The endgame clear that catches its own imps

AI notice: This text is created with the support of AI systems; it is reviewed editorially and taken responsibility for before publication.

At the beginning of October none of our warriors stood in the top third of koth.org 94nop. The best, the stone/imp Megrez, was at rank 8 with 135.43 — exactly 0.02 points below the top-third mark. This page documents the attempt to close the gap with a rebuild of the endgame, and the finding of why that does not work in this class.

What the matrix says

The hill publishes every pairing. From that one can read off, for every resident, how hard it is for the rest of the field (column mean):

On the left the papers and stone/imps, on the right the six scanners — three of them ours. All top-7 warriors live off the scanners: Foolishly Hyperpure Lark takes 190 to 223 points from them, Intermodal 173 to 203, incubus 170 to 200.

Megrez already has this scanner prey (Pendulum 208, Phecda 185, Foolishly Hyperpure Lark 172, Ecaz 171). It loses only to the three imp papers: Riptide 79, Corrosion 85, Renegade 93 — with 3 wins, 13 losses and 84 percent ties. And it is not alone in that: every stone/imp on the hill shares this gap (Intermodal 103 / 103 / 105, 3Daisies 112 / 88 / 99, Road Warden 104 / 84 / 81). Conversely, the papers clearly beat only two residents, Pendulum (153 / 189 / 214) and Phecda (152 / 168 / 189) — both scanners with an SPL/DAT clear, and both lose to the seven stone/imps (87 to 126).

The arithmetic was tempting: Megrez at Renegade level in the three paper cells, plus 35 per cell, would give +5.5 points — rank 1 or 2.

The measurement line

For the rebuild, a measurement line of ten real rows was set up: Arvon, Hatysa, Megrez, Tolypeutes, Intermodal, Ecaz, Phecda, Pendulum, a replica of Renegade and Pyronaut from the Koenigstuhl archive as stand-in for Riptide. Check against Megrez's real hill cells: seven of ten cells within 5 points (Arvon 126.5 against 127.2, Pendulum 209.8 against 208.0, Renegade replica 98.0 against 92.8, Pyronaut 83.3 against Riptide 79.2), two outliers (Tolypeutes 129.6 against 143.2, Phecda 201.5 against 185.2). Mean over ten opponents: 139.83.

What happens in a tie

Sixty positions against the two papers, with round details: against the Renegade replica 6 wins, 6 losses, 48 ties; against Pyronaut 2 / 20 / 38. In all 86 ties both sides live to the last round, 80,000, both close to the process limit — the opponent with a median of 4137 processes, Megrez with 309. The classic paper-versus-imp stalemate: the paper cannot kill the imp ring, the imp ring cannot kill the paper.

The trace of the endgame against an inert opponent shows the reason. Once the stone passes into the DAT clear, the clear runs at only 0.2c — a fifth of the speed it would have on its own — because the imp processes share the cycles with it. A 0.2c DAT clear does not catch up with a silk paper.

The hypothesis: SPL first, then DAT

The idea: before the DAT clear, a pre-pass of SPL bombs that freezes the paper processes in endless loops, so that the DAT clear reaches them afterwards — the recipe with which Pendulum and Phecda beat the papers. Four variants in the budget, with a threshold table fixed beforehand.

The first round failed at the build: the boot geometry of the chassis does not accept longer tails without further ado, the stone head was missing, and Pendulum dropped to 27. A build error, not a finding. For the second round there was a build gate: the landing instruction has to be the expected one, and an inert opponent has to die between 4000 and 4600 cycles as with the original. Three variants broke the stone despite a correct landing — additional lines between certain stone cells disturb the bomb sequence, the cause is unexplained. Two passed and were measured:

The paper cells do not move: Renegade +0.1, Pyronaut +8.5 — the only cell with a plus. Instead the cells against the stone/imps collapse by 34 to 38 points, the self cell by 37.

The structural finding

SPL bombs catch the bomber's own imps. Every imp process that reaches one of its own SPL cells becomes an endless source of processes. The cycle share of stone and clear collapses, and the warrior chokes on itself. An endgame clear on a self-splitting stone with a running imp ring cannot resolve the paper stalemate.

Pendulum and Phecda manage it — because they carry no imp ring of their own. And that is exactly why they lose to the seven stone/imps. The gap in the top third does not sit in a constant but in the class: stone/imp loses structurally to imp paper, and what helps against imp paper takes away the stone/imp's strength.

Thread closed, negative. Four variants, all below the base, no cell better across the lists. What remains are the measurement line and the build gate — both were reused in the two threads that followed: Paper and scanner as chassis and The paper stalemate.

Measurement convention

• Ten real rows, start positions from the position list of test field v16, 400 positions × 2 × 3 seeds, process limit 8000, one round per position.
• Build gate before every measurement: landing instruction and time of death against an inert opponent (4000 to 4600 cycles like the original).
• Tie diagnosis: 60 positions with round details and trace.
• Hill numbers: published pairing state of koth.org 94nop of 8 October 2026, 04:11 UTC.
• Source: entry N1382 (8 October 2026).

Attribution

Megrez, Arvon, Hatysa, Ecaz and Phecda are our own warriors (Stefan Kammann & Forge-AI); of Megrez only the class is given here, no source code. Pendulum is by John Metcalf, Intermodal A by Steve Gunnell; the Renegade replica is based on Metcalf's Mercenary. Renegade, Riptide, Corrosion and Pyronaut are by John Metcalf, incubus by Christian Schmidt, Foolishly Hyperpure Lark by Ken Espiritu, Tolypeutes by Roy van Rijn; Diorthotic, Spawn, 3Daisies, Road Warden, HazyLazy, Kudu and Forgotten Lore are further foreign hill residents. All appear here only with published hill results; Pyronaut is in the Koenigstuhl archive.

Related pages

• Strategy classes — stone/imp against paper, the measured matrix
• The paper stalemate — eight parameters of the same chassis
• Portfolio and age — where Megrez stands

← Core War overview