flemming-it
7dcdf7679f
Regex-based extraction of atomic deontic statements (Pflicht,
Verbot, Erlaubnis) from German legal text. Reuse-lens: contracts,
ISO standards, internal policies. Vorbild: Mercatus QuantGov /
RegData.
Pattern coverage in v0.1.0:
- OBLIGATION: muss, hat zu, ist verpflichtet, ist anzuzeigen,
aufzuzeichnen, muss eingehalten/geführt/...
- PROHIBITION: darf nicht, dürfen nicht, ist untersagt, verboten
- PERMISSION: kann, können, darf
Each emitted Duty carries a confidence ∈ [0, 1]. Pattern hits
default to 0.85; ambiguous structures (bare "kann") drop to 0.70.
Anything below 0.6 is the surrounding flow's signal to re-run
through llm.chat for human-readable refinement and feed the result
to system.approval@^0 — that keeps deterministic extraction
separate from non-deterministic LLM judgment.
Pure in-WASM (regex), zero filesystem, zero network.
Reserved for next versions:
- 0.2: EN / FR locale variants
- 0.3: addressee + frequency + authority detection (Phase 1
relies on flow-side LLM enrichment for those fields)
Signed-off-by: flemming-it <sf@flemming.it>
40 lines
1.2 KiB
YAML
40 lines
1.2 KiB
YAML
schema_version: 3
|
|
provider: chain
|
|
name: text-deontic-extract
|
|
version: 0.1.0
|
|
|
|
# Extracts atomic deontic statements (Pflicht / Verbot /
|
|
# Erlaubnis) from a normalised norm. Reuse-lens: useful far
|
|
# beyond law — contracts, ISO standards, internal policies.
|
|
#
|
|
# v0.1.0 ships a regex pattern list for German legal text
|
|
# (must / hat zu / ist verpflichtet / darf nicht / kann). Hybrid
|
|
# LLM-assist happens in the FLOW via a separate llm.chat step
|
|
# that this module's output feeds. Keeping LLM out of the module
|
|
# preserves determinism + tier markings.
|
|
#
|
|
# Vorbild: Mercatus QuantGov / RegData (US).
|
|
provides:
|
|
- capability: text.deontic_extract
|
|
version: 0.1.0
|
|
|
|
inputs:
|
|
norm:
|
|
type: json
|
|
description:
|
|
en: NormalizedNorm (output of text.akoma_normalize).
|
|
de: NormalizedNorm (Output von text.akoma_normalize).
|
|
|
|
outputs:
|
|
duties:
|
|
type: json
|
|
description:
|
|
en: |
|
|
JSON-encoded { duties: [Duty], confidence: 0..1 }.
|
|
Each Duty: { modality, addressee, action, frequency,
|
|
source_norm, confidence }. Low confidence (<0.6) flags
|
|
an LLM-assist round; the surrounding flow handles it.
|
|
de: |
|
|
JSON-codiertes { duties: [Duty], confidence: 0..1 }.
|
|
|
|
permissions: []
|