A parked note (T-536, Jordan, 1 Oct (Pacific), via the design session). Nothing here is planned or built because of this note. It's the long view, so the cards we do build point the same way. Each step stands on the ones before it. For each step: what it is, the maths it needs, a rough size ("weeks" means a few cards, "months" means a project), and what it unlocks. Our words throughout; never Games Workshop's rules text.
1. The meta: what wins
What. Which armies, detachments, dispositions and units the winning lists take, and how that changes month to month. Has today. grimstat-corpus lists (CC BY 4.0), inclusion.json, metatrack.json, win rates when a source allows them (T-411), the Meta hub (T-312), the patch pulse (T-519). Maths. Counting, shares and their doubt (lists × points, a binomial interval). An army's lists are the denominator, never the whole field. Size. Weeks, mostly sources: more lists means less doubt, so the work is permissions (T-412), not maths. Unlocks. Every step below is checked against it. It's the referee.
2. The mission: what scores
What. Each Force Disposition's primaries and the secondaries, as the VP each job can earn: Hammer kills, Anvil holds, Runner moves and does actions, Banner makes the others better. Has today. The first reading: reports/tiers/2026-10-02-dispositions.md (T-537). The demand per disposition, and a check against what each disposition's lists carry, which fails for now: the army decides the disposition more than the mission decides the list. Maths. VP per scoring line × the times a game offers it, sorted by job. Then a game model's numbers (two objectives, a kill and a half a turn) are replaced, step by step, by measured ones (step 5). Size. Weeks for the table; months to measure it. Unlocks. Job weights with a reason behind them (step 3), and the question "what does my primary pay for?" answered on the page.
3. Scoring: what each unit is worth at its job
What. A unit's letter and its job tag: Offense and Defense as the raw pair, the four jobs from them, quotas by job (T-523). Has today. Value, the live letter; V1, V2 and V2b in the Lab (T-524); every formula in design/scoring.md (T-535); the Banner from effects (T-527); army rules (T-530); archetypes (T-529); Lenses carded (T-531); the calibration score carded (T-533). Maths. Expected damage (ev.js), hits to destroy, percentiles, a composite of peak and breadth, quotas, Spearman against list share and public tier lists. Size. Weeks per piece; it's where most of the work has gone. Unlocks. A letter players argue with for the right reasons, and the lens: "rank them for my disposition" or "for the meta I expect".
4. List building: what goes together
What. From one unit's worth to a list's: synergy (a leader and its squad, an aura and who's in it), coverage (enough anti-tank, enough bodies for the objectives) and the points. Has today. Best legal loadout (T-225), Tune your list (T-414), joined characters (T-145), the Banner method's "with and without". Maths. A list as a bundle: coverage against the archetypes (each target's kill turns), job demand from the disposition (step 2), synergy pairs from the effects files. Then a search over swaps under the points limit (greedy, then a local search; the full problem is a knapsack with interactions, so no exact answer). Size. Months. Unlocks. "Your list is light on Anvil for Take and Hold; these three swaps fit the points." The first piece of advice that isn't about one unit.
5. Testing: does the list do it?
What. Play a list's turns against a target list on a simple board and count what it scores, not only what it kills. Has today. The calculator's volley and To the death (T-214), Army vs Army (the comparator), the game plan sheet (T-159), King of the Hill's bouts (T-521). Maths. Monte Carlo over turns: who reaches which objective when (move, Advance, charge distances as dice), who survives to hold it, the primary's lines scored each turn. The step-2 assumptions get measured here. Size. Months, and it needs a board model that stays simple (objectives as points, distances as turns, terrain as cover or not). Unlocks. "Against that list, yours scores about 32 primary"; and the step-2 numbers measured, not assumed.
6. The whole game: both players' choices
What. Both players choosing each turn (where to move, whom to shoot, what to hold), not a script. Maths. A game tree is far too big; what fits is a policy for each side (a handful of simple rules: hold what you can, shoot what threatens you, screen the home) scored by simulation, and the policies tuned against each other. Expected VP and win chance per matchup, with their doubt. Size. Months to a year. The hardest step, and the one most likely to be wrong in ways that look right; it stands or falls by step 1's referee. Unlocks. Win chances between two lists, which the meta (step 1) can check directly. The board-simulator upgrade (noted 2 Oct, Jordan via the design session). The step-6 version of delivery would be a Monte Carlo of units pathing over real terrain layouts. We didn't start there, for two reasons:
- terrain decides most of it, and we have no terrain data;
- distance constants checked against the lists (V3's standard table, design/scoring.md) get most of the value for a fraction of the work.
The closed-form model's exposure factor (the share of shots that have a line, 1 today) is the hook a board simulator would replace.
7. The counter: what beats what's winning
What. Given the meta's top lists, which lists, units or loadouts do best against them. Maths. Step 6's matchup grid over the meta's lists, then step 4's search with the meta as the opponent field. In game-theory terms, a best response to the field's mix, and a look at whether the field itself is stable. Size. Weeks once steps 4 and 6 exist; nothing before. Unlocks. Sleepers with a reason (T-534): "nobody takes it, and it beats three of the top five lists".
8. Play and learn: the coach
What. A player brings their list and the opponent's; the site says how the game is likely to go, where it turns and what to try. After the game they tell us what happened, and the model learns. Maths. Everything above, plus learning from reported games (a Bayesian update of the step-5 and step-6 numbers by the results players send). Size. Ongoing. Unlocks. Jordan's aim in T-523: "they'll look to us for leads."
What stays true at every step
- The referee is the winning lists and public tier lists (steps 1 and 3, T-533). A step that makes us agree with them less needs a written reason.
- Every number gets a sentence in design/scoring.md before it reaches the site.
- Datasheets stay fetched in the browser, and rules stay in our own words. Missions and dispositions are read, never copied.
- The free calculator stays free. A deeper step can be a supporter perk only when it's our software (a custom lens, a long simulation), never a datasheet.