@daviddao

Eval Wiki · educator audit · round 3

Superseded by round 4. The landing page has since been re-organised around the validity frame (Messick × civic-AI pipeline) as the primary orienting lens. See round 4 report.

Test Report v3

Post-fix audit after (a) renaming Plurality WikiEval Wiki, (b) adding a one-sentence summary beneath every page title, and (c) populating a structured Eval Card on every one of the 71 synthesis pages. Format inspired by evalcards.evalevalai.com.

Auditor: Claude · 2026-04-17 · Reference doc saved to wiki/raw/articles/evalcards-evalevalai-about.md and wiki/raw/notes/evalcards-reference.md. Video: full-audit-v3.webm (1.6 MB). Previous rounds: v2 · v1.

What changed in round 3

  1. Rename — every Plurality Wiki string replaced with Eval Wiki across masthead, title, <title> tag, sources page, and the homepage wordmark.
  2. One-sentence summaries — every metric page now opens with a plain-English definition right under the title (before tags or any other layout). 71 / 71 pages covered.
  3. Eval Card block — between the hero and the prose body, every page shows a structured glanceable card with up to 11 fields: kind, unit of evaluation, inputs, output, range, direction, formula, timing, introduced by (year), license/access, main caveat. Only populated fields render.
  4. Schema — content collection now accepts summary: string and card: { … } frontmatter (optional, backward-compatible).
  5. Three title polishesincidence-rate-from-reports, relative-risk-from-reports, reporting-overrepresentation-test had their formulas baked into the title (# IR_G ≥ (β − γ^FR)/(γ^TR − γ^FR)). Formulas now live in the card's formula field; titles read clean.

Educator-perspective evaluation

A student new to the field should be able to answer — for any metric, within ten seconds of opening its page — these four questions:

  1. What is this metric, in one sentence?
  2. What do I need to compute it?
  3. What does it produce, and in what units/range/direction?
  4. Who introduced it, and what's the main pitfall?

Before round 3, answers required reading paragraph-style prose. Now every page surfaces them in the card grid directly below the title.

Gini page with card
Gini coefficient — the definitional sentence, then kind · unit of evaluation · inputs · output · range · direction · formula · introduced by (1912) · main caveat ("insensitive to tails") — all visible above the fold. Math in the formula row is real LaTeX rendered by KaTeX.
DQI page with card
Discourse Quality Index — a coder-based instrument. The summary sentence names the five dimensions; the card then tells the reader what kind of instrument it is, what inputs it needs (verbatim transcripts), what output it produces, and what the ordinal range looks like. "Introduced by Steenbergen, Bächtiger, Spörndli & Steiner 2003" is one glance away.
NSW page with card
Nash Social Welfare — a non-scalar metric whose "kind" field correctly reads "allocation rule + axiomatic guarantee". The main-caveat field ("Computationally NP-hard; the 2019 paper gives a practical approximation that scales") warns the reader before they over-interpret the result.
AIR page with card
Aggregated Individual Reporting — §11 post-deployment. The card makes clear this is not a scalar metric but a framework. Unit of evaluation ("deployed system over a reporting window"), timing ("post-deployment"), and caveat ("Aggregation requires a faithful ex-ante-policy model; mis-specifying it leaks into every downstream number") warn against naive aggregation.

Card field semantics

FieldPurposeExample value
summaryOne-sentence plain-English definition shown in the hero"The classical inequality index: half the mean absolute difference between every pair of agents' outcomes, normalised by the mean — ranges from 0 to 1."
kindRough type (coder-based index, axiom, framework, impossibility theorem, algorithm family…)"coder-based index"
unit_of_evaluationWhat the metric scores"speech act (sentence / short utterance)"
inputsData needed to compute it"verbatim transcript of deliberation"
outputShape + type of the result"five-dimensional ordinal score per speech act"
rangeNumeric range if scalar"[0, 1]"
directionHigher-is-better / lower-is-better / depends"higher is better"
formulaOperational definition"G = (1 / (2 n² μ)) Σᵢ Σⱼ |y_i − y_j|"
timingPre- / during / post-deployment"post-deployment"
introduced_by (+ year)Canonical attribution"Corrado Gini, 1912"
licenseAccess / licensing model of the metric definition"public"
caveatOne-line main failure mode or pitfall"Insensitive to tail shape — two very different distributions can share a Gini"

All fields are optional. A tension page, for instance, only fills kind, caveat, and introduced_by — rendering a short card with just those rows.


Coverage


Screen recording

32-step tour of the post-round-3 site — focuses on card pages across all 11 families + the two hubs + a tension page + search.

Download WebM (1.6 MB)


Does the structure make sense from an educator's perspective?

pass · information architecture

Ten-second rule: can a student name this metric after a glance?

Across 32 sampled pages, the answer is yes. The layout now reads top-to-bottom as a definitional pyramid: (1) which family? (eyebrow), (2) what is it called? (H1), (3) what is it in one sentence? (summary), (4) tags, (5) card with 9–11 structured fields, (6) prose with history and context, (7) backlinks in the sidebar. Each layer adds context the previous one leaves out.

Bridging-score page
Bridging Score — statement-level scalar in [0, 1], 2014, caveat about minority-suppression.
JR page
Justified Representation — social-choice axiom, 2024 LLM extension, formula in-card.

pass · provenance

Every card names the originator

Like Eval Cards itself, provenance is a first-class field. A student can see in one line that quadratic funding is Buterin, Hitzig, Weyl 2019; that DRO is Hashimoto, Srivastava, Namkoong, Liang 2018; that Ostrom's design principles are 1990 with the 2021 blockchain extension by Rozas et al. Attribution never requires reading the prose.

pass · caveats-first pedagogy

Main failure mode in every card

The biggest pedagogical win: almost every card has a caveat row that names the single most common pitfall. Examples from the corpus:

This is the research-educator move Eval Cards gets right — reporting conditions travel with the number, so readers don't over-interpret.

pass · tensions are first-class

Structural disagreements between families are their own pages, with cards

Four pages (tension-bridging-vs-jr, tension-endorsement-vs-dqi, tension-fidelity-vs-strategyproof, tension-fast-vs-severe-harm-detection) explicitly name when not to trust any single metric. Their cards use kind: structural tension and a caveat field with the recommended resolution. This is the most educator-friendly contribution: it teaches the limits of the ontology, not just the ontology.

Bridging vs JR tension
Tension page — summary names the conflict; card's caveat ("Don't pick one — measure both. The opinion-bridge-rate composite is one proposed resolution") guides the reader toward the right next page.

pass · §11 integrated cleanly

Post-deployment family fits the card format without special-casing

The §11 cluster (AIR framework, IR/RR/over-representation/stopping-time) was added after round 1; every §11 page now has a card in the same format as §1–§10 pages. Fields like timing: post-deployment and formula: IR_G ≥ (β − γ^FR) / (γ^TR − γ^FR) make clear how the family differs from pre-deployment metrics — without visually breaking the ontology.

Incidence rate card
Incidence rate from reports — math-heavy, but the card still fits. Formula renders in-line, range and direction are explicit.

warn · content-layer debt

48 dead-link spans in §11 pages remain

Still unchanged from round 2. The §11 family's prose references concepts/* and sources/* pages that don't exist yet — the renderer correctly marks them as dead-link spans. Fixing them is a content task (create 14 missing concept/source pages), not a renderer bug. The cards themselves use the existing citations.

warn · small polish

A few card fields could use a tighter word budget

Two rows occasionally wrap awkwardly on narrow viewports (e.g. "five-dimensional ordinal score per speech act; aggregate to group mean"). They remain readable, but a future pass could trim to ≤ 70 characters per cell. Not a correctness issue.

new · future work

Add a "download card" action

Eval Cards allows exporting a card as a standalone artefact. We could emit a /{slug}/card.json endpoint per page and a small "Export card" link in the sidebar — so researchers can cite or embed the structured metric definition without scraping HTML. Easy next step; not required for educator use.


All ontology families at a glance (post-round-3)

FamilyMetricsCard coverage
§1 Deliberation quality77 / 7 with summary + card
§2 Bridging, representation, consensus1212 / 12
§3 Summarization faithfulness66 / 6
§4 Digital twin fidelity77 / 7
§5 Mechanism robustness1212 / 12
§6 Distributive justice88 / 8
§7 Meta-evaluation55 / 5
§8 Institutional design11 / 1
§9 Policy uptake11 / 1
§10 Tensions44 / 4
§11 Post-deployment evaluation5 + 1 hub6 / 6
Hub pages (_ontology, _mve-stack)22 / 2
Total7171 / 71 (100%)

Reproducing

cd wiki/frontend
rm -rf .astro dist       # cache-bust
npm run build
python3 -m http.server 8000 -d dist &

# Re-run the educator audit:
bash tests/scripts/tour.sh         # existing tour
# Or just populate the cards end-to-end:
python3 tests/scripts/add-cards.py  # idempotent; skips pages that already have a card

Round 3 auditor: Claude · 2026-04-17 · Artefacts under tests/v3-screenshots/, tests/v3-artifacts/, tests/v3-report-assets/. Reference document saved to wiki/raw/articles/evalcards-evalevalai-about.md.