Eve library

Comparison and configured-target tables · pilot brief

Guide

Revision 1

On this page

Source pin: 1ceae52c7896a59ac74a4429d410f3218f0b91c6.

Visual thesis: open, aligned comparisons keep the identifying row beside its evidence, with calm cream/ink rules and exact numeric alignment. Use a table because readers need exact percentile targets across distinct measurement contracts; a bar chart would overemphasize relative magnitude between unlike intervals. This is a quantitative pilot candidate, not an accepted benchmark.

Content plan: retain the measurement-interval question; distinguish configured targets from sampled results; compare each metric at p50/p95/p99; explain every missing entry visibly; finish with the empty-sample evidence warning. Detailed timestamp definitions, nearest-rank calculations and assessor differences remain in the explanation and current narration. No values or production measurements are invented. No new slide or extra chart is needed for this communication job.

Interaction thesis: keep native keyboard/touch table scrolling, with the row header staying visible; preserve slide navigation, source disclosure and reduced motion. No animation or hover is required to explain a value. PDF retains the complete table and missing-value legend without a scroll-dependent omission.

Quantitative slide brief#

psyche-latency-targets: teach that comparison requires the correct measurement interval and target provenance. Relationship: exact quantitative comparison, not ranking, trend, parts of a whole or measured performance. Emphasis: metric identity and the configured/not-configured distinction. Reading order: question, configured units and legend, metric row with its percentile targets, empty-data caveat. Four rows are the source's four distinct metric contracts, not a quota.

Evidence: PSYCHE_VOICE_FIRST_AUDIO_LATENCY_TARGETS, PSYCHE_FIRST_TOKEN_LATENCY_TARGETS, PSYCHE_AVATAR_FRAME_LATENCY_TARGETS and PSYCHE_ASR_FINAL_TEXT_LATENCY_TARGETS in the pinned latency-dashboard.ts; the complete latency-budget source unit and the existing controlled sample and assessor tests. Values are exact integer milliseconds, without rounding, aggregation, conversion, sampled population or observation window. A missing percentile target is not zero and does not assert an observed latency.

Introduce typed numeric cells and explicit missing-state labels rather than relying on a bare dash string. Bind provenance to cited source, pin and exact symbol order; test every displayed value and missing cell against those source exports. The generic exact-number representation should preserve negative and zero values where another source legitimately supplies them; this specific latency-target table must continue to match its positive source constants.

Shared table behavior#

All native slide tables already have column and row headers and 22px right padding. Preserve those semantics and native horizontal scrolling. On narrow screens keep the first column anchored with an opaque paper background and a restrained separating rule. Ensure no row becomes unidentified at the far edge, all other columns can still be reached, and the same row/column association is available to assistive technology. Do not apply the slide-table behavior to tables inside source/explanation dialogs.

Recheck actual full-size images and text geometry before changing optical spacing. The prior priority-decision ledger's room-maturity touching-column finding is retracted: a new 1440px capture is byte-identical to its saved image, and DOM text ranges show 195.984375px between the final row's substrate text and next cell. The existing 22px cell padding is intact. No spacing fix is warranted by that finding. Mobile row identity loss is separately reproduced.

Review boundary#

Exercise real four-column numeric and three-column qualitative comparisons at desktop, compact, active narration, narrow standalone/embedded and PDF sizes. Test missing/zero/negative values and long labels as labeled synthetic fixtures, not audience evidence. Inspect numeric alignment and the full scroll path, not only page overflow or an attractive screenshot. Whole-library table behavior needs the complete applicable browser/PDF gate because the CSS is shared.

Neutral reader tasks: what does the ASR p99 entry mean; can these values prove a production latency result; which metric still owns a value after horizontal scrolling? Record real willing-reader observations later; this document contains no participant results, listening acceptance or pilot promotion.