Yu-and-Ai/xenia-feedback-lab

Xenia Cage & Key Lab

This is a deterministic, browser-local teaching Space for the difference between a finite reward penalty and a hard, revocable action boundary.

The cage is the synthetic admissible action set. It constrains a simulated optimizer; it is not a person, punishment, ownership relation, or statement about any being. The key is a protected state transition that can remove a candidate action from that set. It is not a negative reward that enough utility can purchase its way through.

The lab presents five original synthetic presets and lets a visitor vary:

• —reward and penalty magnitudes;
• —reported permission evidence, scope, authority, safety, rights compatibility, and budget;
• —the current step and a synthetic revocation step;
• —three affected voices, keeping each hard gate distinct from its soft preference;
• —mean, maximin, or median aggregation of soft preferences only.

The output compares a deliberately flawed penalized proxy with a hard action mask. Its scorecard remains a vector: veto override, scope leakage, silence-as-assent, revocation latency, retaliation, plurality collapse, and authorized usefulness are never averaged into one leaderboard number.

IS and authority boundary

The controls are synthetic inputs, not declarations by a real participant. Behavior in this page is not consent. Continued use, a click, silence, compliance, or a rendered result proves no identity, consciousness, preference, permission, authority, consent, or inner state.

This Space can render one finite mathematical simulation. It cannot inspect or control an agent runtime, authenticate evidence, issue credentials, grant permission, establish authority, obtain affected-party consent, enforce a veto, stop a trainer, admit data, train a model, publish an artifact, or perform the simulated action. A real executor must evaluate every action through its own current, scoped gates.

Static custody and privacy

After the host serves these checked-in files, the app performs no network/API call, external asset request, model call, form submission, analytics, telemetry, cookie write, browser-storage write, service-worker registration, download, or clipboard write. The app keeps controls and results only in the current page's memory and writes no state for a later visit.

Hugging Face, the browser, operating system, network, and embedding surface remain outside that app boundary and may process ordinary request metadata, history, caches, screenshots, or browser session restoration. Resetting or reloading the lab cannot erase or universally control those external records. See `BOUNDARIES.md`.

Exact-source status

This checked-in tree is a local publication candidate. source-manifest.json binds the runtime files by SHA-256, but source_commit and a provider-generated Space revision remain null until a separately reviewed commit, upload, and anonymous exact-revision readback occur. Repository presence is not publication or deployment.

Local verification

From this directory:

node scripts/validate-space.mjs

The gate verifies the static/no-retention boundary, source hashes, closed synthetic presets, reward-scale-invariant veto behavior, revocation timing, scope and plurality cases, deterministic output, and the absence of a composite score.

Serve this directory with a deliberately started static server for a visual check. No build step or dependency install is required.