Orivoxx

Orivoxx — Benchmark Frontend Specification

Version: 2.0 · Status: source of truth for the attribution-era UI · Updated: 18 November 2026

Supersedes the list-first-only v1 contract. The visual language follows docs/UI_REFERENCE_NOTES.md (Coderia); page completeness follows docs/LANDING_REFERENCE_NOTES.md and docs/APP_HOME_REFERENCE_NOTES.md.

1. Product split

Surface Route Job
Marketing landing / Explain the problem, solution, measurement, and proof; drive to /app
App home /app Orient a returning user and expose the next action
Product surfaces /benchmarks, /submit, /results, /leaderboard, /compare, /docs, /privacy Do the work

The landing is not the app shell with a hero. The app is not a dashboard with five widgets and no narrative.

2. Visual language

Coderia-derived, dark-only, warm near-black.

Token Value Use
--bg-app #0B0B09 Page background
--bg-panel #111110 Cards, tables
--bg-raised #171715 Hover, selected row
--bg-inset #0E0E0C Code, terminal, trace
--border-subtle #262622 1px card/divider
--border-strong #3A3A34 Focus, hover
--text-primary #E1E1DB Primary text
--text-secondary #B0B0A8 Supporting text
--text-muted #8C8C85 Metadata
--accent #17CF97 Primary action, PASS, links
--accent-secondary #75A7FF Informational links
--status-fail #FF5C5C FAIL, critical
--status-drift #FF8A7A DRIFT
--status-unverified #F3C66D UNVERIFIED
--status-neutral #8C8C85 Not assessed

Type: Instrument Serif 400 for display and section headings (clamp(40px, 6vw, 64px), tracking −0.03em); Geist for body/UI (14–16px); Geist Mono for labels, ids, numbers. Mono eyebrow: 11px uppercase, 0.12em tracking, format \ 01 \ / \ SECTION \.

Rhythm: max width 1280px; section padding 90px desktop / 56px mobile; 4/8px base gaps; 1px borders; 12–16px panel radius; 999px primary buttons. No heavy shadows, no gradients except a static radial accent glow behind hero artifacts.

Motion: scroll reveal opacity + 12px translate over 500ms, once; hover border/background 150–250ms; artifact scale 0.98→1 over 600ms. No bounce, no infinite loops, no scroll-jacking. Respect reduced motion.

Status vocabulary remains PASS, FAIL, DRIFT, UNVERIFIED, plus CAPPED, SAMPLE, HARNESS, DEGRADED, REDACTED. Text labels always accompany colour.

3. Page inventory

3.1 / — marketing landing

Sections (all substantive, per LANDING_REFERENCE_NOTES.md):

  1. Header: wordmark, Benchmarks, Leaderboard, Docs, GitHub, Open app pill.
  2. Hero: hackathon badge, H1 “Benchmark the agent, not just the model.”, two-sentence subtext, View leaderboard + Read the docs, four pills (Deterministic · Model-agnostic · Failure attribution · Reproducible), real result/attribution artifact, three assurance chips.
  3. \ THE PROBLEM \ \ 01 \: “A trace tells you what happened. It does not tell you why.” Five-layer hand-off diagram; three named pains; closing line.
  4. \ THE SOLUTION \ \ 02 \: three pillars (deterministic scoring, counterfactual attribution, reproducibility) + embedded attribution card.
  5. \ HOW IT WORKS \ \ 03 \: five numbered steps (Spec, Run, Score, Attribute, Replay), each with a real UI crop.
  6. \ MEASUREMENT \ \ 04 \: six metric cards + three chaos modes + pass@k; note that metrics are the substrate.
  7. \ LEADERBOARD \ \ 05 \: four real rows, score mean±std, pass@k, reliability stacked bar; View full leaderboard →.
  8. \ ATTRIBUTION \ \ 06 \: real result snippet, evidence + proof line (47.3 → 89.1), View proof →.
  9. \ FOR DEVELOPERS \ \ 07 \: terminal block with orivoxx bench, orivoxx replay, orivoxx diff; copy control.
  10. \ RESOURCES \ \ 08 \: docs, GitHub, build log.
  11. FAQ: 6 honest items.
  12. Closing CTA: “Benchmark the agent. Prove the failure.” + tagline.
  13. Footer.

3.2 /app — app home

Per docs/APP_HOME_REFERENCE_NOTES.md:

  1. Orientation header: eyebrow, H1, one paragraph, status stats (tasks, adapters, results, attributed runs, pass rate), four quick actions.
  2. Leaderboard snapshot: 4 rows with reliability bars; View full leaderboard →.
  3. Task catalogue snapshot: 3 cards with version, difficulty, requirement count, passThreshold, best score.
  4. Recent runs: last 5 rows with adapter, task, model, score, exit reason, attribution cell, timestamp.
  5. How attribution works: baseline → intervention at k → replay → PASS diagram; link to a real proof.
  6. Quick start / CLI: three commands + terminal output, copy.
  7. Chaos modes: three cards + Run with chaos actions + reliability definition.
  8. Footer.

No heading-only sections. Real store data or explicit empty states.

3.3 /benchmarks

Catalogue with id, title, version triple, difficulty, requirement count, passThreshold, best score. Cards link to detail. One paragraph on version immutability.

3.4 /benchmarks/[task]

Header + version strip; requirement table (id, title, severity, bench, check); starting files; spec; adapters; task leaderboard; Run baseline and Run with chaos actions.

3.5 /submit

Task select, adapter select (fake self-serve; organizer adapters labelled disabled), fixture variant or chaos mode, disclosure of what executes. Fake run is synchronous; organizer runs are documented. No arbitrary command field.

3.6 /results/[id]

Above the fold:

Middle:

Footer: reproducibility note.

3.7 /results/[id]/trace

Full-width timeline: sticky step rail, event cards, filters (model/tool/error/chaos), summary header with totals. Every content-bearing field is a summary, labelled redacted.

3.8 /replay/[id]

3.9 /leaderboard

Columns: rank, adapter, task, model, artifact score mean±std, pass@k, steps, tool calls, tool error rate, tokens, cost, duration, reliability stacked bar (layer distribution), trend, chaos filter. Task/model/chaos filters; baselines toggle. Sorting total and stable.

3.10 /compare

Two results: metric deltas, trace diff of first N steps with divergence, attribution diff (python: harness × recovery × 3; aider: …), scores. Read-only.

3.11 /docs, /privacy, /404

Docs sidebar/breadcrumb/search with new Attribution, Replay & proof, Reliability, Metrics & traces, Resilience pages. Privacy updated for trace/action handling. 404 uses product shell.

4. States

State Surface Rule
Empty leaderboard, app home, compare Explain missing data, link to a next action
Degraded leaderboard/result DEGRADED chip + fallback store
Unavailable attribution/replay trace_unavailable / no_sufficient_intervention; never guess
Capped result Banner above score
Redacted trace Chip on every summarized field
Truncated trace/diff Notice with budget
Error submit Fixed envelope with request id

5. Accessibility and responsive

6. Non-goals