routed_token_selector_audit.py
python scripts/run_stage218_full_replacement_hardware_metric_interpreter.py
python scripts/run_stage219_rope_substitution_gate.py

Legacy Stage 4 hardware packet verifier:

python scripts/verify_stage4_hardware_packet.py

Expected verifier summary:

{
  "pass": true,
  "provider": "ibm_runtime",
  "backend": "ibm_fez",
  "packet_id": "qrope-hardware-73c61893576297ff",
  "job_ids": [
    "d84jbq00bvlc73d4krr0"
  ]
}

For a shorter human entry point, start with REVIEWER_START.md. For the canonical evidence map, see docs/stage-to-claim-map.md.

Canonical Experiments

The full repository preserves the historical stage archive, but the public argument is carried by a small canonical set:

Stage(s)	Role in the argument
5	Shows lookup_mod24 and direct periodic features exactly recover the original synthetic label.
11	Characterizes the fixed score: period 24, 10 values, aliases, and Fourier support.
12	Tests stricter non-phase-cued retrieval and shows fixed PhaseWrap is weak.
30/32 and 219	Show bounded top-1/MRR ranking parity in selected bridges while RoPE retains stronger probability/calibration. This is a ranking-parity bridge, not substitution adequacy.
67	Content-key redesign solves for all methods, including no_position.
74/75	Query-support recovery solves for no_position too.
80/81	Support-routed selectors repair phase-cued retrieval for no_position too.
82/93/94/96	Learned support-to-token binding and promotion gates remain bounded or fail.
216-218	Freeze and interpret the bounded IBM Fez hardware-readout packet.

Everything else is retained for provenance, replication, or historical audit. It should not be read as a separate headline claim.

In short: support-routing repairs solve phase-cued retrieval for all tested methods, including no_position.

Trained Ablations

Archived autograd-based toy transformer and bridge ablations are historical negative-results context only. They are not headline evidence for production transformer behavior. Any future positive transformer claim should be rerun in a current ML framework with preregistered metrics, strong controls, and real-task baselines.

Repository Layout

src/qrope/                 Historical package namespace and verifier code
src/phasewrap/             Compatibility import surface for new code
scripts/                   Reproduction and stage runner scripts
docs/stage-to-claim-map.md Canonical stage-to-claim map
docs/publication/          Methodology drafts, release notes, review matrices
logs/automated_stage_gates Frozen stage outputs and archived evidence
tests/                     Public API, verifier, scoring, qsim, and stage tests

Documentation Pointers

• REVIEWER_START.md: 15-minute no-credential review path.
• docs/stage-to-claim-map.md: canonical experiments and claim mapping.
• docs/publication/phasewrap-methodology-paper-v1.md: methods-paper draft.
• docs/publication/post-publication-novelty-and-prior-work-supplement-v1.md: prior-work and novelty boundary.
• docs/reproducible-environment.md: reviewer environment details.
• docs/api/scoring.md: public scoring API.
• docs/publication/figures/README.md: publication figures, including qrope-full-replacement-metrics-v1.png.
• PATENTS.md: low-prominence legal notice. Patent filings are not used as scientific evidence.

Installation

Reviewer path:

python -m pip install -r requirements-review.txt
python -m pip install -e ".[dev]"

Core package only:

python -m pip install -e .

Provider SDKs are optional and are not required for saved-artifact verification.

python -m pip install -e ".[ibm]"
python -m pip install -e ".[braket]"
python -m pip install -e ".[quandela]"

Citation

Use the archived negative-results release and DOI:

Quantyra. PhaseWrap negative-results evidence and reviewer-verification tooling.
Zenodo. https://doi.org/10.5281/zenodo.20387905

See CITATION.cff for machine-readable citation metadata.

License

Code is released under AGPL-3.0-only. Documentation and research text are covered by the repository notices. See LICENSE, NOTICE, and PATENTS.md.