Contract pricing: from linear-in-views, through convex-in-views, to convex-in-points
Status: decided and implemented (#397, #409).
model/pricing.tscarries the points-based formula below,PRICE_EXPONENT = 1.7, the coefficient derived in the file from the same 11-giant-team target stated here, andTIER_DAYSat 3 / 7 / 14. It importsbasePointsfrommodel/scoring.tsrather than restating the curve, so a change to the scoring shape reaches price with nothing to keep in step. Both the browser and the Worker price from that one module.
ContractPrice has gone through three iterations. This ADR is the single canonical record, superseding revisions are folded in here rather than left as a chain of separate ADRs, so there's one place to read the current formula and the reasoning that ruled out the alternatives.
Current formula
BasePoints(v) = max(0, log2(v / 2000)) if v ≤ 150,000 # scoring-system.md §3
6.23 + (v − 150,000)/50,000 if v > 150,000
ContractPrice = D × BasePoints(Normalized_30dAvg_Views)^k × contract_daysPriced on the 30-day average, never daily views (unchanged since the original design, the anti-jackpot guard: a 1-day spike barely moves the 30-day average, so trades only pay off on sustained multi-week momentum). BasePoints here reuses the scoring curve's exact constants (2000 zero-point, 150k kink, same linear-tail slope) fed the 30-day average instead of daily views, one curve, one place to tune the shape. k and D are pricing-only constants, tuned independently of the scoring curve.
Locked: k = 1.7, calibrated so an 11-giant team costs 1,800 credits at the SHORT (3-day) tier ⇒ D ≈ 2.577. Starting budget: 1,000 credits (see "why not 10,000" below).
| Band (views) | SHORT (3d) | MEDIUM (7d) |
|---|---|---|
| niche, sub-anchor (3,000) | 3.1 | 7.3 |
| rank~1000 (9,000) | 28.9 | 67.3 |
| mid-tier (25,000) | 69.7 | 162.5 |
| giant (130,000) | 163.6 | 381.8 |
| top-3 (300,000) | 338.1 | 788.9 |
| viral/event spike (638,000) | 860.5 | 2,007.9 |
- 9 mid-tier + 2 giants (full 11-slot team) @ SHORT = 954.1: matches "1,000 credits buys a full team with a couple of giants" almost exactly.
- Full 11-giant team @ SHORT = 1,800 (1.8× the 1,000-credit budget): gated behind grinding.
- Contract-duration tiers, locked in days: SHORT = 3, MEDIUM = 7, LONG = 14.
How we got here
v1, linear (superseded): Normalized_30dAvg_Views / 1000 × contract_weeks. Let a day-one player assemble an all-giants 11-slot team for as little as 471 credits at SHORT, under half the starting budget, on day one, zero grinding. Because scoring is cumulative and never resets mid-league, that's a permanent lead. Rejected: giants were exactly as credit-efficient as everything else, so "build toward all-giants" stayed the dominant strategy at every tier, just delayed by budget.
v2, convex-in-views (superseded): C × NormalizedViews^1.5 × contract_days, anchored to preserve the legacy ~9-credit rank~1000 price. Fixed the day-one snowball (giants now cost progressively more per marginal view than mid-tier articles), but live data on real articles exposed a new problem: BasePoints spreads only ~5.7x across the whole competitive range (anchor to viral) while ContractPrice spread 249x, price diverged 44x faster than the value it buys. Root cause: pageviews are Zipfian, so any convex function stacked directly on raw views front-loads nearly all price differentiation onto the extreme head of the distribution (~rank 1–40), leaving rank ~40-down-to-anchor nearly flat. That's what produced "top tier costs way too much, mid tier way too low, all the way down."
v3, convex-in-points (current): apply the convexity to BasePoints (already log-compressed, ~5.7x spread) instead of to raw views (~249x spread for the same range). This gives direct control over the top-to-anchor price ratio via k alone, without a convex-in-views formula fighting the underlying power-law skew. Two options were rejected en route:
- Price = points directly (k=1): removes the anti-snowball property entirely: credits-per- point becomes constant, re-opening the exact all-giants-day-one bug v1 had. Points' own anti-snowball (synergy + log-compression) balances score composition across archetypes; it doesn't gate affordability on its own, so price still needs its own convexity knob.
- Anchoring at the low end (like v2 did): rejected for v3: because points only span ~27x total, any
kthat also preserves the old low anchor leaves even a full 11-giant team costing 118–327 credits, nowhere near gating against a 1,000-credit budget. The anchor must be set from the top (fix the full-11-giant-team price, deriveDfrom it) instead.
Why starting budget is 1,000, not 10,000
v2's CREDIT_SCALE × 10 / 10,000-credit budget existed only to fix a rounding artifact: the views^1.5 formula's tiny coefficient made many low-view articles round down to 0 integer credits. Verified this does not reproduce under the points-based curve, BasePoints already intentionally floors at 0 for sub-2,000-view articles (the documented niche/synergy archetype), so near-zero prices there are correct design, not a resolution bug. No hidden internal fixed-point scale either, credits are integers everywhere already (team balance column, stipend accrual, fee math); threading a hidden multiplier through all of that for a display-only difference isn't worth it. Reverts to the simpler, already-established 1,000, matching the explainability principle for a casual-core audience.
Consequences
model/pricing.ts,docs/domain/scoring-system.md§6.1, andCONTEXT.md's Contract Price entry all need updating to this formula, not yet done as of this ADR.- The Language Scale Factor
L(ADR 0002) still enters superlinearly, now asBasePoints(rawViews × L)^k, same caveat as before: whoever calibratesLfor other languages must account for the exponent, not just the raw view-volume ratio. - Grind-timeline re-derivation (the "~3 weeks to a full giants team, active trader vs. passive saver" narrative) under this curve is still unvalidated, same caveat v2 already carried, inherited here too.
- Still open, raised alongside this decision but not yet resolved: the "sold to system if nobody renews at contract expiry" payout mechanic, and the trending-spike resale risk (buying near a rising 30-day average that may already be about to mean-revert down).