Held-back dials on the cascade survivors: partial, run stopped by design

5 Oct. T-727. tools/heldback.js on tools/cascade2.js's survivors. Snapshot 8dcd24032d, inputs hash dcd477d8a78c (verified). A report: nothing is adopted and nothing in the ledger moves. The run was stopped by design (Jordan's OK, through the supervisor) after about 1 h 35 min of its roughly 5 h. The next step is a screen of every dial in the full inventory (T-726), then pairs, then one sampled search over all live dials. That replaces adding dials to survivors that were chosen without them. Read every number below as covering only the cells listed under "What the finished jobs cover".

In plain English

• What was asked. The cascade (cascade2) searched 26 dials. This run takes the settings that survived it (the main chain's 213 and the clean chain's 240) and adds, one at a time, each of the 87 dials it left out, at two values each (153 values in all). Each survivor keeps each army's own values as the cascade picked them, and is scored as cascade2 scores it, with the same three scrambled letter sets. Those 87 include the four leader-share switches Jordan asked about: leadersLeading, joinedMarginal, joinedShapley (with joinedMarginal) and leaderCredit at 0.5 and 1.
• What the finished jobs cover.
  • Dark Angels: 125 of the 127 dial values that move it, on all 453 survivors, against Track A's survivors without the dial. The two not finished are jailW 0.5 and wKill ×0.5. The other 26 values don't move Dark Angels at all (the probe).
  • The two Gargoyle dials (tools/steal.js): jailW 1 in all 16 armies, jailW 0.5 in 15 (all but Dark Angels), and stealW 0.5 and 1 in all 16. stealW moves only five armies; it changes nothing in the rest.
  • Nothing else. No other dial was scored in any other army, the Tyranids included. So for every dial apart from jail and steal, nothing here says how it does on the chain's other armies, nor whether a survivor that clears Dark Angels with it still clears its earlier bars. The "together" step was never run.
• Dark Angels, the clean chain's wall, falls to several held-back dials, and each beats its scrambled control there. On their own, the survivors' best is 59.3 (main) and 58.1 (clean), against a bar of 60.
  • arrivalKill at 1.5: +2.7 / +3.7 points on average, against +0.3 / +0.2 scrambled. 108 main and 40 clean survivors reach 60, best 62.2.
  • r.buff 2.12 (today 1.12): +3.1 / +3.1 against +1.4 / +1.2 scrambled. 119 and 16 survivors reach 60.
  • charPsychic 1.5: +2.3 / +2.8 against −0.8 / −0.6 scrambled. 88 and 12 survivors reach 60.
  • Also above their scrambled control: shockPressure off, charMove 1.5, charLone 1.5, wKill ×2, armyMarks 1, killByJob, landsNoClamp and r.detachment 2.07.
  • In all, 30 values pass the gate on both survivor sets in Dark Angels (above 0 and above the scrambled mean). Most of them pass by tenths of a point; the three above pass by 2 to 4.
  • Caveat: Dark Angels has one 11e list. Whether these survivors still clear the other armies with the dial added was not measured.
• What costs Dark Angels most: the VP ledger and missionVP 1 (−13 to −14), charLead 1.5 (−11), charEnh 1.5 (−9), leaderCredit 1 (−9 / −5) and joinedMarginal (−7 / −3).
  • Every leader-share switch costs Dark Angels on both survivor sets: leadersLeading −4.1 / −2.2, joinedMarginal −7.1 / −2.8, leaderCredit 0.5 −4.6 / −0.4, leaderCredit 1 −9.2 / −5.3. joinedShapley costs −3.0 on the main set and gains +0.3 on the clean one.
  • The scrambled letters mostly gain under them, so these are real costs, not noise.
  • Death Guard (the Biologus case the supervisor raised) was never reached.
  • supportMeasured and preferredTargets (Track B) cost −1.4 to −4.1 there, again with the scrambled letters level or up.
• Jail and steal are noise-sized everywhere. Over the fair-test armies (the clean chain's seven) and the other nine they move less than ±0.25 points on average.
  • jailW costs the fair-test armies slightly (−0.13 to −0.23) about as much as it costs scrambled letters (−0.08 to −0.15), so it fails the gate there. On the other nine it is +0.03 to +0.13 against −0.02 to −0.07 scrambled: it passes, at a tenth of a point.
  • stealW 0.5 is +0.01 to +0.05 against +0.03 scrambled, and stealW 1 costs Adeptus Mechanicus −1.6.
  • Neither moves Agents (the main chain's wall): its best stays 58.3 / 59.2. Nothing finished here clears Agents.
• One flag, not fixed here. Some survivors' best forms may rest on wargear combinations BSData forbids (T-723 lists them, for example a Death Guard Helbrute with both hammer and scourge, and Thunderkyn mixing bolt cannon and SP conversion beamers). The frozen snapshot predates T-723's fix, so these scores include such forms.

How it was run

• Survivors: recomputed from build/cascade2/*.bin by cascade2.js's own analyse() (a new all option); 213 and 240, as the cascade report says.
• Each compute: cascade2.js's slidersOf() with the held-back dial laid over (a new extra option). Then tools/exam.js runOf(). Then cascade2.js's valuesAndDeltas(), with only the post values the survivors use (need) and the held-back post step added to the grid's own (extra), as the cascade adds its post steps.
• Two shortcuts, both exact:
  • A probe skips a dial value on any army where it changes no Value at Track A or at the flipped background.
  • Survivors that differ only in a game-wide slider that moves nothing in an army share one compute there. armyEnabler and r.synapse, for example, are inert in Dark Angels.
• Check: 815 finished base scores match the cascade's stored records exactly, and cascade2's own Track A check reproduces on all 16 armies with the changed code.
• Run time with 8 workers:
  • probe 4.6 min, then a first scoring pass stopped after 9 min (it lacked the shortcut);
  • grid probe 0.8 min, then 81 min of scoring;
  • about 1 h 35 min in all. 1,924 of 8,355 jobs finished, almost all of them Dark Angels' (its computes take about 4 s each under load).
• Left out of the catalogue, with the reasons (node tools/heldback.js --list):
  • prem: the same as wKill.
  • raw: a units switch for old benchmarks.
  • combo: display only.
  • killMargin: broken (never built).
  • fireB: tested through killOverGame.
  • profileDials: hibernated.
  • profileLines: the carrier for the feature slopes.
  • popshare: popularity, never blended into a letter.
  • The carry sub-dials (inside carryCargo's shape A), the switches' strengths (inside their switches), and options of dials already in the grid.

Tables

Coverage (dial values per army: finished on every survivor; inert there by the probe (change 0); not run):

Dark Angels, every held-back value (mean change in pairwise points over the survivors; scrambled: the three seeds' mean; best: the best survivor's score; bar 60; the survivors' own best 59.3 main / 58.1 clean):

jailW=0.5, per army (finished or inert armies only):

jailW=1, per army (finished or inert armies only):

stealW=0.5, per army (finished or inert armies only):

stealW=1, per army (finished or inert armies only):