This is one of the few LLM-for-markets writeups where the negative result is the valuable one. The v2→v4 sign flip (+0.24 → −0.06 correlation, +3.9% → −2.4% decile spread) deserves to be the headline: most teams would have shipped v4, because every intuition says "closer to how a thoughtful human reads = better." You've rediscovered with an LLM what the event-study literature has said since Tetlock (2007): tone is contemporaneous at best and mean-reverting at worst, and the only text that predicts is text that maps to a dated, quantified surprise. Ke, Kelly & Xiu's SESTM work is the closest academic cousin — they likewise found the predictive vocabulary is narrow and boring.

Two suggestions, offered because this line of work is worth pushing further:

1. On the tight variant (+0.34 on 17 signal days): before trusting any variant that improves by pruning, run an equal-N null — draw random same-size subsets of signal days from the parent set a few thousand times and see where +0.34 falls in that distribution. Sparse-signal "improvements" are reproducible by random pruning alone more often than seems possible, and a CI-includes-zero caveat usually resolves to exactly that.

2. The one that bites if you extend the window: training-cutoff look-ahead. Your six-week live test largely dodges it, but the moment this design is run over 2021–2024 with a current-vintage model, the model is grading headlines about companies whose subsequent history it memorized — Nvidia news from 2023 is not exchangeable with Nvidia news from next week, to a model trained in 2025. Backtests built this way flatter themselves. The fixes are ugly but necessary: evaluate post-cutoff only, or mask entity identity before classification and check the signal survives.

One question: did the strict prompt's negative calls predict downside symmetrically? The neutral-by-default architecture makes negatives the highest-bar calls, so whether they carry signal would say a lot about mechanism versus one-sided news flow.