Unit Economics — pricing, COGS, and margin model

Current pricing model. Two-sided, and the two operator products have fundamentally different margin shapes because they differ on who pays the providers:

	License $3.99/wk	Hosted $5.99/wk
Runtime	self-hosted (Electron)	managed always-on VM
Provider keys	the Director's (BYOK)	ours — usage included
Revenue/mo	$17.29	$25.96
Fixed COGS/mo	$2.01 (Stripe + control plane)	$4.79 (+ VM $2.47)
Provider COGS	$0 — never touches us	variable, we absorb it
Margin	flat ~88%, no tail risk	$21.16/mo of provider headroom, then negative

License is the safe product: margin is 88% regardless of how hard the operator runs the Program, because their spend is billed to their own account. Hosted takes on real, unbounded liability for $8.67/mo more revenue — that gap buys the VM and every token and crawl. It clears comfortably on a typical 3–5 domain Constellation (~$1.69/domain/mo → breakeven near 13 domains) but goes negative on a heavy operator, and there is currently no cap: cadence, breadth and cycle count are all user-controlled. §7 carries the exposure math and the allowance proposal.

(2) subscriber rev-share — the platform takes a ~18% cut (range 15–20%) when an operator charges their audience, ~pure margin (Stripe on the sub is the Director's, via Connect).

How to read the cost tables below: the per-action / per-token figures are the operator's provider spend. On License that is their bill and never ours; on Hosted it is our COGS. The older $20/seat + allowance + metered-overage framing in §5–6 was shelved when everything was BYOK — it is live again for Hosted, which is exactly the flat-fee-plus-allowance problem it solved. The live model is in src/lib/core/economics/unit-economics.ts (run npm run economics).

Purpose. The grounded cost surface of Meridians — every external service the engine touches, its real 2026 price, the per-action token cost of every AI operation, and a worked proposal for token-based ("credit") pricing across the whole app. This is the quantitative companion to the manifesto's Business Model section — which now sells the shipped $3.99 / $5.99 weekly story, matching src/lib/server/gateway/pricing.ts. This file is where the superseded $20/seat proposal was tested against the cost surface; §5–6 keep it for that reference only.

Canonical source of truth for prices: src/lib/core/constants/ai-pricing.ts (MODEL_PRICING) — that is what the app actually bills calls against via calculateApiCost. Where the manifesto and the code disagree, the code wins (noted inline).

The simulation is now a cost unit

For a causal-simulation user, one situation carried through world construction, forked arms, Meridians World Search (MWS) rounds, readings, and debrief is a more useful cost unit than a research crawl. See the simulation leaf. Program crawl economics still apply to Directors who ground worlds in live evidence; they are no longer the modal bill for the participant, reader, learner, or author shapes.

---