Quality machinery

Most “AI writing” tooling guarantees the floor: no em-dash tics, no “delve”, no rule-of-three spam. WRITING AGENT enforces that floor and builds machinery for the ceiling - a thesis, a voice, a risk.

The floor: no slop, enforced twice

1. In every prompt - a fixed lexicon of banned verbs, adjectives, transitions, intensifiers, stock phrases, and openers (plus no em-dashes, no fabrications, concrete over abstract) injected into every writer, humanizer, and critic call. The critic flags violations as blocking, not nits.
2. Deterministically - a lexicon-based AI-tell scanner finds offending sentences without an LLM’s opinion. The surgical humanizer rewrites only flagged sentences, with each rewrite guarded (citations, numbers, and length preserved; tell actually gone) before splicing. Approved prose is never wholesale regenerated.

The ceiling

A thesis worth disagreeing with

Articles open with a generated contestable claim - stakes, arguments, a steelmanned counterargument with rebuttal, and non-goals - injected into every writer and critic call. Coverage without advancement is a blocking issue, and the writer is told to engage the steelmanned counterargument head-on rather than dodge it.

An insight bar

The critic scores every draft 1–5 on insight (plus clarity, structure, evidence). Approval requires insight ≥ min_insight - a correct-but-generic draft gets a sharpening revision note instead of a pass. Deterministic structural metrics (paragraph uniformity, rule-of-three density, specificity density) feed the critic as computed evidence it can’t hallucinate around.

Divergent drafts, judged side by side

First attempts sample multiple variants at different temperatures. A dedicated judge reads them side by side and picks the winner - far more reliable than comparing each draft’s jittery 1–5 self-score - and the winner’s biggest remaining weakness feeds the refinement pass. In manual runs you pick instead. Selection pressure for boldness, not convergence to the safest median. (tournament_judge, default on; route the judge node cross-family for an independent eye.)

Claims checked against their sources

For articles, every in-text [N] citation carrying a specific claim - a stat, date, quote, or named attribution - is checked against the actual source text it cites. Under deep research (full page text) an unsupported claim is blocking; with thin search snippets it’s surfaced as a non-blocking nit, so a good draft is never tanked on weak evidence. This turns the critic’s evidence opinion into a structural check. (verify_claims, default on.)

A voice to match

Drop admired paragraphs in voice/ - or /praise a committed unit you loved - and they’re injected into every writer call as a model to match (showing voice beats describing it). Praised passages also feed the learner as positive exemplars.

Write in any field - the craft engine

The same pipeline writes technical explainers, fiction, academic papers, journalism, copy, poetry, or screenplays - because the anti-slop/craft contract is data, not one hard-coded “researcher voice.” A selectable register (one of 11) inverts the rules per genre: fiction keeps the em-dash, academic requires hedging, copy keeps the exclamation. Each register ships few-shot exemplars + a gold style corpus as a default “match this” anchor, plus genre-aware deterministic metrics the model can’t argue with - so it runs well even on a weak model. Pin it with /set register technical (or leave it to infer), and tune field (structure) and citation_style to taste. (plan §22.)

Personas & emotion - composed, not stacked (the compositor)

Voice layers - register ⊃ field ⊃ persona ⊃ emotion ⊃ skills - are resolved by a single compositor that selects one voice and logs why, never piles them on (more layers is worse on a weak model). Choose a persona (one of 46 - 18 archetypes like wry-skeptic / hard-boiled-minimalist, plus 28 public-domain manners like shakespearean / wildean / poe-gothic; no living authors, original-pastiche exemplars) and it’s dropped-with-a-log if it clashes with the register. Choose an emotion (one of 12) and - instead of a cliché-generating “fear = racing heart” dictionary - its anti-cliché deny-list is wired into the detector (“her heart raced” gets flagged) alongside a show-don’t-name cue. Set them with /set persona wildean / /set emotion grief - see Settings. (plan §23.)

A skeptical reader, in the loop

After a run, a table read cold-reads the whole piece as a skeptical target-audience reader - where it gets boring, where trust breaks, what’s never made concrete, what’s missing - into table_read.md. In autonomous runs you can let it close the loop: table_read_revise applies the single highest-impact fix as one bounded, version-snapshotted revision (off by default, since it edits finished prose).

Clean references & figures, owned by the producer

After assembly a pure-Python pass (no LLM, ~0 tokens) takes sourcing and figures away from the writer so they come out clean and consistent:

• one end ## References list ranked by influence (0–100) - weighted cite count plus title overlap with the thesis and headings, dated, most-influential first (rank_references);
• the inline [N] markers are stripped from the body after scoring, so the prose reads clean and all sourcing lives in the end list (strip_inline_citations);
• stray mid-article reference dumps are pulled back out, and generated diagrams are deterministically reconciled into the manuscript - embedded once, de-duplicated (a diagram the model still drew, its self-numbered caption, and a redundant embedded SVG), and anything unused or dropped is logged to rejected.jsonl (reviewable in the web dashboard’s Rejected tab) rather than silently discarded.

Re-run the whole pass on an already-finished piece with polish - references, citations, figures, and a re-export, with no model call.

Evidence report - proof, not vibes

Every article also ships an evidence_report.md: the thesis it argues plus every source ranked by influence (0–100), generated deterministically from the finished manuscript (no model call). It turns the otherwise-invisible trust machinery into something a reader can check. Regenerate any time with evidence (or Project.evidence_report()).

The learning loop

After every finished piece, the learner distills what the critic kept flagging into craft skills (markdown files) and a watch-list injected into future critic calls. It also mines the model’s own preference data - which draft won a side-by-side judging and why, and which revisions fixed a flaw - for the generalizable craft principle behind each. Like critic findings, that data is model-judged, so it enters as candidate skills only and must still earn promotion; it never bypasses the gate. Thirteen built-in craft skills seed the system (seed-skills).

The system gets better with use by accumulating this on-disk memory - not by retraining or fine-tuning the model. The weights never change; what changes is the craft skills and watch-list the agents are handed each run.

Ablation duels - a causal efficacy signal

The old efficacy signal was confounded: every applied skill got the same chapter-level credit, with no counterfactual to say whether the skill actually helped. v2 fixes that. With skill_duels on (opt-in, default off), whenever a unit still has an undecided skill the writer drafts one extra variant with that skill held out - same prompt, same temperature, only the skill differs - and the critic compares it against the full-skill draft. That is a true cause-and-effect (counterfactual) test of whether the skill earned its place.

Promotion and retirement are then driven by a smoothed, sample-gated duel win-rate measured against a 0.5 coin-flip (50/50) - not against the old “first-pass lift” baseline, which is now only the fallback when no duels exist yet. A skill needs ≥5 duels before it can be decided (MIN_DUELS); a win-rate ≥0.55 promotes candidate → trusted (TRUST_WR) and ≤0.45 retires it (RETIRE_WR). Cost is bounded: a draft is only added (never substituted) while a skill is undecided, and the whole mechanism is off by default. /skills surfaces this as a duels (vs 50/50) column alongside first-pass lift.

Distillation & guarded watch-list

• skill_distill (opt-in, default off) - after learning, deterministically retires near-duplicate skills (textual Jaccard similarity ≥ 0.85), keeping the best duel-scored one in each cluster. It is non-destructive (sets a status field; the markdown stays) and keeps skill retrieval sharp as the library grows.
• watch_blocking (default on) - watch-list violations now block only clear, concrete ones; borderline or stylistic hits are downgraded to a nit. Set it false to make the watch-list fully advisory (nit only).