Skip to content

Instant timing and the entity relational gap

2026-10-20 → 10-22

Does the net play instants well? Humans on 17Lands FDN cast about half their instants on the opponent's turn; the bots cast about a fifth there. Does it take kills when they are available? And can the value evaluation see stack targets, or who blocks whom?

The starting measurement: "Bots cast ~21% of instants on the opponent's turn; 17lands FDN humans cast ~50% (z=-10)". For removal, bots 0–14% against humans 37–61%; for combat tricks, bots 0% against humans 61–84%.

Verdict

  • The relational block was adopted (it is in the 100-wide encoding), but it unlocked no training gain.
  • Rules at play only move casts to main 1; nothing in the training signal rewards holding. The run notes' conclusion: "The loop is flat on every knob tried; it is not an encoding problem and not simple overtraining."
  • This led to the training-side change (held casts recorded as zero-visit candidates), which moved opponent-turn share from 20.9% to 35.2% at equal strength (react).

What was done

  • 1. The relational entity block

    Columns 68–83 of the entity record: targeted by me / by the opponent, target counts by zone, stack X, kicked, combat foes and their power and toughness, dies, kills (architecture). Landed as gorge 333264b3b, with matching changes in the kernel and the trainer; old nets play identically (zero-padded). The run notes call it entity format 3; the specification that defined it calls it format 2.

  • 2. A fresh control series, f84

    26 generations of about 15 minutes each, sims 200, window 4.

  • 3. Knob rungs on that loop

    Value thinning, 400 sims, playout-cap randomisation, window 12, depth, relational gain, a :settle leaf.

  • 4. Rules at play

    instwin=noupkeep and instwin=noupdraw.

  • 5. Behaviour probes

    Kill censuses, forced kills, a stack-target probe.

Results

Try Result
relational block, stack-target probe the net reads the target (98% consistent sign) but prices it at about 2% of the creature
f84 control, 26 generations flat: 0.775–0.794 vs sb-tactical, 0.61 vs g7, 0.50 vs prev
knob rungs all flat or worse, all stopped
behaviour on g14/g26 instants held through ~40% of opponent turns (p4c 38%); lethal burn taken 89–92% when offered, lethal attack 100%
forced kills (:killfirst, 2,000 games) 0.391 ±0.021; own creatures 0.392 ±0.021; opponent's 0.449 ±0.022: taking more creature kills is not better
instwin=noupkeep, 5 generations from g26 strength flat (sb-tactical 0.78–0.80, g7 0.61–0.62); own-upkeep casts 36 → under 1 per 100 games, moved to own draw step and main 1; opponent-turn share 25.5% → 23.4%
instwin=noupdraw, from g4 g7 0.613 → 0.586; sb-tactical 0.78–0.795; opponent-turn share 26.6% → 21.5%
rule at play, head to head (g7, 2,000 games, sims 100) full-combat+resp 0.5035 ±0.0219; inst8-combat+resp 0.4980 ±0.0219

Source: the run notes' table of what was tried, the forced-kill figures from the run's output and the head-to-head from its results. All kept privately.

Data

The line's generations, census tables and archives, the per-class census runs and the forced-kill counterfactuals are kept privately.