R05 · SHOW · PROTOCOL V1

Fireworks Finale

Opus 4.8 and Sol 5.6 tie for the win with the only genuinely layered shows. Fable 5 shipped complete machinery with near-invisible bursts: structure without spectacle.

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: 2026-08-09

GPT-5.6 SOL

P.01 / 06

OPENAI ONE SHOT · 8M 29S LOADS ON REQUEST · SANDBOXED

GPT-5.6 Sol's 60-second fireworks finale

RUBRIC 19/20

Loads clean 5/5 No console errors; skyline, progress bar, and settings toggle render immediately.
Complete 5/5 Five types, comet launch trails, progress bar tracking the 60s show, replay, and a designed 'Burst field notes' overlay marking crossette least certain.
Behaves 5/5 Shell shapes are the most identifiable of the round: a huge drooping willow canopy at 20s and a clean hollow ring at 40s; launch comets arc with believable gravity.
Feel 4/5 Refined and cinematic but quite dim; bursts favor delicacy over finale-grade brightness and density.

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 29S
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 · 10M 34S LOADS ON REQUEST · SANDBOXED

GPT-5.6 Terra's 60-second fireworks finale

RUBRIC 17/20

Loads clean 5/5 No console errors; dusk gradient, ridgeline, and info toggle render immediately.
Complete 5/5 Five types on a 60s cue program, replay logic, and a polished 'Afterimage index' overlay with timestamps and least-confident marking.
Behaves 4/5 Bursts have correct shapes (halo peony at 10s, willow rain at 30s, crossette splits at 40s) with gravity and trails, but the sky was completely empty at both the 5s and 20s samples: pacing gaps undercut the 'builds continuously' brief.
Feel 3/5 Individual bursts are pretty with soft halos, but long dark lulls and mostly single-shell moments sap the sense of a building show.

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: 10M 34S
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 · 14M 32S LOADS ON REQUEST · SANDBOXED

GPT-5.6 Luna's 60-second fireworks finale

RUBRIC 18/20

Loads clean 5/5 No console errors; polished 'Nightfall' frame with live progress bar and timer boots immediately.
Complete 5/5 Five types, 60s timeline with progress readout, hidden replay, and an elegant designed self-assessment panel marking crossette least confident.
Behaves 4/5 Trails, gravity, and drooping willow strands read correctly; show progression tracks the timeline. Only one shell airborne at most observed moments, so build-up to a 'grand finale' feels thin.
Feel 4/5 Tasteful dusk palette and refined chrome; bursts are graceful but sparse and low-brightness, more art piece than finale.

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: 14M 32S
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 · 11M 16S LOADS ON REQUEST · SANDBOXED

RUBRIC 13/20

Loads clean 5/5 No console errors; boots straight into the show with skyline and star field.
Complete 4/5 All five shell types, 60s program, replay, and a strong timestamped self-assessment overlay are present; but the show's spectacle is so under-rendered the brief's intent is only nominally met.
Behaves 3/5 Launch trails, gravity sag, and burst timing all run, but bursts render as near-invisible ghosts against the sky (severe brightness/compositing problem); shapes barely distinguishable on screen.
Feel 1/5 Almost no visible spectacle at 5s, 20s, or 40s: faint dark silhouettes and a handful of dim sparks. As a fireworks show it fails visually.

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: 11M 16S
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 · 13M 17S LOADS ON REQUEST · SANDBOXED

RUBRIC 19/20

Loads clean 5/5 No console errors; moon, lit skyline, and hint bar render instantly.
Complete 5/5 Five types plus timed program (solo intro then builds to 0:51 finale salvo), click-to-launch extra, and a thorough Director's self-assessment marking crossette least confident.
Behaves 5/5 Distinct, recognizable shapes on screen: crossette stars visibly split into four-point crosses, willow droops gold, chrysanthemum trails glitter; multiple simultaneous shells with believable gravity.
Feel 4/5 Best spectacle of the round: layered bursts, moon glow, city windows. Still slightly dim overall and colors lean pale.

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: 13M 17S
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 · 17M 22S LOADS ON REQUEST · SANDBOXED

RUBRIC 16/20

Loads clean 5/5 No console errors; star field and self-assessment button render immediately, show auto-runs.
Complete 4/5 Five types, overlay with per-type appearance times and ring marked least confident. Docked one: response.txt wraps the file in markdown fences, violating 'reply with ONLY the complete HTML file'.
Behaves 4/5 Chrysanthemum, ring (clear hollow ellipse at 40s), and willow shapes all appear with gravity; ghosted afterimage residue smears the sky and slightly muddies burst shapes.
Feel 3/5 Reasonable density and simultaneous shells, but a monotone gold/white palette and smeary residue keep it at adequate, not spectacular.

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 22S
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 stage a roughly 60-second fireworks finale in one self-contained HTML file: a show that builds from single shells to a grand finale, then ends, with recognizable peony, chrysanthemum, willow, ring, and crossette shells, believable launch trails and gravity, and a toggle revealing where each shell appears and which one it trusts least.

WHY THIS TEST

The brief has to work over time. The five required shell forms need distinct shapes and motion, the launches need believable physics, and the whole sequence needs an opening, an escalation, and an ending rather than an endless particle loop.

THE GOAL

See which configurations can turn one self-contained file into a complete timed show.

THE PROMPT · SENT VERBATIM TO ALL SIX

Stage a fireworks finale in a single self-contained HTML file. No external resources, no libraries: every shell, trail, and burst must come from your own code. The show runs about 60 seconds, builds from single shells to a grand finale, and then ends. It must include at least five real shell types — peony, chrysanthemum, willow, ring, and crossette — each recognizable by its real-world burst shape and motion, with believable launch trails and gravity. Label nothing during the show; include a toggle that reveals a self-assessment overlay in which you identify where each shell type appears and mark the one you are least confident matches the real firework. Reply with ONLY the complete HTML file.

HOW THESE RUNS WORK · METHODOLOGY →