Self-Explanation Experiment — Demo Launcher

Configure and launch any condition. All links include clear=1; data saves through the Cloudflare Worker into card-games-staging/results_gallery/. Append &save=local to force a browser download for dev tests.

⚠ You opened this file directly from disk (`file://`)

The experiment loads data via fetch(), which browsers block on file:// URLs. The launch buttons below are disabled until you serve this page over HTTP. From the card-games/ folder, run:

python3 -m http.server 8080

Then open http://localhost:8080/self-explanation-experiment/demo.html

Quick Guide: What to Demo (v2 mature design)

The v2 condition set has 4 locked conditions along the verbalisation axis (none / description / counterfactual / periodic rule articulation). peer and token-type are kept in the codebase for archival replication but are deprecated for new pilots.

Condition	What participant does	Key theoretical question
`silent` control	View gallery, classify silently	Baseline: rule inference without verbalisation
`describe` control	Describe at least 3 cards before classifying	Attention control: does verbalising attention alone help?
`type` explain	Periodic recorded rule articulation at trial checkpoints	Top-down hypothesis externalisation
`token` explain	For each hand, recorded counterfactual reasoning (which cards would have to change to flip it).	Contrastive token-level reasoning
`token-type` deprecated	Per-hand counterfactual + periodic rule articulation (v1)	(Not in v2 design)
`peer` deprecated	Judge a fictional participant's verdict, then explain (v1)	(Not in v2 design)

Feedback defaults to off in v2. tokenSelection defaults to off in v2 (was on in v1). Toggle either via the parameter builder for legacy A/B comparisons.

SE Snapshot

Required for v2. The snapshot drives rule selection, exemplars, and the adversarial trial pool. Build via python -m se_compare.build_snapshot_from_pilot ... from rule-gallery/analysis/.

Loading snapshot manifest...

🔍 Rule Explorer (per-rule diagnostic-hands viewer)

Per rule, shows the hands each of the 4 strategies produces and how Gemini Pro/Flash classified each.

Open Rule Explorer →

Adversarial-hand strategy

Drives ?adversarialStrategy= on every preset launch and the custom-launch URL. Pilot accuracies are Gemini Pro / Flash, mean across the 10 rules in this snapshot.

A · Entropy (default) — probes representational ambiguity (LLM 0.83/0.81) B · Misclassification — probes collective error (0.79/0.56) C · Flip from random winning — probes feature-locality (0.84/0.87) D · Flip from exemplar — probes feature-locality from gallery (0.90/0.75)

Quick-Launch Presets

Custom Configuration

Condition

Group

Feedback

Token Card Selection

Re-engagement

Tutorial

Trials / Rule

Rules

Gallery Time (s)

Min Record (s)

Checkpoints

Accumulate History

Describe Min Cards

Launch Experiment

Console DevTools

Open the browser console (Cmd+Option+J) and type these commands:

`DevTools.skip()`	Skip current trial/screen (clicks Winning or Continue)
`DevTools.skipAll()`	Blast through multiple screens rapidly
`DevTools.skipTutorial()`	Mark tutorial done and reload (jump to game)
`DevTools.skipGallery()`	Skip gallery countdown
`DevTools.info()`	Show current condition, phase, and state
`DevTools.setCondition('type')`	Reload with a different condition
`DevTools.fastMode()`	Reload with galleryTime=3, nTrials=3, nRules=2