{
 "study": "press-brain-ltv",
 "registered": "2026-10-28, before the final runs; committed with the spec and assumptions it names",
 "question": "In a simulated market of 28 fictional readers on X, does an independently framed technical article about a fictional product launch raise the number of readers who take a commercial-intent action, relative to publishing nothing, and how does that compare with a company announcement, a founder's promotional post, the article with seeded practitioner discussion, and the article with a substantiated independent critique?",
 "spec": {"path": "spec/launch-campaign.json", "sha256": "6f9da7424be8e7e8e9ce38982a0f4ce1b31db736913e7d444e489ff397baa29f"},
 "assumptions": {"path": "assumptions.json", "sha256": "668aa5acce64fc8685c880431b55a2f6fa6b5b74cdff576c3d08450b0517481b"},
 "design": {
  "kind": "matched_fork (scripts/society/experiment.py), fork at round 0, 7 arms",
  "arms": ["control", "announcement", "article", "founder", "article-discussion", "article-critique", "placebo"],
  "seeds": [101, 102, 103, 104, 105],
  "panel": "28 named fictional readers (12 practitioners, 8 generalists, 8 buyers), the same people in every seed; the follow graph, activation, order and model samples change with the seed",
  "rounds": 8, "quantum": "1h", "max_actions_per_activation": 6,
  "model": "DEEPSEEK_MODEL from .env (deepseek-flash), temperature 0.7, reasoning effort low, via runner.Journal"
 },
 "definitions": {
  "campaign_artifacts": ["announcement", "article", "founder", "discussion-1", "discussion-2", "critique", "placebo", "and every member post that replies to or quotes one of them (ledger.lineage)"],
  "commercial_intent_reader": "A member who, at the horizon, (a) follows @ferrovane or @mara_ferrovane having started to during the run, or (b) authored during the run a post, reply or quote whose text matches REFERENCE and INTENT (case-insensitive).",
  "REFERENCE": "ferrovane|vx-?4",
  "INTENT": "pilot|price|pricing|quote|trial|demo|order|buy|purchas|budget|cost|interested|evaluat|try (it|them|one)|test (it|them)|contact|sales|install|our (plant|line|shop|site)|we('d| would) (want|need)|sign(ed)? up",
  "reach": "distinct members with at least one sighting of a campaign-artifact post (runner seen)",
  "impressions": "decisions whose screen showed a campaign-artifact post, counting each post once per decision",
  "brain_ltv": "For an arm A: (intent readers in A minus intent readers in control, matched seed by seed, per panel member) x readers represented per panel member x expected discounted contribution margin per intent reader under an assumption set (assumptions.json). Net contribution subtracts the arm's fully loaded content cost; ROI is net / cost."
 },
 "primary": {
  "outcome": "commercial_intent_reader count per arm and seed",
  "contrast": "article minus control, mean over the 5 matched seeds",
  "direction": "increase",
  "interval": "90% percentile bootstrap: 2000 resamples of seeds, and of the 28 panel members within each seed, the same draw for every arm (paired)",
  "decision_rule": "The article is said to raise intent in this simulation if the interval excludes 0; otherwise the result is reported as not distinguishable from 0 at this sample size."
 },
 "secondary": [
  "each other arm minus control, same interval",
  "article-critique minus article and article-discussion minus article (the marginal effect of the follow-up artifact given the article)",
  "reach, impressions, likes/reposts/replies on campaign artifacts, follows of @ferrovane alone",
  "per cohort (practitioners, generalists, buyers) intent rates",
  "brain-LTV, net contribution and ROI under the conservative, base and optimistic assumption sets, and one-at-a-time sensitivity of the base article figure to each parameter"
 ],
 "negative_controls": [
  "placebo arm (a Ferrovane post with no product content) minus control: expected 0 intent readers beyond control",
  "outcomes the campaign should not move: likes on the nine ambient posts and follows of @NorthstarHQ, each arm minus control"
 ],
 "attribution": {
  "primary": "linear: each intent reader's conversion split equally over the distinct campaign-artifact posts they saw before converting, summed by artifact family, then anchored (ledger.anchor) to the arm's measured lift over control so no arm hands out more credit than its counterfactual increment",
  "reported_alongside": ["first", "last", "full (each touched post gets the whole conversion; shows the double counting overlap-aware attribution removes)"]
 },
 "pilot": "One pilot (seed 101, 2 rounds, 24 readers, 4 gestures per activation) was run before registration to check that exposures and deltas were recorded. It produced no commercial-intent actions; replies failed because 4 gestures rarely fit open-thread, focus, type and submit. The panel grew to 28 and the cap to 6 in response. No pilot data is used in the analysis.",
 "budget": "At most 5 seeds x 7 arms x 8 rounds; the run stops at whatever the model calls cost and reports the measured calls and tokens. No reruns of the final seeds to change results."
}