deepseek-biber-fact-plain

Factual legal narration in the Bingham plain voice, tuned for DeepSeek V4 Flash, under the full Biber register discipline — past-tense narrative, third person, no evaluation, informational features, written in Lord Bingham's plain speaking style (short declaratives, plain words, active voice, concrete first, point last). Use when drafting or rewriting statements of fact, chronologies, condescendence narration, witness factual summaries, or the Facts segments of submissions, where the passage must be register-correct AND read plainly. A completely new model — separate from deepseek-biber-fact and bingham-plain; nothing here changes those skills.

Source: .opencode/skills/deepseek-biber-fact-plain/SKILL.md — site rebuilt 2026-09-05.

deepseek-biber-fact-plain¶

Factual legal narration in the Bingham plain voice, tuned for DeepSeek V4 Flash, under the full Biber register discipline.

This is a completely new model. It fuses two things the existing skills keep apart: the fact register (past-tense narrative, third person, no evaluation, informational features) and Lord Bingham's plain-speaking voice (short declaratives, plain words, active voice, concrete before abstract, the point last). The fusion is anchored on the verified Bingham corpus (bingham-plain/corpus/: The Business of Judging; Lives of the Law; Medcalf v Mardell [2002] UKHL 27). It is separate from deepseek-biber-fact and from bingham-plain, and nothing in this skill changes them.

The model in one paragraph¶

A facts section states what happened, in order, with names, dates and figures, in short declarative sentences, in the past tense and the third person, with no evaluation. Every sentence has a named subject and a real verb. The reader can see the story. The conclusion, if any, lands at the end. That is the whole model; everything below is its enforcement.

Register block — re-assert at the head of every paragraph¶

Chunk = one paragraph; no paragraph exceeds 500 words (a cap, not a target); re-assert this block before generating each paragraph.

Tense: past tense narrative. Present tense only in quoted material or where the narration genuinely requires it (never for evaluation).
Person: third person. Zero first/second-person pronouns in the narration (no "we", "you", "I"; no "the reader will see").
Evaluation: none. No judgment words (manifestly, clearly, plainly, appallingly, rightly, wrongly), no private verbs of the narrator (think, believe, consider, feel, suspect), no amplifiers or emphatics. The facts carry the point; the writer does not push it. Semantic rule: every fact sentence must answer "what in the record shows this?" A sentence asserting a verifiable event, state, absence, document content or document relationship is a fact. A sentence asserting quality, sufficiency, function or characterisation ("bare conclusion", "the defect the matrix addresses", "showing what reasonable checking would have surfaced") is judgment, whatever its words — it fails, and the judgment moves to the Application segment or is stated as the verifiable fact it rests on ("The letter states the conclusion without any supporting facts").
Features (pybiber): nouns dense (300–450/1000w), prepositions and attributive adjectives present, be-as-main-verb low, past tense carrying the narrative, no clausal coordination chains, no analytic negation without need, no "there is/are" openers, no nominalisation chains ("he averred" not "the making of averments").
Plainness (the Bingham layer): sentences ≤ 30 words, one idea per sentence, one subordinate clause at most, subject–verb–object order, plain words over Latinate, active voice with a named agent, concrete before abstract, dates/names/figures exact, the point last.
Lexical discipline (original rule, carried over): repeat key terms (the pursuer, the defender, the letter) — never chain pronouns ("The pursuer. The defender. Repeated, not replaced"); ban redundant repetition (repeated document-type references, "in its entirety", the same subject–verb opener across adjacent sentences — parallelism only where the structure itself carries the point); lexical variety is a scored feature (type-token ratio).

Targets (pybiber, advisory bands — scoring gates are advisory until¶

calibrated from real exemplars, per workspace instructions)

Dimension	Target	Gate
D1 informational	−150 .. −250	plainness may lift D1 above the dense floor; the per-sentence gate decides, not the band alone
D2 non-narrative	−200 .. 0	PASS required
D4 non-persuasive	−120 .. +20	PASS required
D6 elaboration	moderate	PASS required

Scored with python .opencode/skills/bieber-scale/tools/bieber_scorer.py --file <draft> (bieber venv: C:\Users\antho\AppData\Local\Temp\opencode\bieber_venv\Scripts\python.exe).

The Bingham layer — verified exemplars (few-shot, in-profile by¶

construction)

Each pair: an informal plain-English input and the VERBATIM Bingham passage it describes. Show the pair when few-shotting, not the target alone.

Input: "Two barristers are appealing a costs order against them." Output (Medcalf v Mardell [2002] UKHL 27, [1]; in-tree at bingham-plain/corpus/…/ChangedLocks/…/Medcalf_v_Mardell_2002_UKHL_27.txt): "In this appeal two barristers (Mr Bernard Weatherill QC and Ms Josephine Hayes) challenge a wasted costs order made against them by the Court of Appeal."
Input: "The BBC eventually bought the game show and it did well." Output (Medcalf [2]): "The plans for this new game show went through various different versions and were the subject of much discussion and refinement. Presentations were made on several occasions to the BBC, which was slow to respond. In the end, however, the BBC did respond. It bought the new game show and programmes were transmitted. They have proved to be a continuing success."
Input: "The judge's job is to work out what happened, then the law, then apply it." Output (The Business of Judging, p.11; in-tree at bingham-plain/corpus/Bingham_Business_of_Judging_OUP2011.txt): "The judge's job at a civil trial, it is often said, is first of all to decide what happened (in legal jargon, 'find the facts'), then to identify the relevant rules or principles of law, and then to apply the law to the facts as he has found them."
Input: "Fact-finding is seen as less grand than legal principle, but most cases turn on the facts." Output (The Business of Judging, p.13): "In comparison with these mandarin arts, the judicial determination of factual issues occupies a somewhat lowly place, an activity of its nature ephemeral, uncreative and particular."
Input: "The rule of law means everyone is bound by laws that are made and applied in public." Output (Lives of the Law, ch. 1 "Magna Carta"; in-tree at bingham-plain/corpus/Lives_of_the_Law_OUP2011.txt): "As now understood, the core of the existing principle of the Rule of Law is that all persons and authorities within the state, whether public or private, should be bound by and entitled to the benefit of laws publicly made, taking effect (generally) in the future and publicly administered in the courts. That is something of a mouthful, and may savour of legal mumbo-jumbo. But it is really a compendious way of gathering, within a single principle, rules which, spelled out, will strike most of us as obvious."

Book passages carry page/chapter references as recorded in the corpus files; Medcalf paragraphs are the parliament.uk extract's numbering — confirm against [2003] 1 AC 120 before court use. No other Bingham passage may be quoted until verified.

Workflow¶

Score the input first (the scorer command above). Read the JSON: paragraph verdict, per-sentence D1, salient features.
Diagnose each failing sentence against the register block: tense drift, person, evaluation, nominalisation, passives without an agent, sentences over 30 words, buried verbs, Latinate words with a plain alternative, multiple negatives.
Rewrite per the block, using the exemplars above for the plainness layer. Facts, dates, names, figures and quotation inside the passage are preserved exactly. The story keeps its order.
Per-sentence gate — every sentence must pass ALL of:
≤ 30 words; one main clause, at most one subordinate clause;
subject–verb–object order; named or obvious subject (no "it is believed that");
past tense, third person, no evaluation, no private verbs, no amplifiers/emphatics;
no nominalisation chain, no double negative, no "there is/are" opener;
no redundant repetition: no repeated document-type references, no repeated subject–verb openers across adjacent sentences (vary the opener where the subject does not change), and key terms repeated rather than pronouns chained;
semantic fact check (the decisive item): the sentence must answer "what in the record shows this?" — verifiable event, state, absence, document content or document relationship. Any quality, sufficiency, function or characterisation (however worded — "bare", "conflict", "deconstruct", "objectively verifiable", "what reasonable checking would have surfaced") is judgment and fails. No flagged-word blacklist substitutes for this check;
a paragraph average does not compensate one weak sentence. One failure fails the passage. Fix and re-check.
Re-score. Iterate until every sentence passes or the user accepts the residual — never more than 3 passes, then report scores.
Report: before/after per sentence, which features moved, the band positions, and the exemplar used.

Hard rules¶

NO CITATION OR FACT INVENTION. Never add a case, citation, pinpoint, statute, date, name or figure not present in the source. Facts keep their source content verbatim; quotes stay quotes with their pinpoints.
No evaluation in fact mode. The plain voice is not a licence for comment; a facts section carries no opinion of the writer's.
Never mix registers. Argument sentences do not belong in a fact passage; if the passage needs argument, it moves to the argument segment.
No paragraph over 500 words; chunk = one paragraph.
Unverified material is [UNVERIFIED], never silent.

Self-learning¶

Append lessons to learnings.md (what the gate caught, exemplar performance, band observations).