{"path":"research/reading-the-text-itself.md","content":"# Reading the Text Itself — The Analytic Half of Premise Recovery\n\n**Date**: 2026-08-27\n**Occasion**: The founder's correction of a phrasing of mine. I had written that the unwritten structure is recovered *\"not by reading the text harder, but by reading something other than the text\"*, and he answered that it is **also** — and centrally — a matter of reading the text itself with philosophical, linguistic and semantic care, dissolving a compound message into its extricable aspects so each can be judged on its own merits.\n**Status**: he is right, the phrasing was wrong, and the literature is more supportive of his position than the corpus had recorded.\n\n---\n\n## 1. What was wrong with the phrasing\n\nThe contrast I drew was between *reading harder* and *reading something else*, and it smuggled in a strawman: \"harder\" implied **more effort at the same operation**. What he means is **more apparatus** — theory-informed analysis of the same sentence. Those are not the same thing, and collapsing them made the analytic half of the project sound like the part being set aside.\n\nThe honest trichotomy:\n\n| | What changes | Finds | Built? |\n|---|---|---|---|\n| **Same object, better apparatus** | the analytic instruments brought to the sentence | smuggled modifiers, presuppositions, thresholds, frames | **the largest gap** |\n| **Different object** | the pair; the vocabulary distribution across two texts | classificatory and frame-level premises | cross-source pass built and run |\n| **Different question of the object** | what the descent asks | the deleted counterweight in a weighing | shipped in the question set |\n\nThe first row is where the founding conviction lives. It is *close reading with instruments*, and the eight axes in [decomposition-axes.md](decomposition-axes.md) are its inventory.\n\n## 2. The measured reason the two halves cannot substitute for each other\n\nThe cross-source run produced **eleven premises across four adversarial pairs, and every one is frame-level** — the unit of moral analysis, the reference class, the burden of proof under uncertainty, the governing body of law, the market model. **Not one is a within-sentence smuggled modifier.**\n\nThat is not a coincidence and not a shortcoming. The pass reads **stored claims**, not raw sentences, so an evaluative adjective fused into a single claim is invisible to it by construction. Symmetrically, no amount of close reading of one text surfaces a premise whose absence is only legible against the *other* text.\n\n**The two mechanisms are structurally blind to each other's finds.** Neither is the whole method.\n\n## 3. What the literature actually says, read rather than summarised\n\nThe corpus had recorded the decompose-then-verify results as a caution against maximal decomposition. Reading the primary source changes that, and in the founder's favour.\n\n**Wanner, Ebner, Jiang, Dredze & Van Durme, *A Closer Look at Claim Decomposition*** (JHU; the group whose DnDScore the pipeline is built on) propose **an LLM-based decomposition guided by Bertrand Russell's logical atomism and neo-Davidsonian semantics**, and report that it *\"gives far more subclaims than other methods while maintaining high coherence with the claim being decomposed.\"* It wins their decomposition-quality measure.\n\n**So the apparatus is theory in the prompt, not a parser replacing the model.** Their pure-syntactic arm — PredPatt over a dependency parse — scored **worst** on subclaim support (80–94% against 93–99% for the LLM methods), because syntactic sub-structures are often not grammatical sentences and drift from what the original said. A machine-readable warning against the intuition that a formal parser would be the rigorous option.\n\n### The vocabulary the corpus lacked, and should adopt\n\nThey name **three independent quality dimensions**, and the Babbage example teaches all three at once:\n\n| Dimension | Failure looks like |\n|---|---|\n| **Coverage** | *\"Babbage was a mathematician\"* alone — the sentence also said philosopher, food critic, French. Claims omitted. |\n| **Coherence** | adding *\"Babbage was an engineer\"*, *\"a dancer\"* — content the original never asserted. Invention. |\n| **Atomicity** | *\"Babbage was a French mathematician\"* + *\"a philosopher and food critic\"* — separable things still fused. |\n\n**Low atomicity is a named failure mode**, which settles a misreading this corpus briefly carried. *Molecular, not atomic* is about **decontextuality** — do not strip the context a reader needs — and says nothing against pulling separable aspects apart. Atomicity and decontextuality are two axes, and the target is **maximal separation with every piece still standing alone**. That is the founding position (*\"until we aren't muddling properly separable aspects of reality that can and should be judged separately\"*) arriving from the measurement side.\n\n**Coverage is the dimension Deliberus has no instrument for at all.** The completeness oracle checks whether a claim's *supporting structure* is exposed; nothing checks whether a decomposition dropped something the parent asserted. That is the recomposition check already proposed in [decomposition-axes.md](decomposition-axes.md) §5 as a colimit condition, and it now has a name and a benchmark from outside.\n\n## 4. Where this leaves the build\n\nThe corpus has dealt with this in depth and the depth is the point — the pieces are specified and unbuilt rather than unconsidered:\n\n- **The predicate–argument axis is the largest gap**, and its corpus example is exactly the founder's case: *\"minimum wage laws cause unintended and **unnecessary** hardship for the low-skilled workers the laws are intended to assist\"* — `unnecessary` is a value premise smuggled inside an empirical claim, and a reader can accept every fact while rejecting it. Nothing in the pipeline surfaces it ([decomposition-axes.md](decomposition-axes.md) §3).\n- **Theory in the prompt is the shape that works**, per Wanner et al. — a prompt-layer change rather than a new subsystem, sitting beside the standing rule that the prompt layer carries the project's practical wisdom. **BUILT 2026-08-27**: `deliberus/extraction/decompose_claim.py`, with both theories' *operational tests* in the prompt rather than their names, both claim forms kept per part, and coverage checked by a separate call. First live run recovered the smuggled `unnecessary` modifier this doc uses as its example, plus an attribution shell the axes table had not listed. Not wired to an endpoint — machine-proposed decomposition entering the correction UX is a founder decision.\n- **The strength-layer guard comes first regardless**: an unaligned split still multiplies energy (one claim at 0.70 contributes 0.06; twelve children contribute 0.72), so evidence routing and `support_interpretation` tagging precede any new axis. That ordering is about the *number*, never about the knife.\n\n## Sources\n\n- Wanner, Ebner, Jiang, Dredze & Van Durme, *A Closer Look at Claim Decomposition*, [arXiv:2403.11903](https://arxiv.org/abs/2403.11903) — read in full from the PDF: abstract, method (§4.2 shallow semantic parsing, §4.3 prompting-with-parse), the coverage/coherence/atomicity framing, and the per-method support tables.\n- White et al. (2016), Zhang et al. (2017) — PredPatt. Goyal & Durrett (2020) — dependency-arc decomposition.\n- Gunjal & Durrett (2024), *Molecular Facts* — decontextuality versus minimality, the axis distinguished above.\n\n## See also\n\n[decomposition-axes.md](decomposition-axes.md) (the eight axes and the alignment warning) · [cross-source-premise-run.md](cross-source-premise-run.md) (the other half, measured) · [the-load-bearing-unsaid.md](the-load-bearing-unsaid.md) (why the structure is missing in the first place) · [scheme-bounded-decomposition-and-evidence-as-subgraph.md](scheme-bounded-decomposition-and-evidence-as-subgraph.md) (the three auto-decomposition layers and the pruning insight) · [../depth.md](../depth.md)\n"}