asingh15/qwen36-27b-amazon-candidate-diversity-explorer

Amazon Qwen27 candidate-diversity explorer

Public explorer. Repository visibility and source-publication eligibility are managed by the repository owner. The bundle validator enforces technical integrity and redaction safeguards; it does not make a licensing determination.

This static explorer makes the completed Qwen3.6-27B candidate-diversity experiment easy to inspect. It presents a deterministic 250-target showcase with all 40 candidates for each of three unprivileged variants:

• —the original unprivileged baseline;
• —latent-state directions inferred from the reviewer's demonstrations; and
• —generated review-type directions instantiated as candidate reviews.

Each target includes its product context, reference review, eight demonstrations, per-target diversity metrics, candidate responses, judge scores, and—where the collection generated them—the complete plan fields and one normalized reviewer summary for each demonstration count 0 / 1 / 2 / 4 / 8. Raw judge reasoning, duplicate generation wrappers, and source UID/product/post identifiers are deliberately omitted.

The showcase is category-balanced and selected independently of model outcomes. Within each Amazon category, targets are ordered by the SHA-256 of a fixed versioned salt plus the internal target identity, and the category quota is taken from that order. Scores, paired outcomes, and spread quantiles do not affect membership. Outcome and quantile labels remain available as filters for inspection, so the showcase must not be used to estimate their population prevalence. Aggregate cards and plots use the full 15,168-target leakage-safe report, not the showcase.

The browser loads only local files. It has no server component, external API calls, analytics, credential storage, or runtime secrets. data/manifest.json pins the exact inventory and SHA-256 of every file plus the certified source bundle digest. The local validator rejects extra files, changed content, symbolic links, raw source-identifier fields, common credential formats, and absolute private filesystem paths.

Build and validate locally with:

python scripts/build_amazon_qwen27_candidate_diversity_space.py build
python scripts/build_amazon_qwen27_candidate_diversity_space.py validate

Publishing and repository visibility are intentionally handled by a separate deployment command.