[byte271]

New executable research mechanism

This is the next experimental step toward LLM text-watermark removal, not a detector product and not another Unicode layer. It is not yet a demonstrated watermark remover.

• Add a dependency-free finite candidate-lattice planner under research/reconstruction/.
• Track cross-sentence n-gram boundaries and optimize three normalized-text views jointly under a fixed edit budget.
• Use source-sized denominators so output padding cannot dilute the objective.
• Pareto-prune only equal-tail states, replay the complete output, and fail closed on state/evaluation limits. Exactness is limited to the supported tokenizers, supplied lattice, and objective.
• Add a deliberately narrow English reporting-frame frontend: entity, metric, literal amount, unit, year, and negation must reparse identically and in order. Unsupported text is returned unchanged, not silently treated as successful removal.
• Provide runnable CLI, receipts, frozen development protocol, greedy comparator, exhaustive references, and a legacy-carrier inverse control.

What is and is not measured

The corpus comprises 128 newly synthesized controlled-grammar documents. It is not LLM-generated watermarked prose. The metric is retained source-window overlap, not a watermark detector score or removal rate. Regex word/token views are not GPT-2/BPE or separate commercial tokenizer families.

Local check	Result
New tests	102 passed
Existing CLI, CLI I/O, protected-machine-text tests	70 passed
Combined targeted tests	172 passed
Joint search strictly improves worst-view retention over identical-budget full-output coordinate greedy	51/128
Ties on worst-view retention	77/128
Mean worst-view retention: greedy / joint	0.6205182055 / 0.5987639731
Exhaustive benchmark comparisons	24/24
Randomized exhaustive-search tests	48/48
Grammar-frame/form roundtrips	2,048
Legacy outputs restored by deleting known carriers	128/128
New outputs unchanged by NFKC plus known-carrier deletion	128/128

Local runtime: CPython 3.13.5 on Linux. These are targeted local results, not a claim that full GitHub CI is green. The comparison is against the included coordinate-greedy baseline, not the repository's historical scheduler.

Canonical benchmark rows SHA-256: 36d173f4e1521f74a963754c917724eaf4e5ba37b62201feb662a5510f495736.

Limits

No real SynthID/KGW/proprietary-detector evaluation was run: the local runtime lacks pretrained tokenizer/model assets and outbound DNS was unavailable. The current end-to-end frontend handles only the documented reporting grammar; arbitrary English/Chinese prose is not supported. Visible wording changes. Frame equality is not universal semantic proof. Receipts explicitly retain watermark_removal_verified: false.

Scope control

Based on 09c8cbf184e37e2e934e133b74a5db9af67b4f97, independent of import-boundary PR #124. Only 10 new research/test/documentation files are added. No production algorithm, CLI default, hidden-carrier stack, frozen evidence, dependency, version, or workflow trigger changes. No model-generation reruns were manually requested. Keep draft until full-checkout CI and scientific review are complete.

Reproduce

python -m pytest -q tests/test_reconstruction.py
python -m research.reconstruction --text "Orion Labs reported a revenue of 120 USD in 2025. Vega Systems did not report a cost of 42 EUR in 2024." --json
python -m research.reconstruction.benchmark --output geometry-128.json
python -m research.reconstruction.legacy_control --output legacy-control.json

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn