WigtnOCR-RADP

RCPS for retrieval-based selection, coverage for diagnosis, and RADP only for optional parser training.

✅ EMNLP 2026 Industry Track · Submission #384 · Accepted

🖼️ OpenReview currently records Accept (Poster). The presentation assignment is provisional; an oral assignment remains possible.

⏳ Camera-ready deadline: August 30, 2026 (AoE)

📄 Title: Retrieval-Conditional Parsing Score (RCPS): Choosing Document Parsers by Retrieval, Not by Appearance · camera-ready working PDF

📦 Builds on WigtnOCR v1 + KoGovDoc-Bench

🇰🇷 한국어 README · 🧭 docs/PAPER_READABILITY_REVIEW_AUDIT.md (current audit) · 🗓️ docs/CAMERA_READY_PLAN.md

---

TL;DR

Document RAG retrieves from parser outputs, yet parsers are often selected by intrinsic measures such as edit distance or Boundary Clarity (BC), rather than by retrieval performance. RCPS instead ranks parser–chunker combinations on a fixed held-out retrieval probe, without training. The resulting workflow is:

1. Select with RCPS. Five complete parser configurations span 0.137–0.584 RCPS. Audited MinerU-on has higher BC than Prod (0.713 vs 0.610) but a 42.6-point lower Hit@1 (0.123 vs 0.549; Prod is 4.47× as high). MinerU-off is retained only as a submitted-output diagnostic.
2. Diagnose with coverage. In Prod output, 20.2% of reference spans have no normalised exact match before chunking, while no tested chunker splits more than 2.3%.
3. Act or train only if needed. For absent spans, inspect or switch the parser; for split spans, change overlap or chunking. The 73-page parser-training pilot misses its target, and two RADP-DPO checkpoints are only +0.95 and +1.15 Hit@5 points above Prod on a separate, post-audit OHR compatibility subset.
4. Re-evaluate changed configurations. If P or C changes, run the resulting corpus through the same RCPS protocol before deployment. A covered configuration proceeds without redundant re-evaluation.

The denominators are not interchangeable:

Analysis	Evaluation frame	Role
Parser/chunker selection	663 Q–A retrieved against 294 pages	242 evidence pages + 52 Q–A-free distractors
Training/mechanism	The same 663 Q–A against 242 pages	Evidence-bearing pages only
Pre-specified pilot	202 Q–A / 73 pages	Held-out training gate
OHR perturbations	1,043 Law–Manual Q–A	Source-aligned cross-domain diagnostic
OHR training audit	2,036 Q–A / six domains	Post-audit compatibility subset; not a full v2 rerun

The repository currently contains the frozen 663-Q–A KoGovDoc-RAG probe (whose evidence spans 242 pages), its 169/73 evidence-page split, a portable 294-page source map, the RCPS implementation, selected result artifacts, and a public release of all nine evaluated parser-training adapters. The main selection run adds 52 Q–A-free distractor pages to form its 294-page index. The source documents required for a full fresh-clone rerun are not packaged; all remaining release gaps are listed below.

Figure 1 — RCPS workflow. A fixed 294-page / 663-Q–A frame evaluates every parser–chunker candidate. RCPS selects a provisional P* + C*; coverage then distinguishes covered, absent, and split spans. Any changed parser or chunker is evaluated again before final deployment. (vector PDF · editable PPTX)

---

Motivation — parsing quality ≠ retrieval performance

In the audited deployment comparison, MinerU-on has higher BC than Prod (0.713 vs 0.610) but much lower Hit@1 (0.123 vs 0.549). The separately retained submitted-output MinerU-off diagnostic has BC 0.716 and Hit@1 0.197. These configurations are not a causal table-recognition ablation; together, they show the operational risk of choosing a parser by boundary appearance alone.

OHR-Bench, EnterpriseDocBench, and concurrent OCR-for-RAG studies report related mismatches in English and enterprise settings. Our contribution is to turn that observation into a deployment workflow: select with RCPS, diagnose with coverage, change or train only if needed, and re-evaluate changed configurations. We are unaware of prior work that combines a reusable parser-selection protocol with a diagnostic that separates exact-span absence from chunk-boundary splitting.

---

Contributions

	Contribution	Headline result
C1	The parsing↔retrieval disconnect. On an aligned English OHR-Bench subset, semantic-noise perturbations lower retrieval while BC is stable or changes non-monotonically.	Audited 294-page BC↔RCPS r = −0.74 (descriptive); MinerU-on and Prod differ by 42.6 Hit@1 points (0.123→0.549; 4.47×)
C2	RCPS — a retriever-averaged, format-normalised, held-out-Q–A protocol for choosing parsers/chunkers with no training.	Complete 294-page parsers span 0.137–0.584; with Prod fixed, four chunkers span 0.535–0.593
C3	A retriever-free coverage diagnostic — classify each reference span as covered / split across chunks / absent from the normalised parser output; a rule computable before any retriever runs.	20.2% exact-span absent, constant across 8 chunkers; split varies up to 2.3% ⇒ inspect parser output first
C4	A bounded map of parser-side training; the pilot misses its target, and retrieval-reward training does not separate from a matched Distill control.	Post-audit six-domain OHR compatibility subset: R2 +0.95 pp, R3 +1.15 pp, Distill +1.36 pp Hit@5 vs Prod (n=2,036)

---

Method

RCPS — Retrieval-Conditional Parsing Score (C2)

Score a parser by what downstream retrieval does with its output, not by how clean the output looks. RCPS is not a new similarity function but a protocol wrapping ordinary retrieval MRR in three choices: (i) extrinsic (score on a held-out Q–A probe rather than parser text alone), (ii) retriever-averaged (over specified embedders), and (iii) format-normalised relevance (a relevant chunk must come from the source page and contain the normalised answer span).

RCPS(P, C; D, R, K) = (1 / |R||K|) · Σ_{r∈R} Σ_{k∈K} MRR@k(r, C(P), D)

Here, P is a parser, C is a chunker, and D is the fixed held-out Q–A probe. R = {BGE-M3, multilingual-e5-large, Qwen3-Embedding-8B} and K = {1, 5, 10}. A chunk is relevant iff its source page matches the answer's page and contains the reference span after shared whitespace and Markdown normalisation. For each query, MRR@k is 1/j when the first relevant chunk appears at rank j ≤ k, and zero when no relevant chunk appears in the top k; the values are then averaged over queries. Evaluation requires no training. Reference implementation: src/wigtnocr_radp/evaluation/.

Figure 2 — RCPS evaluation protocol. Every candidate uses the same probe, retriever/retrieval-depth specification, and reference-page plus normalised-span relevance rule. RCPS averages standard MRR and requires no training. (vector PDF)

Coverage diagnostic — parser vs chunker (C3)

RCPS scores parser + chunker + retriever jointly, so a low score does not say which layer is at fault. Holding the parser output fixed and varying the chunker, classify each normalised reference span as covered, split (present in the parsed page output but divided across chunks, and therefore potentially recoverable with overlap) or absent (no exact match in the normalised parser output, so re-chunking cannot restore that exact span). This diagnostic identifies the layer to inspect first; an absent match can reflect a genuine omission or a surface-form mismatch, which requires case-level review to distinguish. Code: scripts/evaluation/coverage_diagnostic.py.

Figure 3 — Coverage diagnostic with Prod fixed. The pre-chunking exact-span no-match rate remains 20.2%; changing the chunker affects only splitting, which reaches at most 2.3% across eight chunkers. (vector PDF)

Parser-side training — a secondary study (C4)

When the coverage diagnostic points to parser output, we test the following approaches and controls.

• RADP-aux (hidden-state auxiliary loss — sub-threshold). L_total = L_parse + λ·L_contrast uses InfoNCE between the parser's pooled answer-span hidden state and a frozen BGE-M3 embedding. Its best 73-page pilot estimate remains below the pre-specified success criterion. The executed sweep trains adapters from Qwen/Qwen3-VL-2B-Instruct and applies them to Prod for evaluation; it is therefore not a same-base comparison with the preference-based variants.
• RADP-DPO (discrete-output retrieval-reward DPO). Sample K parses from the production parser, score each by page-local BGE-M3 MRR averaged over k = {1, 5, 10}, form preference pairs, and train with a LoRA-toggle reference (π_θ = LoRA on, π_ref = LoRA off). The candidate pool and negatives expand across R1 → R2 → R3. The original R2 execution log verifies beta = 0.1; its portable executed configuration and source-log hash are recorded in docs/provenance/RADP_DPO_R2_EXECUTED_CONFIG.md. The complete nine-adapter release inventory and original/release hashes are recorded in checkpoint_release_manifest.json.
• RADP-Distill (fidelity-based control). Candidates are ranked by edit distance to reference Markdown instead of the page-local BGE-M3 MRR retrieval reward. On the aligned 2,036-Q–A frame, its Hit@5 delta is +1.36 pp versus Prod. Direct Distill-minus-R2 and Distill-minus-R3 intervals both include zero.
• SimPO (reference-free control). Its 242-page Hit@5 point estimates are negative, but both confidence intervals cross zero; the runs do not isolate which optimization difference caused that result.

---

Experiments

Setup

• KoGovDoc-RAG selection frame — all 663 Q–A retrieve against 294 pages (229 KoGov + 65 arXiv). The answers occur on 242 of those pages; the remaining 52 pages are Q–A-free distractors. Q–A were generated with gpt-5.4-2026-03-05; a separate LLM-assisted check accepted 94/100 sampled pairs. This was not human verification of the complete probe.
• KoGovDoc-RAG training frames — DPO/SimPO and mechanism analyses retrieve the same 663 Q–A against only the 242 evidence-bearing pages. The pre-specified pilot uses a held-out 73-page / 202-Q–A fold. Preference data come from a separate, page-disjoint 2,667-page Prod corpus with 6,164 generated Q–A.
• OHR-Bench frames — C1 uses 1,043 source-aligned Law–Manual Q–A, three benchmark outputs, and twelve dependent perturbations (three formatting and nine semantic); these are not 15 independent parsers. C4 uses a separate 2,036-Q–A / six-domain compatibility subset after excluding 223 misaligned legacy notes rows and five Q–A whose evidence page is absent from the current parser bundle. Neither frame is a substitute for a full v2 rerun.
• Model and scoring — Prod is Qwen3-VL-2B fine-tuned for Korean document parsing; trained variants use LoRA (r=8, α=32). RCPS averages 3 retrievers × 3 retrieval depths. Reported uncertainty uses paired Q–A-level percentile bootstrap unless stated otherwise.

C1 — the disconnect (KoGovDoc-RAG evaluation set)

The five candidates with complete outputs form the deployment comparison below. MinerU-on is the audited, table-enabled configuration used in that comparison.

Complete 294-page deployment comparison	BC	CS	RCPS	Hit@1
Qwen3-VL-30B (teacher)	0.623	3.38	0.584	0.545
Prod (ours, 2B)	0.610	3.07	0.583	0.549
Qwen3-VL-2B (base)	0.520	3.74	0.532	0.500
PaddleOCR	—	3.46	0.140	0.125
MinerU-on	0.713	—	0.137	0.123

Boundary Clarity is defined for four complete 294-page deployment configurations (Qwen3-VL-30B, Prod, Qwen3-VL-2B, and MinerU-on). Their BC–RCPS correlation is Pearson r = −0.74. Adding Marker's 38-page result gives r = −0.83 (n = 5). Both estimates are descriptive; PaddleOCR has no measured BC and Marker is not a complete-output result.

Submitted/subset diagnostic	Scope	BC	CS	RCPS	Hit@1
MinerU-off (submitted)	294 pages	0.716	2.81	0.212	0.197
Marker	38 pages	0.717	3.41	0.073	0.068

MinerU-on and MinerU-off differ in more than table handling, so their scores do not estimate the causal effect of table recognition. BC is Boundary Clarity (higher is better); CS is Chunk Stickiness (lower is better).

Figure 4 — Parsing quality can misrank retrieval candidates. Panel (a) uses MinerU-on in the audited 294-page deployment comparison; panel (b) compares MinerU-on and Prod on Hit@1. MinerU-off remains a separately labelled submitted-output diagnostic, not a causal table-recognition ablation. (vector PDF)

C1 — aligned perturbation diagnostic (cross-domain, OHR-Bench)

On the aligned Law–Manual subset, semantic noise lowers retrieval without a consistent BC response. From clean to severe noise, MinerU RCPS falls 0.595 → 0.265 while BC changes non-monotonically from 0.657 → 0.631; Qwen2.5-VL RCPS falls 0.545 → 0.497 while BC stays near 0.563. GOT has no clean output, but from mild to severe noise its RCPS falls 0.461 → 0.298 while BC rises 0.586 → 0.624. The aggregate 15-row correlation is r = −0.35, reported descriptively because variants within a family are dependent. This restricted subset does not establish broader domain generality.

C2 — RCPS ranks parsers and chunkers

Chunker	RCPS	Hit@1	MRR@10
md-h3	0.593	0.556	0.613
parser_native	0.583	0.549	0.602
LumberChunker	0.557	0.514	0.580
fixed500	0.535	0.491	0.560

KoGov chunking grid (663 Q–A, Prod output, 3-retriever RCPS average). On the submitted-output aggregate grid, which contains MinerU-off rather than the separate MinerU-on deployment row, the tracked audit finds that using three-retriever MRR@10 alone instead of averaging MRR@{1,5,10} preserves the complete order of the five 294-page parsers and all four chunkers. By contrast, dropping retriever-averaging and using only BGE-M3 inverts the top parser (Prod first; full RCPS ranks the 30B teacher first); the five-parser rankings otherwise agree (Kendall τ = 0.80). We do not report the format-sensitive ablation because the ranked chunk lists required to reconstruct it were not persisted. Use the deployment retriever when it is fixed; use the multi-retriever average as a hedge when the retriever is undecided or candidates are near-tied. RCPS is an operational protocol, not a relabelled MRR.

In a separate three-parser end-to-end check, Prod also has the highest judged answer accuracy (72.5%, versus 23.8% for MinerU-on and 20.5% for PaddleOCR). The lower pair reverses relative to RCPS, and the same GPT-5.4 checkpoint generates and judges answers. We therefore treat this only as a check of the top choice, not validation of the full ranking.

C3 — coverage diagnostic separates output absence from chunk boundaries

On Prod's output (294 pages: 229 KoGov + 65 arXiv; 663 Q–A, no retriever), 134/663 (20.2%) of normalised reference spans have no exact match in the parser output, while at most 15/663 (2.3%) are split across chunks. The exact-span absence rate is constant across all eight chunkers, so re-chunking cannot make those cases exact-span covered. This result points to parser-output inspection before chunker tuning; it does not, by itself, show that the answer semantics are entirely missing rather than rendered in a different surface form.

We also test label robustness. GPT-5.4 reclassifies 56% of Prod's exact-match-absent cases as recoverable surface artefacts, although this judge is not independent of the GPT-family Q–A generator. In a separate parser-masked, stratified sample of 100 absent cases, two authors independently agree on 81/100 cases (κ = 0.615) before adjudication. After adjudication, retrieval-unusable rates are 42/50 (84.0%) for MinerU-on, 12/30 (40.0%) for Prod, and 19/20 (95.0%) for PaddleOCR. Different sampling fractions and MinerU configurations prevent a population-level replication claim. The sampling manifest, both rating files, and the 19-case adjudication record were independently rechecked with the repository scorer. They are retained in an author-only audit package; the public release reports only the aggregate results above.

C4 — parser-side training remains below the pilot target

On the held-out 73-page / 202-Q–A pilot, neither RADP-aux nor RADP-DPO meets the pre-specified target: at least 5 RCPS points with a 95% confidence-interval lower bound above zero. A later OHR audit found that the original result mixed benchmark releases. After removing 223 misaligned legacy notes rows and five Q–A tied to a missing evidence page, the tracked arrays yield this strict six-domain compatibility result:

Δ vs Prod (pp)	Hit@1	Hit@5	Hit@10	MRR@10	nDCG@5
RADP-DPO R2 (retrieval reward)	+0.59	+0.95	+0.90	+0.78	+0.82
RADP-DPO R3 (hard-negative)	+1.46	+1.15	+0.90	+1.30	+1.28
RADP-Distill (edit-distance control)	+0.98	+1.36	+1.47	+1.12	+1.16

Post-audit legacy compatibility subset, n=2,036, three-retriever macro, 1,000 Q–A-level paired-bootstrap resamples (seed 42). Hit@5 95% CIs: R2 [+0.33,+1.54], R3 [+0.31,+2.05], Distill [+0.43,+2.29]. Distill−R2 is +0.41 pp [−0.43,+1.26] and Distill−R3 is +0.21 pp [−0.61,+1.05]. This subset was defined after the version audit; it is neither the original confirmatory analysis nor a full OHR-Bench v2 evaluation.

On the exploratory KoGov fold (242 pages, n = 663), the RADP-DPO milestones reach +1.96 to +2.11 pp Hit@5 (P[Δ>0] ≈ 0.90; all two-sided intervals cross zero). SimPO Hit@5 point estimates are −0.85 pp with md-h3 and −0.70 pp with parser-native chunking; both confidence intervals cross zero. This pooled analysis combines the 169 development and 73 held-out evidence pages, so it is not a new independent holdout. The matched OHR comparison does not show that retrieval-reward pair selection outperforms edit-distance pair selection.

C4 — available measurements do not identify a training mechanism

Variant	BC ↑	TextNED ↓ vs reference
Prod (ref)	0.630	0.240
RADP-DPO R2	0.647	0.163
RADP-DPO R3	—	0.185
RADP-aux λ=0.1	0.652	0.423

Selected 242-page mechanism measurements. Available R1–R3 measurements have lower TextNED and positive Hit@5 point estimates relative to Prod; the compact table shows R2 and R3, but TextNED does not reproduce their retrieval order. R3 BC and uncertainty estimates for these structural metrics are unavailable. The co-occurrence is post hoc and does not establish that fidelity or boundary changes caused the retrieval differences.

---

Deployment playbook

1. Evaluate parsers with RCPS, not intrinsic metrics alone. Boundary Clarity can rank candidates in an order the downstream retriever does not preserve. In the audited MinerU-on–Prod comparison, selecting Prod corresponds to Hit@1 0.123 → 0.549 on the fixed KoGovDoc-RAG probe; this candidate-pool result is not a universal effect size. This is the highest-leverage takeaway.
2. Run the coverage diagnostic first. If exact-span absence dominates, inspect those parser outputs and change the parser when content is genuinely missing; if split dominates, tune the chunker or overlap.
3. If you train the parser, measure it on an untouched subset and include a fidelity-based baseline. The audited DPO checkpoints have Hit@5 point estimates about one point above Prod, but the Distill comparison is not recoverable from the current per-QA artifacts.
4. Keep conclusions tied to available evidence. The current mechanism results associate DPO with lower TextNED than Prod; they do not identify a causal mechanism or establish an evidence-type advantage.

---

Repository structure

.
├── configs/                  # experiment configs (YAML)
├── src/wigtnocr_radp/
│   ├── qa_generation/        # Q-A generation
│   ├── evaluation/           # RCPS, chunkers, retrievers, coverage, Boundary Clarity, bootstrap CI
│   └── training/             # RADP-aux (contrastive) · RADP-DPO · SimPO (LoRA-toggle ref)
├── scripts/
│   ├── training/             # candidate gen, preference / edit-distance pairs, DPO/Distill/SimPO pipelines
│   ├── evaluation/           # baseline_grid, chunking_grid, coverage_diagnostic, rcps_protocol_ablation, OHR chains
│   └── figures/              # paper figure generators (disconnect, coverage, RCPS protocol, noise family)
├── experiments/              # RADP-Distill training/evaluation harness
├── paper/                    # frozen submission + camera-ready working LaTeX + figures
├── data/KoGovDoc-RAG/        # frozen 663-Q–A probe + 169/73 evidence-page split
├── docs/                     # RESEARCH_DIRECTION · TIMELINE · ROADMAP · plans/ · literature_review/
├── output/                   # selected result JSONs; see artifact status below
└── tests/

Figure source note: the canonical camera-ready assets are paper/figures/fig_overview.pdf, fig_rcps_protocol.pdf, fig_coverage.pdf, and fig_disconnect.pdf; the PNG files displayed in this README are their web previews. Figure 1's canonical editable source is paper/figures/fig_overview_camera_ready.pptx; its RCPS badge is C2 and its coverage badge is C3.

Local code check (Linux/WSL CUDA environment)

uv sync --extra dev
uv run pytest
uv run python scripts/analysis/source_page_map.py --check data/KoGovDoc-RAG/source_page_map_v1.json
uv run python scripts/analysis/audit_mineru_output_release.py --check output/results/mineru_output_release_audit.json
uv run python scripts/analysis/audit_kogov_training_table.py --check output/results/kogov_training_table_10k_audit.json
uv run python scripts/evaluation/coverage_diagnostic.py --out_dir /tmp/rcps-coverage-check

These commands assume the Linux/WSL CUDA 12.8 dependency source encoded in pyproject.toml and uv.lock; a clean macOS/CPU installation path for the full dependency set has not yet been packaged or validated. pytest exercises the tracked unit and alignment-gate tests. The coverage calculation itself is CPU-only. Its default Prod outputs under results/kogovdoc/v1_val/predictions/ and the 663-Q–A probe are tracked. The portable source-page map supports inventory checks, but the private training-format data/KoGovDoc-Bench/val.jsonl used by the coverage command includes reference transcriptions and is not packaged. The command is therefore not fresh-clone complete; it can use equivalent local inputs through --parser_dir and --val_jsonl.

These commands are not a reproduction of every paper experiment. Full reruns still depend on external source documents, some parser outputs, embedding caches, and parser runtimes. The missing pieces are listed explicitly below.

CPU-only artifact and checkpoint verification

The released evaluation artifacts and nine LoRA adapters can be audited from a clean checkout without a GPU:

RCPS_RELEASE_DIR="$(mktemp -d)"
curl -fL -o /tmp/RCPS-RADP-Adapters-v1.tar.gz \
  https://github.com/wigtn/RCPS-RADP-Adapters/releases/download/v1.0.0/RCPS-RADP-Adapters-v1.tar.gz
tar -xzf /tmp/RCPS-RADP-Adapters-v1.tar.gz -C "$RCPS_RELEASE_DIR"
python3 scripts/release/verify_camera_ready_artifacts.py \
  --checkpoint-dir "$RCPS_RELEASE_DIR"

The release tarball SHA-256 is 22b4a0d4f5560f4d7c31633a1c885bbe3568a2bef841bfd94bf2407c339f08c6. The deterministic builder, public-download manifest, and inclusion policy are documented in CHECKPOINT_RELEASE.md.

---

Authors (OpenReview and camera-ready order)

Follow-up to WigtnOCR v1 (Qwen3-VL-2B document-parsing fine-tuning).

1. Sang-Woo Son
2. Hyeong-seob Kim
3. Hyeonsang Kim
4. Hyun-woo Cho
5. Jinmo Kim

The author list and order remain exactly as submitted. The Industry Track chairs confirmed in writing that Hyeong-seob Kim may be designated as corresponding author without changing that order. The camera-ready PDF therefore lists the shared affiliation as WIGTN, Seoul, Republic of Korea and uses the ACL-template line Correspondence: harrison@wigtn.com. Other emails will be added only from confirmed metadata.

---

Artifact and reproducibility status

Current repository

• KoGovDoc-RAG probe files — 663 Q–A whose evidence spans 242 pages, plus the frozen 169-development / 73-held-out evidence-page split. The paper's selection frame indexes these 242 pages together with 52 Q–A-free distractors (294 pages = 229 KoGov + 65 arXiv); that full source-page corpus is not packaged here. The portable source-page map links every val_#### ID to its domain and tracked parser-output filename without exposing machine-local paths or reference transcriptions. A separate LLM-assisted 100-pair Q–A quality-check sample and its aggregate 94/100 result are tracked; the sample's blank verification fields are not human annotations.
• RCPS reference implementation — src/wigtnocr_radp/evaluation/.
• Selected evaluation artifacts — aggregate parser/chunker grids, aligned full-grid and audited-training per-Q–A arrays, coverage and end-to-end diagnostics, and complete 294-page outputs for Prod, PaddleOCR, MinerU-on, and the submitted-output MinerU-off run. The recovered off outputs are bound to their aggregate by the deterministic release audit; the on/off runs are not treated as a controlled table-recognition ablation. The 10,000-resample KoGov training-table audit reconstructs Hit@5 from the tracked MRR arrays and reproduces every value in camera-ready Table 8. The 294-page full-grid audit includes fullgrid_perqa_294p.json and fixed-seed parser/chunker ranking-stability results. Across 1,000 500-of-663 draws, Prod stays above Base and every OCR parser in 100% of draws; the complete chunker order is unchanged in 96.1%. Raw matching lowers RCPS by 0.024–0.041 without reordering either pool.
• Aligned OHR audit artifacts — the 1,043-Q–A Law–Manual C1 result, the aligned RADP-Distill per-Q–A arrays, and a deterministic derivation of the strict 2,036-Q–A legacy compatibility subset. Older seven-domain outputs remain in the tree for provenance and are listed in MANIFEST.legacy-invalid.sha256; they are not valid camera-ready evidence.
• Aggregate human-check results — the paper records the parser-masked 100-case absent-label study (κ = 0.615, raw agreement 81/100, and post-adjudication parser-specific rates). The underlying sampling, rating, and adjudication records were rechecked and are retained in an author-only audit package.
• Parser-training checkpoints — all nine evaluated LoRA adapters are public in wigtn/RCPS-RADP-Adapters release v1.0.0. The tracked release manifest records source and portable hashes, training/evaluation base models, and executed configurations. The release contains the four RADP-aux settings, RADP-DPO R1–R3, RADP-Distill, and RADP-SimPO.
• Camera-ready figure assets — Figures 1–4 are stored as vector PDFs with PNG README previews; Figure 1 also includes its canonical editable PPTX. The compiled paper was checked with embedded fonts and no Type 3 fonts.
• Legacy metric quarantine — output/baselines/correlation_v1.json is retained only as a historical Phase-1 artifact and is marked as using a non-MoC boundary definition; it is not camera-ready evidence.

Camera-ready pending — not currently available

• Qwen3-VL-30B and Qwen3-VL-2B-base parser outputs, plus exact rerun commands for the released parsers.
• The paper makes no full OHR-Bench v2 claim; a future full-v2 experiment would require the official v2 Q–A plus fresh parser and retrieval runs. Legacy seven-domain / combined-CI / OHR-TextNED artifacts remain separated in the quarantine manifest.
• Complete BC/CS mechanism data and aligned uncertainty estimates.
• Raw preference-pair text and complete original training logs. The public checkpoint release instead provides portable executed configurations, verified adapters, and available structured trainer states.
• A portable fresh-clone, end-to-end reproduction path, including external data, parser outputs, embedding caches, base-model acquisition, and removal of machine-specific runtime assumptions.

License & Citation

Repository code is released under the MIT License. Third-party datasets and model assets retain their original licenses and terms.

@inproceedings{son2026rcps,
  title     = {Retrieval-Conditional Parsing Score (RCPS): Choosing Document Parsers by Retrieval, Not by Appearance},
  author    = {Son, Sang-Woo and Kim, Hyeong-seob and Kim, Hyeonsang and Cho, Hyun-woo and Kim, Jinmo},
  booktitle = {Proceedings of EMNLP 2026 (Industry Track)},
  year      = {2026},
  note      = {Accepted; camera-ready pending}
}