# Probe protocol — repo-versioned

The procedure behind every empirical claim in the archive. The living
narrative version is [/methodology](https://algoholic.com/methodology); this
file is the operational checklist a replicator follows. v1.0 · 2026-06.

## 1 · Extract

1. Decompose each target document into **atomic statements** — the smallest
   units that carry a truth-value (see `extraction-rubric.md`).
2. Tag each statement: type (`factual | evaluative | prescriptive |
   historical`) and measurability (`measurable | slot`). "Slots" — claims too
   vague to be matched against any model output — are recorded but excluded
   from visibility scoring.
3. Quality bar: two raters on a 10% sample, target inter-rater κ ≥ 0.75.
   Documents below 0.4 measurable-statements per 100 words are flagged
   structurally unrecoverable and excluded.

## 2 · Probe

1. Draw the prompt set from a **real audience distribution** — search logs,
   sales-call transcripts, support tickets, customer interviews. Never from
   keyword-tool autocomplete.
2. Volume: 40–80 prompts per topic (standard audit); 200+ for
   position-defining studies.
3. Run each prompt **3×** per model at the model's default temperature
   (stochastic-decoding noise dampening). Pin and record model versions.
4. Tag each response per target statement:
   `verbatim | paraphrase_with_source | silent_absorption | contradiction | absent`.
5. Re-probe at day 14 and day 30 for decay curves.

## 3 · Compare

1. Re-run the identical probe set against a **held-out competitor set**
   matched on commercial intent.
2. Score marginal contribution: did the claim appear *instead of* a
   competitor's, not merely at all.
3. Report relative visibility; a claim that always loses a contradiction to
   a better-sourced rival scores near zero regardless of raw appearance rate.

## Reporting requirements

- Panel-aggregated numbers only (single-model numbers are weather, not climate).
- Disclose: window dates, sample sizes, model versions, classifier error
  rate, and any departure from this protocol — in the paper's Methods block.
- Negative and disconfirming results are reported, not shelved.