my gt research: an outside review took its best shots, and the paper is still standing (week five)

week five, in two sentences: i ran the paper through an independent review with five separate reviewer personas — benchmark coverage, contextual-integrity theory, judge methodology, adversarial security, and a skeptical venue reviewer — and worked through every finding worth acting on. the two sharpest open questions from last week both got closed with real experiments this time, not better wording.

last week ended with the paper mostly done and two loose threads i didn't love leaving open. this week i stopped writing around them and just ran the numbers.

the near-tie, rerun at scale

last week's honest worry was a result that looked too convenient: a single full-information llm call matched ciguard's two-stage semantic leak rate almost exactly, on 150 messages, one model. if a single call ties the two-stage system, the paper's core architectural claim gets shaky.

i reran it at n=1,972, the same scale as the headline results table, and the near-tie doesn't survive:

the gap holds on both metrics and widens at scale instead of shrinking. that's a properly powered result behind the architecture now, not an assertion from a small sample, and it closes the open question outright.

isolating where the task-success gain actually comes from

separately, the paper had been crediting a task-success improvement in the cross-model live setting to the rewrite "decluttering" the message — plausible, but asserted rather than shown. so i ran a declutter-only ablation: a rewrite told to clarify and de-redundant the message, but never to remove any field.

style cleanup alone doesn't reproduce the gain — it's slightly worse than doing nothing. so it's the removal of over-shared, task-irrelevant content doing the work, not the rewrite's prose style, which also rules out the worry that a grader was just rewarding familiar phrasing.

the rest of the review

the other findings were smaller but all worth fixing: a novelty claim against a comparison system got reframed around what actually differs instead of overstating the gap, the multi-level lattice contribution got correctly situated in the decentralized information-flow-control literature instead of undersized against a narrower baseline, a real gap in what the judge-validation setup does and doesn't cover got named directly instead of left implicit, three new attack surfaces went into future work, and about a dozen new citations went in after being checked individually against the live api.

on the process side, a mid-review figure swap got reverted back to the version i actually preferred, and the diagrams and result plots got resized relative to each other so the paper reads like a systems paper instead of a plot-first one.

where it actually stands

the paper is merged to main, holds at exactly 12 pages, and compiles clean. what's left is scope (bundle everything into one submission or split off the thinner theoretical thread), a title decision, and a final read-through — all decisions, not open research questions.

next

ten days to the july 24 deadline. the two things that were genuinely unresolved a week ago are now measured and closed, so what's left is polish and sign-off, not more experiments.