{"path":"research/the-jhu-decomposition-line.md","content":"# The JHU Decomposition Line — Two Papers Read, and What They Cost Us\n\n**Date**: 2026-08-27\n**Occasion**: The founder asked for a survey of the group behind *A Closer Look at Claim Decomposition* and the two most relevant papers read properly.\n**Group**: Miriam Wanner, Benjamin Van Durme, Mark Dredze, Zhengping Jiang, Seth Ebner and collaborators (Johns Hopkins CLSP). **This is the group whose DnDScore the extraction pipeline is built on**, so their results are not adjacent literature — they are the pipeline's own foundations.\n\n**Selection note**: their output spans claim decomposition, factual-precision scoring, Wikipedia grounding, LLM critiques of papers, linguistic calibration, and a separate line on reasoning-model honesty. The two below were chosen because each names a defect Deliberus **currently has**, rather than a capability it might want.\n\n---\n\n## Paper one — CORE: the splitting inflation is a published attack, and it has a defense\n\n*Jiang, Zhang, Weir, Ebner, Wanner, Sanders, Khashabi, Liu & Van Durme, ACL Findings 2025 ([arXiv:2407.03572](https://arxiv.org/abs/2407.03572))*\n\nThe abstract, verbatim on the part that matters:\n\n> \"We observe that popular metrics using the Decompose-Then-Verify framework, such as FACTSCORE, **can be manipulated by adding obvious or repetitive subclaims to artificially inflate scores.** This observation motivates our new customizable plug-and-play subclaim selection component called CORE, which **filters down individual subclaims according to their uniqueness and informativeness.**\"\n\n**This is the exact defect measured in Deliberus's own strength layer the same week**, arriving as a named, published attack on the whole metric family. Sibling energy is additive with no correlation term: one claim at 0.70 contributes 0.06, three children contribute 0.18, twelve contribute 0.72 — and nothing new was learned between those rows. What the corpus had recorded as an unexamined house choice is, in the literature, **score manipulation with a known method**.\n\n**Three consequences.**\n\n1. **It changes the adversary class.** The corpus's standing calibration is that no instrument here has met an interested party. This is the first case where the literature says the interested party's move is already documented and cheap: add obvious or repetitive children. Splitting-inflation is therefore **strategy-class**, not an entropy-class tidiness issue.\n2. **Deliberus already ships half the defense, and it is nearly inert.** (Corrected 2026-08-27: 132 of 149 decomposition edges untagged, the stress suite having tagged 17 — so the parameter is lightly exercised rather than untouched. Live: `uv run python scripts/graph_facts.py`.) `support_semantics.necessary` removes the payoff entirely — three parts and twelve parts both score 0.502 — which is the payoff-removal shape the scale ladder recommends reaching for *before* a detector. All 131 decomposition edges are untagged, so it protects nothing today.\n3. **CORE is the other half, and it is a different mechanism**: filter by *uniqueness and informativeness* rather than cap by joint requirement. The two compose — necessity handles \"these parts are jointly required\", informativeness handles \"this part adds nothing\" — and Deliberus has no instrument for the second at all. The nearest thing is the worth-asking layer, which prices *questions* by outcome-swing; the same idea applied to *children* is the missing piece.\n\n## Paper two — DnDScore: the pipeline's own basis, and the tension it names is unhandled here\n\n*Wanner, Van Durme & Dredze, EMNLP 2025 ([arXiv:2412.13175](https://arxiv.org/abs/2412.13175))*\n\nThe tension, in their words:\n\n> \"Their conflicting purposes can create tensions: **decomposition isolates atomic facts while decontextualization inserts relevant information.** Furthermore, a decontextualized subclaim presents a challenge to the verification step: **what part of the augmented text should be verified as it now contains multiple atomic facts?**\"\n\nTheir answer is not to choose a point on that trade. **They keep the pair** — the atomic subclaim *and* its decontextualized form — and verify the atom **in the context of** the decontextualized version.\n\n**The gap this exposes, checked against the schema rather than assumed.** `AtomicClaim.text` is documented as *\"The claim, decontextualized to pass the stranger test\"*, and the stored `Claim` node's properties are `claim_id, text, short_title, type, claim_kind, epistemic_status, evidence_type, confidence, source_span, attributed_to, argument_label, reasoning, cq_template_id, group_id, embedding, recorded_at`. **There is no field for the pre-decontextualization atom.** Pass 2a produces it and Pass 2b overwrites it.\n\nSo the paper the pipeline is named after has an architecture Deliberus cannot express: the graph keeps the augmented form only.\n\n**Why that is not merely tidiness.** Every claim in the graph carries inserted context. When a user attacks a claim, **which part are they attacking** — the atomic fact, or the context the pipeline added to make it stand alone? Nobody has named that ambiguity, and it sits underneath addressability, which is the substrate everything else rides on. A disagreement aimed at inserted context is a disagreement with *the extractor*, not with the author.\n\n**One honest limitation of theirs, which bounds how much we can borrow.** They state plainly: *\"DnDScore is not intended to handle debatable claims, such as opinions.\"* Their evaluation is generated biographies against Wikipedia. Deliberus's corpus is mostly the case they exclude, so the *tension* they name transfers and their *verification method* largely does not.\n\n## What this corrects in our own docs\n\n**A phrasing of mine from earlier today, in `depth.md`, overstated.** I wrote that atomicity and decontextuality are two axes and that maximal separation with self-contained pieces \"is not a tension; it is the target.\" The first half stands and the second half is wrong: DnDScore measures a **genuine mechanical trade** — inserting the context that makes a piece self-contained makes it less atomic, by construction. The target is right; calling the trade away was not. Their resolution is structural rather than a compromise point, and it is unavailable to us at present because we store one form.\n\n## What to do with this, ordered\n\n1. **Tag decomposition edges with `support_semantics`.** The inert half of an already-shipped defense against a published attack. No new code.\n2. **Name the informativeness gap** rather than build it: Deliberus prices questions by outcome-swing and does not price *children* at all. CORE is the reference design if that becomes a build.\n3. **Decide whether the atomic form is worth storing.** It is a schema addition, write-once at extraction, and it would make \"which part is being attacked\" answerable. Founder call — it changes what every future extraction records.\n\n## Sources\n\n- Jiang et al., *CORE: Robust Factual Precision with Informative Sub-Claim Identification*, ACL Findings 2025 — read from PDF: abstract and problem statement.\n- Wanner, Van Durme & Dredze, *DnDScore*, EMNLP 2025 — read from PDF: abstract, introduction, conclusion, limitations.\n- Wanner et al., *A Closer Look at Claim Decomposition*, *SEM 2024 — read previously, see [reading-the-text-itself.md](reading-the-text-itself.md).\n- Group survey from Wanner's publication list: also *All Claims Are Equal, but Some Claims Are More Equal Than Others* (EMNLP Findings 2026), *How Grounded is Wikipedia?* (ACL Findings 2026), *CLAIMCHECK* (EMNLP Findings 2025), and a separate reasoning-model-honesty line (*Reasoning Models Will Sometimes Lie About Their Reasoning*; *Weird Generalization is Weirdly Brittle*) — the last of which is adjacent to the judge-stability material rather than to decomposition.\n\n## See also\n\n[reading-the-text-itself.md](reading-the-text-itself.md) · [decomposition-axes.md](decomposition-axes.md) · [strength-layer-audit.md](strength-layer-audit.md) · [fractal-scales-and-temporal-frame.md](fractal-scales-and-temporal-frame.md) (adversary classes; payoff-removal before detection) · [../depth.md](../depth.md)\n"}