Capabilities and evidence
Practice you can inspect.
Glasshouse is a synthetic educational exercise. This engineering candidate has not passed independent public-release qualification.
Shannon Brown directs the product. Implementation and automated review are AI-assisted. No employer deployment, agency endorsement, certification, real loss savings or improved professional performance is claimed.
| Area | Implemented mechanism | Evidence and limitation |
|---|---|---|
| Decision integrity | One deterministic transition kernel; explicit posture, treatments, scope, authority and action lifecycle. | Automated semantic tests, full-state replay tests and 1,000-seed branch sweep. |
| Evidence | Source, observation and receipt times, corrections, authorized actor knowledge and immutable decision snapshots. | Known-then and correction fixtures. Interpretive reasoning remains for human review. |
| Recovery | Transactional local journal, writer lease, explicit takeover, validated backups and corruption quarantine. | Automated IndexedDB tests including aborted writes, stale saves and two tabs. Legacy records are preserved as read-only evidence. |
| Review | Eight observable dimensions, complete event trace, same-seed alternative, HTML, JSON and text PDF. | Process and modeled outcomes are separate. No global leadership grade. PDF tagging is not claimed; HTML is the accessible alternative. |
| Access | Account-free local play, pausable time, transcripts, keyboard controls, complete schematic and optional 3D. | Automated browser checks supplement pending manual NVDA, VoiceOver/Safari, device and accessibility reviews. |
| Learning impact | A repeatable exercise and reflection mechanism. | Independent rubric calibration, formative studies and learning transfer have not been established. |
How to interpret feedback
Structured rules can check whether evidence existed, approval was recorded, a control completed and a review condition was retained. They cannot establish the quality of free-form reasoning. “Not observed” is different from an observed omission. A good outcome cannot make a weak process sound.
The monetary ledger uses one synthetic exposure and one declared horizon. Avoided loss minus incremental treatment cost equals net modeled benefit. Zero-cost ROI is not applicable. Neither money nor completed practice establishes competence.
Local data and synthetic boundaries
The flagship sends no decisions or microphone data to a service. It needs no account, analytics, cloud model, microphone or paid backend. Settings provide backup and deletion of Glasshouse data only. Shared-origin browser storage is not isolation from neighboring portfolio applications; use fictional, minimal notes.
Client-side world data and unsigned records can be inspected or changed. Import replay checks and digests help detect accidents; they are not proof of authorship or tamper resistance.
Qualification still required
Three independent domain and exercise reviewers must calibrate the content and rubric. Ten formative participants must test entry, comprehension, replay and accessible completion. Named physical devices, assistive technologies, motion/audio states, offline failure recovery and a public rollback rehearsal require recorded evidence before G2 promotion.
Additional cases, facilitated sessions, networking, enterprise controls and AI expansion remain gated by the PRD. A later learning study would need separate consent and authorization.
Source and release history · Campaign scenarios · Glasshouse · Shannon Brown