Stratagems in Value, held out on every army (T-742's step)

6 Oct 2026, Pacific (T-742; Jordan approved the step on 6 Oct: the test that decides whether stratagems belong in a unit's value). Frozen snapshot b3c416933e, inputs hash 1b88e955b9b3 (node tools/ledger-inputs.js --verify: 27 inputs match). Base: step 41, Track A2, the benchmark (tools/ledger-benchmarks/2026-10-06-step41-stable.json). Tools: tools/stratagems.js (--step), tools/stratuplift.js, tools/recipestep.js (the protocol), tools/postchain.js. Scratch numbers in build/strat/ (git-ignored). The prediction and the keep rule were written in reports/tiers/ledger-steps.md and committed (fe34cf96ca) before anything was scored. Nothing on the site, in docs/data or in the benchmark moves.

Detachments are named by the ledger's slugs; stratagems only by our ids.

In plain English

The verdict: reverted. Stratagems priced into Value don't agree better with the reviewers, and they do worse than the same credit handed to random units.

• Held out on the 15 exam armies: −0.6 points [−1.8, +0.4] over step 41. 4 armies up, 10 down: only 5 of 15 don't fall (the rule needs 9). Leave one army out: −0.9 [−2.3, +0.3]. Gemini at half weight: −0.1 [−1.2, +1.0].
• Placebo credits do better. The same credit per army, handed to random units of that army, 20 draws through the same fit: their held-out gains run from −1.3 to +0.4, mean −0.2. The real credit (−0.6) beats only 3 of the 20. The rule asked for it to beat 19 (their 95th percentile is +0.4).
• Shuffled letters: −0.1, +0.6 and −1.1 held out. The real result sits inside the noise, on the wrong side of 0.
• It fails every part of the keep rule: the gain is negative, the interval's low end is −1.8 (the rule needs above −0.5), 5 of 15 armies not falling, and below the placebo's 95th percentile.

Why it fails: the whole pull comes from one army, the Leagues of Votann.

• Votann gain a lot: +12.4 at face value and +21.9 at w 2. That is the pilot's result again, on one list of 22 graded units.
• Every other army loses. Leave Votann out and every w loses on the other 14 armies: −0.2 at w 0.25, −0.6 at w 1, −1.0 at w 2, −1.8 at w 8. At every w, at most 7 of those 14 don't fall.
• So the folds pick w 2 whenever Votann is among the 10 armies they train on (6 of 9 folds), and step 41 itself when it isn't. On the held-out side Votann never gains: when Votann is held out the fold picks the base. That is the signature of one army's noise carrying a fit, and the held-out exam is built to catch it.

The two fitted parameters.

• w (strength): 2 in 6 of the 9 training folds, 0 (step 41) in the other 3. Leave one army out picks w 2 for 14 armies and 0 when Votann is the one left out. It is stable only because Votann dominates the training score.
• The shoot-back share: 1 (the whole extra phase) in every fold that took any credit. No fold chose 0.25 or 0.5. At small w the share hardly matters (each army's change is the same to ±0.1 across the three shares). Only Votann separates them (+15.2 at share 0.25 and 0.5, +21.9 at 1, w 2), and that is what picks share 1.
• So the share isn't settled by the reviewers' letters. It is settled by one army's one list, like w.

Face value (w 1, share 1, no fitting): +0.3 [−1.0, +1.7] over the 15, 4 up and 10 down. Gemini at half +0.5. The placebo credits' face value runs −1.9 to +1.0 (mean −0.4), so the real credit is above their average (it beats 17 of 20) but not clear of them. Without Votann it is −0.6. Tyranids 82.6 → 82.5 on the three lists, consensus 90.4 unchanged.

What moves, and why the reviewers disagree (at w 2, share 1, the folds' pick):

• Captains, by their CP. A Captain's 3 CP a game buy up to 104 points a game in anvil-siege-force. The credit is generator credit only, and it lifts every Captain:
  • Space Marines: Captain 38th → 7th, in Phobos Armour 75th → 14th, in Terminator Armour 26th → 6th.
  • Dark Angels: Captain 47th → 10th, in Terminator Armour 50th → 11th, in Phobos Armour 82nd → 34th, graded D and F.
  • Blood Angels: the Captains 15th, 17th → 9th, 11th.
  • Lieutenants carry no CP rule and don't move.
  • The reviewers grade the Space Marine Captains A/B, but the Dark Angels' ones D and F.
• The shoot-back users rise most, mostly against their letters:
  • Firestrike Servo-Turrets 7th → 1st (B/B/D), Centurion Devastators 13th → 3rd (C/B/C), the Gladiator Valiant 6th → 2nd of 93 (C), the Baal Predator 14th → 4th (B), Kapricus Defenders 13th → 4th (C/B/C).
  • With the Tankbustas 9th → 1st (A) and the Rubric Marines 8th → 5th (S), they go toward their letters.
• Votann, the disputed units:
  • Thunderkyn 5th → 1st (S/A/S): right.
  • Sagitaur 15th → 7th (C/A/D/A): mixed.
  • Hearthguard 16th → 11th (A/A/A/B): right, but still low.
  • Hekaton 11th → 9th (S/A/S/S): right, but small.
  • Kapricus Defenders 13th → 4th (C/B/C): wrong.
  • Yaegirs 18th → 10th (S/A): right.
  • What falls in their place: the Arkanyst Evaluator, the Einhyr Champion and the melee Steeljacks.
• Tyranids barely move.
  • Deathleaper 3rd → 1st, Norn Assimilator 14th → 11th, Old One Eye 19th → 16th.
  • Hive Tyrant 2nd → 3rd, the Swarmlord 4th (unmoved), the Winged Hive Tyrant 5th (unmoved).
• Death Guard:
  • Plague Marines unmoved (8th): their credit, 118 points saved a game in champions-of-contagion, is priced on the led squad and doesn't reach their ranking form.
  • The Biologus Putrifier unmoved (31st → 32nd, graded S).
  • The Defiler 3rd → 2nd.

What it means.

• The catalogue and the CP economy are a good description of what stratagems do: the points per CP below.
• As a line in a unit's Value, they don't predict how reviewers grade units. They pay most to units the reviewers don't rate highly: the shoot-back tanks, the Captains as CP batteries, the Kapricus.
• One reading: reviewers grade the unit, and the stratagems they expect to spend sit with whichever unit is best at the time, not with a fixed carrier.
• Stratagems stay out of Value. The catalogue stays as a report, and could feed the calculator's stratagem chips (T-742's later cards).

1. The test, as written before scoring

From reports/tiers/ledger-steps.md, row "stratagems" (committed in fe34cf96ca before any score):

• The step:
  • Base: step 41 (Track A2) with its post step.
  • The candidates add the stratagem credit as a post step chained after it (tools/postchain.js).
  • The credit: Value′ = Value + w × (Kill rate × kill credit + Soak contribution × Soak credit + the army's median Kill rate × generator credit), each unit's best detachment (the pilot's formula).
  • The spend has the core re-roll in its menu (the catalogues' "+core" reading, the pilot's 44b).
• The new shared parameter, share:
  • It is the part of a whole extra phase that a shoot-back stratagem counts: a version with again, the unit shooting in the enemy's phase or once more.
  • It is applied inside the spend, so the CP economy re-buys at each value.
  • 24 catalogue entries carry again: the four Codex ones (anvil-siege-force, ironstorm-spearhead, firestorm-assault-force, headhunter-task-force), repeated in each of the four Marine armies, plus unforgiven-task-force's, eradication-cohort's, ordo-hereticus-purgation-force's, three of the Votann's, wreckas' and rubricae-phalanx's.
  • No fight-again stratagem is priced; all of those are listed by kind ("fights on death"), so the share reaches shooting only.
• The grid: w on STEP_W {0.25, 0.5, 1, 2, 4, 8} × share on AGAIN_SHARES {0.25, 0.5, 1}: 18 candidates and the base, which comes first and wins ties.
• Measure: tools/recipestep.js's protocol.
  • The 15 exam armies, 3 seeds × 3 folds: chosen on 10, scored on 5.
  • Leave one army out.
  • Gemini at half weight.
  • The Tyranids' three lists and their consensus beside.
• Controls:
  • (a) Shuffled letters, seeds 11 to 13, through the same fit.
  • (b) Placebo credits: each army's adds (the Value each unit gains, so the same total credit) handed to its ranked units at random, PLACEBO_SEEDS 101 to 120, each through the same held-out fit (args.placebo).
  • This placebo permutes the adds unit by unit. The pilot's permuted form ids, which only roughly kept the total.
• Keep rule: keep only if all four hold:
  • the human held-out gain over step 41 is positive;
  • its 95% unit-bootstrap interval's lower end is above −0.5;
  • at least 9 of 15 armies don't fall;
  • the gain is above the placebo credits' 95th percentile (the 19th of the 20 held-out gains).

Otherwise revert.

• Prediction:
  • Held out −0.5 to +1.0, the interval's low end below −0.5.
  • 8 to 11 of 15 not falling.
  • The share at 0.25 or 0.5 more often than 1.
  • The real gain inside the placebo's spread, below its 95th percentile: revert.
  • Face value −1 to +1, Votann up 3 to 12, the Marines within ±2, Tyranids within ±0.5.
  • Shuffles about 0.
• Right:
  • the verdict and the placebo comparison;
  • the gain's range (−0.6);
  • face value (+0.3), Votann (+12.4);
  • the Tyranids (−0.1);
  • the shuffles;
  • Thunderkyn, Sagitaur and Kapricus up; Hive Tyrant and Swarmlord unmoved.
• Wrong:
  • The count: 5 of 15 not falling, below 8 to 11.
  • The share: always 1, never 0.25 or 0.5.
  • Blood Angels: −4.0 at face value, outside ±2.
  • Plague Marines: unmoved, not up.
  • Captains: they rose far more than "up for their CP" suggested.

2. The held-out exam

Each candidate fixed, over step 41, the 15 (human, change [units], up/down; Tyranids' three lists and consensus):

The fitted parameters, fold by fold

• Leave one army out picks w 2, share 1 for 14 armies, and step 41 when Votann is left out.
• Stable in value, but carried by one army. The three folds that pick step 41 (3, 4 and 9) are exactly the three that hold Votann out of training.

3. The placebo credits (20 draws, the same fit)

• Held out, the real credit is worse than the typical random placement.
• At face value it beats most random placements, by about 0.7 points on average, but stays inside their range.
• Four placebo draws pick step 41 in every fold (+0.0); a few beat the real credit by 0.5 to 1.0.

4. By army

• Gains: World Eaters (Forgefiend 18th → 8th, A/S/S; the Master of Executions 4th → 2nd), Thousand Sons (Ahriman 10th → 2nd, A; Rubric Marines 8th → 5th, S), Agents (Imperial Navy Breachers 19th → 13th, A; Sanctifiers 7th → 5th, A) and Genestealer Cults (Acolytes with autopistols 11th → 4th, A/A).
• Biggest losses:
  • Blood Angels: the Baal Predator 14th → 4th (B) and three Captains (B, B, C) climb over the Death Company (A, B).
  • Drukhari: the Corsair Voidscarred (ungraded) and Kabalite Warriors (C) and Wracks (D) rise past the Incubi (S) and Drazhar (S).
  • Space Marines: the Servo-Turrets, Centurions and Captains climb past the Eradicators (S/A/B), Hellblasters and Aggressors.

5. The disputed units (w 2, share 1; rank of the army's ranked units, Value, the reviewers' letters)

• Outside Votann, seven shoot-back users take the credit. Two move toward their letters (Tankbustas, Rubric Marines); the other five are graded B to D and rise to the top of their armies.
• At share 0.25 they rise less, but the folds never chose it: at small w the share hardly changes any army's score.

6. Points per CP, by detachment (snapshot 1b88e955b9b3, +core, share 1)

10 CP a game (CORE_CP 2 a battle round × ROUNDS 5), the typical list by Value within 2,000 points. Average = points ÷ CP spent; margin = what the next CP would add (0: every priced use is bought, more CP buy nothing in our catalogue). Best first by average.

The rows hardly differ from the catalogue reports' (old snapshot): the bug-fix snapshot moved the typical lists of a few armies only (the Poxwalkers, the Cult Icon squads). A detachment's average of about 9 to 12 with margin 0 is the core re-roll alone: 10 CP of re-rolls and nothing better to buy.

7. What changed in the tools

• tools/stratagems.js:
  • ARMIES names all 16.
  • AGAIN_SHARES and the spend's share: a stratagem with an again version counts that share of its points, before the greedy buys.
  • With args.share the post step runs the spend itself from build/ledger/strat-uplift-<army>.json; without it, it reads the committed credit file, as the pilot did.
  • args.placebo: the army's adds permuted among its ranked units.
  • --step: the grid STEP_W × AGAIN_SHARES, recipestep.js's evaluate, and the 20 placebo grids through the same held-out fit.
• tools/postchain.js: a step's result is kept per input run and step, so 379 candidates that share the stable layer run it once. Every step is a pure function, so the numbers don't change. The whole step, 20 placebo grids included, ran in 15 seconds.
• tools/stratfx/credit-[core-]<army>.json: rewritten for all 16 armies on 1b88e955b9b3. The 14 new armies' uplifts were computed here for the first time on this snapshot (about 5 minutes for the Space Marines, 7 for the Dark Angels).
• Tests:
  • stratagems.test.js: all 16 wired; the share scales only again stratagems; the placebo keeps the army's total.
  • scoringdoc.test.js now checks stratagems.js's constants are named in design/scoring.md.

8. Next, if anyone returns to it

• The step is reverted, so stratagems stay out of Value. Two variants might be worth a declared step one day:
  • Without the generator credit. The Captains' rise is the clearest wrong-way move: Dark Angels' Captains are graded D and F.
  • Credit to the receiving role, not the unit. The reviewers may price "a good stratagem target" into the army, not the unit.

Neither is declared. The Votann-only gain warns that any variant needs the placebo test.

• The catalogue's natural home is the calculator (stratagem chips under the detachment), T-742's later card.