Evidence-Quoted Text Coder — Bring Your Own Labels

Codes text against your own labels — and when it can't, it tells you what it can't, instead of guessing.

Give it a codebook (your labels, each with a few marker phrases) and some passages. For every passage it either returns the label with the verbatim quote that earned it, or it hands back the named residual: the two labels a passage fell between when the readings split, or an honest abstention where no quote held. That "I can't tell" isn't silence — it's a row you can read, and it's usually where your own labels overlap, which is the real finding a confidence number would have destroyed.

No AI calls. No training step. No model weights. No API keys. Nothing leaves the container.

Why it's different

Most classifiers always answer — they hand you a label and a probability even when there's nothing there. This one is built the other way: each passage is read two independent ways, and a label is emitted only when both readings agree, nothing else competes for it, and the evidence is a verbatim span — not a vibe. When the readings disagree, it says so instead.

It needs no model and no training because everything it reads is what you handed it — your labels, their markers, your passages. Nothing is carried in from anywhere else, which is exactly why it cannot tell you anything your own evidence doesn't already carry. Run it twice on the same input and you get the same rows.

Your markers are the gate. A label's marker phrases are what can refuse it: a passage containing none of them is refused for that label, which is why every coded row can carry a quote. A label with no markers cannot be refused by its own words — it is admitted on agreement alone, so its rows carry a weaker guarantee than the rest. The run warns you which labels those are.

Input

Field	What it is
codebook	Your labels: matching_modes → {dimension: {label: {lexical_markers: [...]}}}. Optional theme_definitions.
passages	The text to code — strings, or {text, source_url, source_title}.
datasetId	Optionally pull passages from another Actor's dataset.
question	The question you're asking of this text. Stamped on every row, so an audit can be checked for whether it answers the question actually asked.
exemplars	Passages you've already confirmed, keyed "dimension::label". Read as more of that label's evidence, so it sharpens. Nothing is stored between runs — they stay yours.

Output

One row per passage per dimension — the codings and what could not be coded, side by side.

{ "outcome": "coded",         "dimension": "...", "label": "...",   "evidence_quote": "<your passage>", ... }
{ "outcome": "fell_between",  "dimension": "...", "between": ["autonomy", "flexibility"], ... }
{ "outcome": "no_quote_held", "dimension": "...", "label_considered": "earnings", ... }

• coded — both readings agreed and the label's words are in the passage, which is quoted.
• fell_between — the two readings named different labels. Not a tie to be broken: usually it is where your own two labels overlap in this material, and it is the finding a confidence score would have hidden.
• no_quote_held — a label was reached for and not one of its words is in the passage, so it was refused.
• no_reading — nothing here to read.

Every row carries a lens tag (the labels-and-markers it was read through — unchanged when you add confirmed examples, so a session proves same frame, more ground), a provenance tag anyone can recompute, and a second independent egress tag from the outgoing door. CODING_REPORT in the key-value store holds the totals and open_addresses — which label pairs your passages kept falling between. A live read-only window shows the run's shape as it forms; counts and label-pair names pass the glass, never your text.

Pricing

One meter, one price. You pay per row — and a residual row (fell_between, no_quote_held) costs exactly the same as a coded one, because the residual is half the deliverable, not a discount.

Meter	When it fires	Price
passage-coded	Once per emitted row — a label with its verbatim quote, or the named residual	$0.05 / row ($50 / 1,000)

Count your rows before you run. A passage is read once per dimension, and each reading is a row. Ten passages against a codebook with three dimensions is thirty rows, not ten — $1.50, not $0.50. That is the honest arithmetic and it is better to have it here than in an invoice.

Silence is free: an empty passage forms nothing and is never billed. There is no per-run fee and no subscription — this boots a standard-library process, and a startup charge would be a number nothing derived. You pay only for rows that formed, and never for a guess, because it never makes one.

How this price was derived

Priced by derivation, not by market-positioning. A careful human coding text with evidence runs $0.20–$1+ per item; this codes deterministically — with the verbatim quote and the honest abstention no probabilistic model can offer — at $0.05/passage, an intro rate well under the human labor it stands beside, normalizing toward the fair recovery rate over time. The competitors bill a subscription for a probability; you pay per formed row for a guarantee.

The discipline

• Fail-closed: the coder proves its own contract before it reads a single passage.
• Model-free: pure standard library. No numpy, no ML runtime — you can list the container and see there is nothing in it to trust.
• Yours: the codebook is your meaning-structure. The engine never speaks meaning you didn't author; it only finds, or fails to find, your labels' evidence in the text.