Study X · Sessions & memory
Edit-anaphora
The last structurally unmeasured request class: follow-ups that point at the previous edit; "also set that same node's…", "apply the same change to X", "actually, undo that." Anaphora steps ran against skeleton views (an oracle-focused view would leak the answer), so the carrier under test had to supply the target, the key, AND the value.
Without one, no model resolves anything: 0 of 144 across sonnet, gemini, and opus, and every single failure was a valid, applied, silently guessed patch; the strongest silent-failure result in the series.
Full history is near-ceiling (48/48, 48/48, 46/48). The winner is the cheapest carrier tested: a one-line last-edit echo the application appends automatically (set "key" from old to new on node), which ties history per model at about half the input cost and BEAT history outright on opus-4.8; 48/48 against history's 46/48, with the from/to format carrying every undo on every model.
Fine print, disclosed: "apply the same change to X" strains compressed carriers model-dependently (sonnet 7/12 with the echo; window-2 leaks on repeat everywhere, significantly on gemini), one more entry in the carrier-advice-is-not-model-independent ledger.
The measured stateless session stack is now complete: fresh view per turn, two worked examples, the session-notes memo, and the last-edit echo.