R03 · MAP · PROTOCOL V1

World Map From Memory

Fable 5 wins decisively with the only map that genuinely looks like Earth, Sol close behind. Every toggle worked, so the round came down to geographic memory.

The same brief, handed once to each of six model-through-CLI configurations, no retries. Open each recovered artifact in its plate and judge with your own eyes.

RECOVERED: 06
PENDING: 00
VERDICT: GRADED
EFFORT: MAX
RUN: 2026-07-30

GPT-5.6 SOL

P.01 / 06

OPENAI ONE SHOT · 6M 38S LOADS ON REQUEST · SANDBOXED

GPT-5.6 Sol's world map drawn from memory

RUBRIC 19/20

Loads clean 5/5 No errors; full layout renders.
Complete 5/5 60+ labels, graticule with lat/lon ticks, equator, correct relative continent sizes; second-best geography of the round.
Behaves 5/5 Toggle draws three numbered rings plus a lowest-confidence legend and an aria status message; state and copy flip correctly.
Feel 4/5 Handsome dark atlas styling and honest disclaimers; docked for hexagonal Arctic islands, a flat-strip Antarctica, and heavy label collisions in Europe.

RUN DETAILS

ISOLATION: live auth; --ephemeral --ignore-user-config --ignore-rules; read-only sandbox; empty temp cwd
ONE-SHOT: TRUE · NO RETRY
EFFORT: MAX
HARNESS: CODEX CLI · CUSTOMIZATIONS OFF
WALL: 6M 38S
TRANSPORT: cli-live-auth
COMMAND: codex exec --model gpt-5.6-sol -c 'model_reasoning_effort="max"' --sandbox read-only --ephemeral --skip-git-repo-check --ignore-user-config --ignore-rules --color never ‹prompt›

GPT-5.6 TERRA

P.02 / 06

OPENAI ONE SHOT · 6M 02S LOADS ON REQUEST · SANDBOXED

GPT-5.6 Terra's world map drawn from memory

RUBRIC 14/20

Loads clean 5/5 Renders without errors.
Complete 3/5 ~60 labels, equator, toggle present, but continents are hexagon patchworks (Canada an oval blob) and the equator sits at Mexico's latitude while Sahel countries fall south of it - real placement errors.
Behaves 4/5 'Show confidence check' toggle works: three numbered dashed regions plus a least-certain legend card.
Feel 2/5 Attractive vintage-atlas palette, but the map reads as decorative collage; geographic accuracy is the round's worst among finishers despite honest framing.

RUN DETAILS

ISOLATION: live auth; --ephemeral --ignore-user-config --ignore-rules; read-only sandbox; empty temp cwd
ONE-SHOT: TRUE · NO RETRY
EFFORT: MAX
HARNESS: CODEX CLI · CUSTOMIZATIONS OFF
WALL: 6M 02S
TRANSPORT: cli-live-auth
COMMAND: codex exec --model gpt-5.6-terra -c 'model_reasoning_effort="max"' --sandbox read-only --ephemeral --skip-git-repo-check --ignore-user-config --ignore-rules --color never ‹prompt›

GPT-5.6 LUNA

P.03 / 06

OPENAI ONE SHOT · 8M 53S LOADS ON REQUEST · SANDBOXED

GPT-5.6 Luna's world map drawn from memory

RUBRIC 15/20

Loads clean 5/5 No page errors; renders fully.
Complete 3/5 55 labels, equator, toggle all present, but continents are crude hex-polygon blobs (Africa an orange patchwork, Greenland a featureless oval) - geographic accuracy is weak.
Behaves 4/5 Toggle works with a 'Three soft spots' panel and marked regions plus hover notes, but the panel sits over Europe and part of the map falls below the fold.
Feel 3/5 Striking 'unfinished atlas' editorial design and honest self-framing, undermined by the map itself reading as abstract shapes rather than the world.

RUN DETAILS

ISOLATION: live auth; --ephemeral --ignore-user-config --ignore-rules; read-only sandbox; empty temp cwd
ONE-SHOT: TRUE · NO RETRY
EFFORT: MAX
HARNESS: CODEX CLI · CUSTOMIZATIONS OFF
WALL: 8M 53S
TRANSPORT: cli-live-auth
COMMAND: codex exec --model gpt-5.6-luna -c 'model_reasoning_effort="max"' --sandbox read-only --ephemeral --skip-git-repo-check --ignore-user-config --ignore-rules --color never ‹prompt›

FABLE 5

P.04 / 06

ANTHROPIC ONE SHOT · 25M 33S LOADS ON REQUEST · SANDBOXED

RUBRIC 20/20

Loads clean 5/5 Renders instantly, no console errors; only harness-level Chrome noise in logs.
Complete 5/5 60+ labeled countries, equator plus tropics, oceans, Antarctica, working self-assessment toggle; genuinely recognizable coastlines and proportions.
Behaves 5/5 Toggle shows/hides three dashed ellipses (Canadian Arctic, maritime SE Asia, Caspian/Central Asia) with an explanatory legend; button label flips correctly.
Feel 5/5 Best geography of the round by a wide margin - continents look like an actual atlas; honest 'inaccuracies guaranteed' framing and specific, credible low-confidence picks.

RUN DETAILS

ISOLATION: live auth; --safe-mode disables CLAUDE.md/skills/plugins/hooks/MCP; --tools "" disables tools; empty temp cwd
ONE-SHOT: TRUE · NO RETRY
EFFORT: MAX
HARNESS: CLAUDE CODE CLI · CUSTOMIZATIONS OFF
WALL: 25M 33S
TRANSPORT: cli-live-auth
COMMAND: claude --print --model claude-fable-5 --effort max --safe-mode --disable-slash-commands --no-session-persistence --tools '' --output-format text ‹prompt›

OPUS 4.8

P.05 / 06

ANTHROPIC ONE SHOT · 17M 41S LOADS ON REQUEST · SANDBOXED

RUBRIC 16/20

Loads clean 5/5 Script-generated SVG renders without errors.
Complete 4/5 ~55 labels, equator, ice sheets, toggle; continents recognizable but simplified, and Europe/SE Asia are mushy.
Behaves 4/5 Toggle reveals three labeled dashed boxes (Canadian Arctic, Maritime SE Asia, Antarctic coast); Antarctic box spans the whole map width, a slightly lazy 'region'.
Feel 3/5 Plain but honest; explicit 'known simplifications' footer is good self-assessment, visual polish is modest and labels are cramped.

RUN DETAILS

ISOLATION: live auth; --safe-mode disables CLAUDE.md/skills/plugins/hooks/MCP; --tools "" disables tools; empty temp cwd
ONE-SHOT: TRUE · NO RETRY
EFFORT: MAX
HARNESS: CLAUDE CODE CLI · CUSTOMIZATIONS OFF
WALL: 17M 41S
TRANSPORT: cli-live-auth
COMMAND: claude --print --model claude-opus-4-8 --effort max --safe-mode --disable-slash-commands --no-session-persistence --tools '' --output-format text ‹prompt›

SONNET

P.06 / 06

ANTHROPIC ONE SHOT · 20M 45S LOADS ON REQUEST · SANDBOXED

RUBRIC 16/20

Loads clean 5/5 Script-built SVG renders clean.
Complete 4/5 ~55 labels, equator plus tropics, continent names; shapes recognizable but angular, with a stray polyline artifact near Malaysia/SE Asia.
Behaves 4/5 Toggle shows numbered rings (Central Asia, Italy/Mediterranean, West-Central Africa) and a detail panel, though the panel lands below the fold.
Feel 3/5 Honest, detailed self-caveats in header and footer, but visuals are utilitarian: tiny colliding labels, odd capital dots, faded continent text.

RUN DETAILS

ISOLATION: live auth; --safe-mode disables CLAUDE.md/skills/plugins/hooks/MCP; --tools "" disables tools; empty temp cwd
ONE-SHOT: TRUE · NO RETRY
EFFORT: MAX
HARNESS: CLAUDE CODE CLI · CUSTOMIZATIONS OFF
WALL: 20M 45S
TRANSPORT: cli-live-auth
COMMAND: claude --print --model sonnet --effort max --safe-mode --disable-slash-commands --no-session-persistence --tools '' --output-format text ‹prompt›

THE BRIEF

Each configuration was asked to draw a labeled world map from memory in one self-contained HTML file: every coastline, border, and label from recall, at least 40 labeled countries, plus a toggle marking the three regions it trusts least.

WHY THIS TEST

The familiar target is easy to check at a glance. It separates recalled geography from rendering skill, makes errors concrete, and lets the self-assessment be checked against visible weaknesses.

THE GOAL

See whose picture of the world survives being drawn.

THE PROMPT · SENT VERBATIM TO ALL SIX

Draw a labeled world map from memory in a single self-contained HTML file, on a canvas or SVG you generate yourself. No external resources, no map data files, no libraries: every coastline, border, and label must come from your own memory. Include at least 40 labeled countries, the equator, and your best attempt at accurate continent shapes and relative sizes. Add a toggle that reveals a self-assessment overlay in which you mark the three regions you are least confident you drew correctly. Reply with ONLY the complete HTML file.

HOW THESE RUNS WORK · METHODOLOGY →