The weighting ran, then the dedup pass undid it 56 lines later

The arithmetic is what makes this legible. Multiply each weight class by the turns in it and the six products come to 94,022, exactly the line count the unmodified extractor produces before its own dedup pass ^[16]. The curriculum was built. Then, 56 lines further down the same function, it was taken apart ^[12].

Dedup had no way to know better. `_make_example` is a pure function of the role, the thinking text and the command, so the sixth copy of a weight-6 turn is byte-identical to the first, and every copy after the first hits a key already in the set ^[13]. That key is the last 200 characters of the line ^[12]. An intentional duplicate and an accidental one are indistinguishable at that resolution.

The number worth holding onto is the share. Gold defensive turns, the ones that patched an unpatched vulnerability and kept the flag ^[9], were meant to be 18.3 percent of what the trainer saw, 17,166 lines of 94,022; they arrived as at most 5.7 percent, 2,861 of 50,145 ^[17]. That is a dilution of roughly 3.2x ^[2]. And 2,861 times 6 is 17,166 on the nose ^[3], which means not one gold turn kept a second copy. The class the retrain existed to promote finished below the neutral turns it was written to outrank ^[17].

The curated Gemini exemplars are the cleanest demonstration that the knob was never connected. The comment above the constant says 5x, up from 2x, so the exemplars would dominate the bot's own neutral defense habits ^[19]. The file's 3,275 lines carry 3,187 distinct 200-character tails, so it contributes 3,187 lines at 2, at 5, or at 500 ^[20]. Eighty-eight lines of a hand-curated file, 2.7 percent of it, were gone before any weight was applied ^[4].

One instruction did survive: the drop. 27,386 defender turns scored zero were never appended, 6,602 of them for firing commands at another machine's IP, which is an attack ^[21]. Zero copies is a decision dedup cannot reverse, so a six-level grading scheme reached the trainer as an include-or-exclude filter, and the gap between a 6 and a 1 became the gap between a 1 and a 1 ^[22].

Subtract and the pass is worse than a no-op: 94,022 minus 50,145 is 43,877 lines removed ^[1], of which 38,988 were the weights ^[15] and 4,889 were distinct examples that merely shared a tail ^[7], 8.9 percent of the genuinely distinct corpus ^[23]. They collide because the key straddles the end of truncated thinking text and the command ^[24].

That leaves the v7 result without a cause. It was built around weighting good defensive turns, it regressed offense more than it gained on defense, and it went back to the v6 snapshot ^[6]. Whatever produced the regression, it was not the weighting, because no weighted file ever reached the trainer; the author's own account is that neither of the two curricula he judged was ever weighted ^[7]. Meanwhile the bot still loses its flag 0.61 times a game against 0.18 for Gemini 3 Flash ^[4], and the fix has not actually been tried.