Your review desk is ready
Add a prediction or the guided cases to begin.
What does this person do?
Describe the work, then see both the suggestion and the next step.
Try a sample description
Adds three fictional cases. Open Review to inspect the accepted suggestion and two escalations.
Scope: selected software, finance, healthcare, education, and trades occupations. Unrelated text can still receive a supported label. Use fictional, non-sensitive text.
A suggestion is not a final label
Unresolved records keep their final labels empty. Records belong to this browser tab.
| Record | Suggestion | Confidence | State | Final label | Source |
|---|
Review controls
Roles are simulated. Anyone can switch identities. This is not authentication.
Export completed human feedback
JSON contains resolved human decisions and history. It does not retrain the model.
Session storage and reset
This free hosted version uses temporary sessionStorage in this browser tab. There is no SQLite server, real identity, shared queue, cross-user lock, or tamper-proof audit. Closing the tab normally ends its session; browser restore settings may retain it. Clear it explicitly when finished.
Measured on held-out occupations
173 task statements from 10 occupations. Benchmark results, not measured business outcomes.
Frozen policy: auto-accept at 52% or above; ordinary review from 50% to below 52%; escalate below 50%. No known vocabulary always escalates. Thresholds use validation only. Probabilities are uncalibrated.
| Category | Precision | Recall | F1 | Test tasks |
|---|
Read the weak spot: SOFTWARE recall is 53.8% on just 13 tasks from one occupation. The overall score does not establish real-job accuracy.
Complete evaluation report
The five categories are project decisions. Other occupations are excluded; there is no OTHER class. Related wording can remain after exact deduplication. The test has limited independent occupational diversity.
Open the model, weights, provenance, and full model card ↗Follow the decision through the system
This hosted Static Space: downloaded JSON weights → JavaScript TF-IDF/logistic regression → simulated review rules → browser sessionStorage. No backend or paid compute. Updates are synchronous in one tab; they do not provide multi-user transactional guarantees.
The original Python companion uses Gradio and SQLite. The diagram below describes that companion, not the storage architecture of this Static Space.
Download the original architecture diagramWhat carries across
The same trained weights, probabilities, frozen routing thresholds, withheld unresolved labels, senior-only escalations, immutable final decisions, reasons, provenance, and simulated role boundaries. Browser inference is parity-tested against Python.
What remains proposed
Authenticated identities, server-enforced permissions, durable shared storage, backups, cross-user concurrency, SLA alerts, reassignment, reopening, and retraining. There were no stakeholder interviews or business sign-off.
A business process, with a model inside it
Independent case study with simulated business requirements. Prepared for PabloRod11. A hypothetical workforce analytics team needs to classify descriptions without silently turning uncertain model guesses into final business labels.
Users and exception paths
An ordinary reviewer handles routine items, a senior adjudicator handles escalations and unsupported work, and an audit reader inspects decisions. All identities here are simulated. Missing input is rejected; unknown vocabulary escalates; an unresolved record has no final label; out-of-scope closure issues no category.
Traceable requirements
- R01 — Reproduce the trained model. Browser probabilities must match Python within 1e-10. Test: browser/Python parity fixture over all validation and test texts plus token edge cases.
- R02 — Withhold unresolved labels. Human routes stay empty until resolution. Test: guided-case workflow and human override.
- R03 — Enforce simulated review rules. Claim before deciding, require a reason, restrict escalations to seniors, reject stale revisions and completed edits. Test: workflow denial and revision checks.
- R04 — Preserve decision history. A failed browser storage write must leave in-memory state unchanged. Test: injected persistence failure. This is one-tab behavior, not a database transaction guarantee.
- R05 — Handle input and output safely. Reject blank and overlong text; escalate zero-vocabulary input; render user text as text. Test: input guards and browser rendering.
- R06 — Disclose limitations. Simulated requirements, browser storage, taxonomy scope, uncalibrated scores, and weak category performance stay visible. Check: manual hosted walkthrough.
Failure → correction → retest
The original notebook silently routed NaN to escalation. The Python version now rejects malformed confidence and tests boundaries. A later review found a wrong standalone model path; the loader was corrected and retested. This browser port additionally checks every held-out probability against the Python implementation.
Three-minute demonstration
- Add guided cases and open Review.
- Try claiming an escalation as ordinary reviewer; the queue stays intact and explains the denial.
- Switch to senior, claim, give a reason, and resolve. Inspect HUMAN provenance and the trail.
- Export feedback; explain that it does not automatically retrain.
- Use Evidence and System to explain measured results and deployment tradeoffs.
Download the full Python project, BSA requirements, tests, and design review