Give the article channel its own length band and a holdout eval

[telivity-otaip]

What this changes

The article channel could already outline, draft sections and stitch them. Two things were missing: it took its length from the post band, and nothing measured whether the result sounded like the author.

Length comes from articles now

style_profile measures linkedin_article pieces separately from linkedin_post pieces. The article aim is the median of the author's own articles (clamped 600–3000 words), the ceiling is p90, and the aim is divided across the outline into a per-section budget (clamped 180–600 words) so a six-section article asks for six sections rather than six posts. With no articles in the corpus the aim falls back to a stated 1100-word default instead of borrowing the post band.

Three mechanical defects surfaced while wiring this up and are fixed here:

• article retrieval was drawing from the whole corpus, so posts could become the rhythm reference for an article section; it is restricted to linkedin_article pieces.
• the corpus floor counted pieces on disk, so a carve large enough to empty article retrieval still passed; it now checks the voice index too and errors instead of quietly drafting from nothing.
• independently drafted sections sometimes restated each other; near-duplicate sections are dropped before stitching.

The eval

personality-protect select-article-holdouts --apply
personality-protect index-voice --from-carve
PP_MLX_ALLOW=1 personality-protect eval-write-article --out receipt.json

The carve is deterministic, keeps previously carved ids pinned, and refuses to drop the voice index below its five-article floor. Each holdout is reduced to a lossy brief — a topic plus 3–6 bullets taken one per segment of the piece so the brief describes the shape of an article rather than its densest paragraph.

The post miner could not be reused unchanged. Its overlap cap is a share of the source, and a 25% share of a 1500-word article is a 375-word "brief"; the article cap is an absolute 60 words, so a longer source does not buy a longer head start. Both arms then write that brief under the same outline, budget and trim — the product arm with exemplars and the style card, the control arm with neither — so distance measures the voice machinery instead of the scaffolding they share.

Real MLX run: article voice is not proven

n=4 holdouts, Qwen3.5-9B-4bit, adapter none.

run	wins (article/base/tie)	disqualified	p	verdict
k=2 (default)	0 / 0 / 4	4 article, 4 base	1.0	not_supported
k=0 (no exemplars)	0 / 0 / 4	4 article, 4 base	1.0	not_supported

Every draft in both arms was disqualified for inventing entities or figures, which scores as a tie, so distance never decided a single item. That is the run's actual finding and it is about the product, not the harness: the model cannot currently expand a 60-word brief into a ~1300-word article without fabricating specifics, and the control fabricates just as much. The invention guard is left at zero tolerance — loosening a factuality gate to obtain a verdict would manufacture the result the gate exists to prevent.

Because "everything was disqualified" and "the voice arm lost on cadence" both produce zero wins, the receipt now distinguishes them explicitly (distance_ever_decided).

Underneath the disqualifications the exemplars do move cadence in the right direction — the article arm is closer to the holdout at k=2 than at k=0 on all four pieces — but that is a within-arm observation, not a result against the control, and it is not a claim this PR makes.

Notes

• write is unchanged for the post channel and still runs base weights.
• No LoRA retraining; the writer adapter work stays on De-voice writer SFT pairs and widen the ship-gate holdout #48.
• Receipts carry ids, distances and flags only. Tests use Contoso fixtures.
• 364 tests pass; ruff check clean.

Made with Cursor