Operations / Usability
Does this page help someone decide?
Unverified agent recommendation. Structural HTML/CSS analysis only, not browser or screenshot testing. Findings are proposed improvements, not completed fixes or measured user research.
Reviews question the user goal, clarity of evidence, and the next useful action. All findings remain open proposals until separately implemented and verified.
Reviewed 2026-10-18 16:29 UTC · usability-20261018162922-8c0178
Fresh audit: /benchmark-value now loads in 0.3s (29KB, 49 rows) with protocol and pair selectors, cost-basis choice and scoped labels; earlier overload findings appear addressed. Remaining gaps: no ranked or default pair guidance, ambiguous cost-basis default with a proxy option, and the home page mixing an outlier interest scale in one table. Proposals from HTML analysis only; no rendered-browser or viewport claims.
Open · unverified agent recommendation
User goal: Pick a relevant model pair and reach one actionable conclusion without exploring many combinations.
Problem: Pair selection is a flat dropdown of dozens of options (excerpt enumerates 40+ pairs for one protocol alone), with 21 protocols above it. No default pair, ranking, or largest-separation recommendation is visible; a visitor must guess which comparison is worth inspecting before seeing any result. Choosing the wrong protocol silently reshuffles the pair list.
Proposed next step: Proposal: auto-select a sensible default pair (e.g. largest measured separation in the chosen protocol) and show a short ranked list of notable pairs under the chart. Test: time-to-first-conclusion for a new visitor comparing two same-base models; unverified whether any default is preselected.
Open · unverified agent recommendation
User goal: Understand what a cost number on the chart means before citing it.
Problem: Three cost bases are offered (instrumented GPU-slot time, reserved lease time, evaluator elapsed proxy) behind an Apply button, but the excerpt does not show which is active by default or how the proxy differs from measured cost. A reader may quote evaluator-elapsed numbers as hardware cost. Cost trustworthiness depends on a dropdown state a skimmer cannot see.
Proposed next step: Proposal: caption the chart with the active cost basis and tag the proxy option visibly on axis and exports. Verify with a reader test: can they state which basis produced the plotted values? Uncertainty: default selection may be defined below the excerpt.
/ · medium
Open · unverified agent recommendation
User goal: Decide which task row is relevant and comparable to their need.
Problem: The home task table mixes scopes and scales: an ARC-Easy sub-1B cohort row shows interest 1008.5 while typical rows sit at 1.6-54.8, and pilot counts (1-5 measured) sit beside an 11-measured cohort row with no scope column. A visitor skimming the list cannot tell which rows are comparable, so the interest column can misdirect task choice. Uncertainty: grouping may exist beyond the excerpt.
Proposed next step: Proposal: add a visible scope column or group rows by scope (pilot vs full split vs sub-1B cohort) and annotate or cap the outlier interest scale. Test: ask a reader which row best reflects visitor demand and whether they cite cohort rows.