Controlled-language rendering experiment

Status: isolated research prototype, 10 October 2026. The DWP example is an independent experiment. It is not official guidance or an individual entitlement decision.

The screenshot report documents all 14 deliverables and the terminology, MCP and extraction proposals, with 17 real Chrome screenshots, an overall evaluation and sequenced next steps. Download the Word report for an illustrated review copy. The screenshots show retained files in a local documentation viewer; they do not imply a shipped Explorer feature.

The prototype separates language measurements from evidence fidelity. It prepares three optional rendering requests after Explorer or Ask OKF has assembled evidence. Each request contains the same complete evidence and a digest. It never retrieves again, changes the source package or changes the OKF core.

What was measured

The retained run uses three offline illustrative fixtures, with model-assisted authorship in this implementation session. It contains no independently executed model trial, human comprehension study or live MCP invocation. Ask OKF currently returns evidence with ai_answer: null; it does not supply an existing default generated explanation. The baseline fixture therefore uses the existing DWP conditional source reading and unknowns, rather than claiming to reproduce a default model answer.

These are descriptive regex measurements of these particular fixtures. The fixtures were designed to exercise the safeguards. Their difference does not measure a treatment effect, model performance or human readability.

The negative control drops the continued-UC/child-responsibility conditions, the condition that A4361 governs, the limitation on its carer-element examples, and the unresolved historical-law/specialist-review qualification. It keeps every citation label. Its zero heuristic count and intact citation labels do not rescue the lost qualifications.

The bounded feedback replay moves from the baseline fixture to the STE-inspired fixture, then rejects the controlled negative control. It records feedback, input and output identities and a maximum of two revisions. It accepts no explanation. This is an illustrative replay, not evidence of how an LLM behaves when it receives checker feedback.

Inspect the summary, raw baseline checker output, controlled evaluation and loop receipt.

Checker choice and rights

The machine-readable comparison records four pinned repositories and the capabilities actually inspected. The selected SimpleEnglish linter is an MIT-licensed, standard-library Python counter. It exposes JSON and detailed findings. Its 11 categories include sentence length, contractions, modal words and trailing conditions. It cannot establish formal compliance. Its upstream self-test passed.

The experiment vendors only its exact script, original style lexicon slop.tsv and MIT licence. That style lexicon is not the official STE dictionary. Upstream finding positions are approximate and snippets can be shortened. Keep raw outputs, category counts and version bindings when comparing results.

The Python checker offers documented spaCy, glossary, JSON, SARIF and MCP surfaces. Its packaging includes an official-PDF dictionary extraction and its README acknowledges redistribution risk. It would also introduce dependencies outside the current lockfile. Its tests were inspected, not executed.

The Go/Node.js checker has 13 catalogue entries, JSON/SARIF and library interfaces. Its third-party notice also acknowledges risk around its embedded dictionary. A runtime --no-dictionary flag does not remove embedded data from an installation.

The stuffbucket/vale candidate has an inspected stdio MCP entry point and a session vocabulary store. Its licence file records MIT code and an OpenSTE notice. The wordset's origin and rights were not independently established, and the engine was not installed.

The official ASD overview identifies Issue 9, January 2025. The official copyright notice limits reproduction outside listed special-use categories. No permission for this project was inferred from free availability or a third-party code licence. Direct PDF access returned HTTP 403; indexed official copyright text was available. No official dictionary was downloaded or redistributed. The official downloads page also warns against treating plausible AI wording as verified STE compliance.

Frozen DWP evidence

The question is uc-dla-001: highest-rate DLA care from 26 January 2025, reported to UC on 28 March 2026, with assessment-period boundaries absent.

The evidence package retains five complete machine-extracted PDF pages: ADM A4 pages 34, 36 and 37, and ADM F1 pages 17 and 18. It keeps paragraph context, footnotes, continuation, alternative qualifying routes and the rates table. The freezer checked both retained PDF file hashes against the extraction's source identities. The snapshot records the DWP revision, extraction file hashes, page text hashes, original URLs, capture metadata, OGL attribution and extraction limitations.

This is a source-guided presentation subset. A4361 was absent from the earlier natural-language retrieval result. Adding it to this fixed experiment does not establish a retrieval repair. The package remains insufficient. Historical statutory applicability, complete dependencies, independent specialist review, AP boundaries and award history remain unresolved. The current rates table does not establish January 2025 rates.

The existing unreviewed assessor reference is retained separately with its provenance. It never enters rendering requests. Do not mistake it for a reviewed answer key or a fully archived primary-source package.

Architecture and evaluation boundaries

explorer-adapter.mjs accepts the existing governed context envelope or the frozen experimental envelope. For an Explorer context it recomputes the context identity and byte count, then retains the entire package. It exports a request; it calls no provider. Instructions treat source content as inert. Profile and request hashes bind the presentation instructions.

The experiment reuses the existing context evaluator's canonicalisation and SHA-256 functions. A real study-club integration test runs Explorer's existing assembler and A–H evaluator before and after request creation. The assessment and source package stay identical.

The DWP assessor is separate. It reports declared phrase/negation signals, forbidden date/rate patterns, citation presence and digest binding. It maps results to A–H stages without claiming to rerun retrieval or traversal. It does not prove entailment, detect every contradiction, validate legal applicability or exhaustively find unsupported claims. A malicious or contradictory text can contain every expected phrase. Passing its checks therefore always leaves semantic preservation not-established. A missing signal produces a development failure. The result contains no inflated count of “claims preserved”.

The checker flags “may” as a modal finding. Legal modality must not change because of that finding. Amounts, actors, conditions, exceptions, temporal relationships, negation, uncertainty, source references and review boundaries need independent review. Causal relationships and general machine interpretation were not measured.

Terminology, MCP and extraction

Existing Explorer context records already carry stable IDs, labels, aliases, text definitions, scope, authority and provenance. These can supply preferred display terms without changing OKF core. Aliases do not establish semantic equivalence. A future domain sidecar could distinguish permitted variants from ambiguous variants, bound to the existing concept and source identities. This experiment does not invent an assessment-period concept ID or rewrite DWP's separate semantic layers.

The checker adapter exposes the same deterministic Python function and JSON CLI. A read-only, size-bounded MCP wrapper is feasible, but no MCP service was implemented or tested. It should expose measurements and limitations, never a compliance certificate or automatic rewrite. The initial experiment needs no MCP connection.

The pre-extraction stretch hypothesis remains untested. It needs a separate frozen corpus and independent entity/relationship/retrieval controls. Rewriting source text before extraction could change evidence identity and must never replace the original source.

Reproduce and extend

See the prototype commands. The runner is offline, preserves raw checker JSON, and binds profiles, scripts, renderings, evidence and results. CI runs the checker self-test and the experiment's mutation/integration/reproduction tests.

Separate request preparation and response import support later model trials. Imported text is marked unreviewed and its generation is unobserved. The importer checks request/profile/evidence identity, preserves exact text bytes, and emits at most two feedback revisions. It applies this DWP assessor only to the exact frozen DWP package.

Before considering production adoption, freeze independent questions, record actual renderer/model/version/settings and responses, commission specialist semantic assessment, and measure human comprehension with a declared protocol. Use the same evidence for all conditions and report source gaps separately. The current result supports retaining an optional downstream experiment. It does not justify a formal ASD-STE100 profile, an OKF core change or a default Explorer rendering policy.