NRR-Phi: A Typed External Text-to-State Interface and Update Contract for Inspectable Ambiguity-State Maintenance
NRR-Phi turns retained-state maintenance into a typed, executable, and falsifiable interface. It maps extracted items into operator-addressable (v, c, w, m) records and declares how later updates may retain, dampen, integrate, persist, or commit them. Across 580 constructed states/pairs and 2,740 operator--state measurements, the tested non-violating transitions and checks show 0% record-weight entropy violations, while uniform subtraction violates the criterion at 1.7%, 6.1%, and 17.8% as subtraction grows. A separate 68-input construction audit shows that the reported extraction procedures populate the typed schema with multiple positive-weight records (H = 1.087 bits).
The operator result is executable contract-conformance evidence with a discriminating negative control. Phi exposes two maintenance layers separately: operator definitions state which records are carried forward, and the entropy criterion detects concentration in normalized record weights. The entropy criterion does not by itself certify record identity or cardinality. Semantic adequacy and end-to-end behavioral improvement remain separate questions.
Every conformance row is executed, but its inferential basis is explicit: dampening and state calibration have analytic entropy results; deferred resolution is an identity check; integration and persistence carry records forward by definition while their entropy rows remain tested conformance results; uniform subtraction is the designed negative control.
The formal extractor now makes both branches return the same typed tuple:
TypeSegments assigns a declared rule score and marker-derived item type, while
TypeCandidates parses the LLM confidence field and candidate class. Core and
Phi share nonnegative, not-necessarily-bounded weights; entropy is evaluated only
for positive-total states. Dampening is defined on lambda in [0,1], and the
entropy guarantee is nondecrease; it is a strict increase only for lambda > 0
with a nonuniform normalized input weight vector. The
cross-context iota utility is a tagged disjoint union with no unconditional
entropy-nondecrease claim.
Quick links
EN/JA query terms
early commitment=早期確定ambiguity-preserving inference=曖昧性保持推論
Part of the Non-Resolution Reasoning (NRR) research program. This repository presents the text-to-state mapping line and bundles the manuscript, transcript-audit artifacts, and verification scripts for that line.
For the cross-paper map and current series links, start here:
Version mapping source of truth: VERSION_MAP.md
For the current package snapshot and transcript-audit artifacts, start with the sections below; VERSION_MAP.md remains the broader version history reference.
NRR is not an anti-LLM framework.
NRR does not replace standard LLM use.
NRR optimizes when to commit and when to defer, under explicit conditions.
Series numbering policy: paper3 is permanently skipped and never reused.
- Archive concept DOI: 10.5281/zenodo.18793562
- Latest inspected archived snapshot: 10.5281/zenodo.18866847
- Earlier exact snapshot retained for provenance: 10.5281/zenodo.18793563
nrr-phi/
├── data/
│ ├── rule_based_data.json
│ └── operator_validation_states.json
├── results/
│ ├── rule_based_output.json
│ ├── operator_validation_results.json
│ └── llm_rerun_sanity_check_summary_20260419.json
├── experiments/
│ ├── rule_based_experiment.py
│ └── run_operator_validation.py
├── prompts/ # Prompt+response transcripts for the main-text LLM table
│ └── llm_audit_manifest.json # Included sentence IDs and aggregation contract
├── notebooks/
├── manuscript/
│ ├── current/
│ │ ├── paper2_nrr-phi_v58.tex
│ │ └── paper2_nrr-phi_v58.pdf
│ ├── figures/
│ │ ├── figure1.png ... figure5.png
│ ├── checksums_active_review_surface_sha256.txt
│ ├── checksums_current_package_sha256.txt
│ └── archive/
│ └── ... # Full-repo history
├── scripts/
│ ├── audit_llm_transcripts.py
│ ├── check_operator_summary.py
│ ├── README.md
│ ├── build_current_manuscript.sh
│ ├── run_primary_checks.sh
│ ├── verify_active_review_surface.sh
│ └── verify_current_package.sh
├── VERSION_MAP.md
├── reproducibility.md
├── requirements.txt
├── README.md
├── LICENSE
└── results/
├── llm_rerun_sanity_check_summary_20260419.json
├── llm_transcript_audit_summary.json
├── operator_validation_results.json
└── rule_based_output.json
Published and local version mapping is maintained in VERSION_MAP.md.
- Current public arXiv line:
2601.19933v6 - Current local revision candidate:
manuscript/current/paper2_nrr-phi_v58.tex - Preserved public-v6 source-line baseline:
manuscript/archive/local-v48_pre_related_work_2026-07-13/paper2_nrr-phi_v48.tex - Preserved pre-revision baseline:
manuscript/archive/local-v47_pre_tmlr_revision_2026-04-18/paper2_nrr-phi_v47.tex - Blind TMLR submission variant:
manuscript/archive/local-v48_tmlr_blind_2026-04-18/paper2_nrr-phi_v48_blind.tex - Earlier package versions remain recorded in
VERSION_MAP.md.
Primary package entrypoints:
bash scripts/build_current_manuscript.shbash scripts/verify_active_review_surface.shbash scripts/verify_current_package.shbash scripts/run_primary_checks.sh
pip install -r requirements.txt
python3 experiments/rule_based_experiment.py > results/rule_based_output.txt
python3 experiments/run_operator_validation.py \
--data data/operator_validation_states.json \
--output results/operator_validation_rerun_summary.json
python3 scripts/check_operator_summary.py results/operator_validation_rerun_summary.jsonSee reproducibility.md for environment, fixed settings, commands, and artifact mapping.
Deterministic reruns in this package cover the rule-based mapping and the main operator-conformance validation.
The operator runner fails closed outside the declared executable domains: state
weights must be finite and nonnegative, entropy requires a positive-total state,
dampening requires lambda in [0,1], proportional stripping requires
0 <= b < max(w), and persistence requires 0 < gamma < 1. The primary-check
wrapper also asserts the complete fixed operator row set, sample counts, violation
counts, rates, published mean-entropy-change values (absolute tolerance 1e-12),
and 2,740-measurement total.
The bundled prompts/ files plus prompts/llm_audit_manifest.json,
scripts/audit_llm_transcripts.py, and
results/llm_transcript_audit_summary.json form the bundled audit path
for the main-text LLM extraction table and combined H = 1.087 line.
For the archived LLM-handled cases, the package also bundles
results/llm_rerun_sanity_check_summary_20260419.json, which records the fixed
18-set current-API sanity-check rerun as a bounded stability note: nine sets were
consistent, eight were family-preserving reshapes, and one was divergent. It is
not full live reproducibility.
I support written technical Q&A, concept clarification, and small evaluation design.
Typical flow:
- you send questions and context,
- I return a structured technical response,
- if needed, I provide an English-ready version for external sharing.
Scope: research interpretation and evaluation planning.
Out of scope: production integration, implementation outsourcing, ongoing operations, and SLA/deadline commitments.
Contact: kei.saito.research@gmail.com
CC BY 4.0. See LICENSE.