>_ mohar@portfolio:~/experiments/legal-atomizer$

← all experiments

paragraphs to atomic facts.

Splitting LiveLaw legal-news paragraphs into atomic, standalone legal propositions. Three configurations, toggled below: a three-stage pipeline; a merged stage that answers atomicity, support and coverage in one call and repeats until it stops making new facts; and a stage-1-only version that moves the split into the extraction prompt and does the whole job in one call per paragraph. The first two with mechanical guards no model can bypass. Built to find out where a decomposition pipeline actually breaks.

pipeline

now showing v5 · three stages — extract, then atomicity + split, then support + coverage. Three calls per paragraph.

12 paragraphs / 39 llm calls / 56 facts / 7 depth 2+ / 0 unresolved / 2 gaps / 56 manually verified / 0 manual errors / 0 not yet checked

The manual: correct and manual: incorrect tags are my own reading of each fact against its source paragraph, not a model verdict. inferred marks a fact I accept but the paragraph never states outright. 56 of 56 facts checked so far, 0 of them wrong.

the original paragraph

p811 extracted · 11 final · 0 gap(s) · coverage closed · 11/11 verified by handextremely fine facts

The Court noted that the revised chapter had been approved, circulated to students and was in classroom use. After perusing it, the Court opined that it presented an objective, factually accurate and balanced account of the Judiciary, setting out its constitutional role and safeguards of independence and accountability without either glorifying or unfairly disparaging the institution.

v5 · three stages made 11 facts here — identical claim set to v6 · merged loop

the three stages, one paragraph

One call per stage per paragraph, so 36 of the 39 calls; the rest go to paragraphs that fail the coverage check. Stage 2 nests children under the claim they came from; stage 3 adds a support verdict and the coverage result.

stage 1 · extract (11)

  • p8-l1-f1not judgedmanual: correct

    The revised chapter was approved.

  • p8-l1-f2not judgedmanual: correct

    The revised chapter was circulated to students.

  • p8-l1-f3not judgedmanual: correct

    The revised chapter was in classroom use.

  • p8-l1-f4not judgedmanual: correct

    The Court perused the revised chapter.

  • p8-l1-f5not judgedmanual: correct

    The Court opined that the revised chapter presented an objective account of the Judiciary.

  • p8-l1-f6not judgedmanual: correct

    The Court opined that the revised chapter presented a factually accurate account of the Judiciary.

  • p8-l1-f7not judgedmanual: correct

    The Court opined that the revised chapter presented a balanced account of the Judiciary.

  • p8-l1-f8not judgedmanual: correct

    The revised chapter sets out the constitutional role of the Judiciary.

  • p8-l1-f9not judgedmanual: correct

    The revised chapter sets out safeguards of independence and accountability of the Judiciary.

  • p8-l1-f10not judgedmanual: correct

    The revised chapter does not glorify the Judiciary.

  • p8-l1-f11not judgedmanual: correct

    The revised chapter does not unfairly disparage the Judiciary.

stage 2 · atomicity + split (11)

  • p8-l1-f1atomicmanual: correct

    The revised chapter was approved.

    It is a single proposition. · The text states the revised chapter had been approved.

  • p8-l1-f2atomicmanual: correct

    The revised chapter was circulated to students.

    It is a single proposition. · The text states the revised chapter was circulated to students.

  • p8-l1-f3atomicmanual: correct

    The revised chapter was in classroom use.

    It is a single proposition. · The text states the revised chapter was in classroom use.

  • p8-l1-f4atomicmanual: correct

    The Court perused the revised chapter.

    It is a single proposition. · The text states the Court perused the chapter.

  • p8-l1-f5atomicmanual: correct

    The Court opined that the revised chapter presented an objective account of the Judiciary.

    It is a single proposition with a reporting frame. · The Court opined that it presented an objective account.

  • p8-l1-f6atomicmanual: correct

    The Court opined that the revised chapter presented a factually accurate account of the Judiciary.

    It is a single proposition with a reporting frame. · The Court opined that it presented a factually accurate account.

  • p8-l1-f7atomicmanual: correct

    The Court opined that the revised chapter presented a balanced account of the Judiciary.

    It is a single proposition with a reporting frame. · The Court opined that it presented a balanced account.

  • p8-l1-f8atomicmanual: correct

    The revised chapter sets out the constitutional role of the Judiciary.

    It is a single proposition. · The text states the chapter sets out the constitutional role of the Judiciary.

  • p8-l1-f9atomicmanual: correct

    The revised chapter sets out safeguards of independence and accountability of the Judiciary.

    It is a single proposition with a compound modifier. · The text states the chapter sets out safeguards of independence and accountability.

  • p8-l1-f10atomicmanual: correct

    The revised chapter does not glorify the Judiciary.

    It is a single proposition. · The Court opined that the chapter does not glorify the institution.

  • p8-l1-f11atomicmanual: correct

    The revised chapter does not unfairly disparage the Judiciary.

    It is a single proposition. · The Court opined that the chapter does not unfairly disparage the institution.

stage 3 · support + coverage (11)

  • p8-l1-f1atomicsupportedmanual: correct

    The revised chapter was approved.

    It is a single proposition. · The text states the revised chapter had been approved.

  • p8-l1-f2atomicsupportedmanual: correct

    The revised chapter was circulated to students.

    It is a single proposition. · The text states the revised chapter was circulated to students.

  • p8-l1-f3atomicsupportedmanual: correct

    The revised chapter was in classroom use.

    It is a single proposition. · The text states the revised chapter was in classroom use.

  • p8-l1-f4atomicsupportedmanual: correct

    The Court perused the revised chapter.

    It is a single proposition. · The text states the Court perused the chapter.

  • p8-l1-f5atomicsupportedmanual: correct

    The Court opined that the revised chapter presented an objective account of the Judiciary.

    It is a single proposition with a reporting frame. · The Court opined that it presented an objective account.

  • p8-l1-f6atomicsupportedmanual: correct

    The Court opined that the revised chapter presented a factually accurate account of the Judiciary.

    It is a single proposition with a reporting frame. · The Court opined that it presented a factually accurate account.

  • p8-l1-f7atomicsupportedmanual: correct

    The Court opined that the revised chapter presented a balanced account of the Judiciary.

    It is a single proposition with a reporting frame. · The Court opined that it presented a balanced account.

  • p8-l1-f8atomicsupportedmanual: correct

    The revised chapter sets out the constitutional role of the Judiciary.

    It is a single proposition. · The text states the chapter sets out the constitutional role of the Judiciary.

  • p8-l1-f9atomicsupportedmanual: correct

    The revised chapter sets out safeguards of independence and accountability of the Judiciary.

    It is a single proposition with a compound modifier. · The text states the chapter sets out safeguards of independence and accountability.

  • p8-l1-f10atomicsupportedmanual: correct

    The revised chapter does not glorify the Judiciary.

    It is a single proposition. · The Court opined that the chapter does not glorify the institution.

  • p8-l1-f11atomicsupportedmanual: correct

    The revised chapter does not unfairly disparage the Judiciary.

    It is a single proposition. · The Court opined that the chapter does not unfairly disparage the institution.

coverage

closed · 0 gap(s) found · 0 filled

call ablation

What each stage costs, and what breaks if it is removed. Deltas are arithmetic on the measured 39-call run, not new runs.

ablationcallsdeltawhat is lost
current pipeline39—-
drop the split step36-37 atomic child facts, every compound fact unresolved
drop the verify stage27-12all fidelity and coverage signal
drop the extract stage27-12nothing downstream can run; extraction is generative
a 100M policy clears all-atomic paragraphs33-6this is the saving a local policy would chase