PRIVATE DATA WARNING — Preview interface only. AI generation, authentication and secure file handling are NOT connected. Do not enter or upload confidential information.
Optional local parsing server reads Office/PDF files in memory. Extraction is not semantic verification; a human must validate claims and citations. No uploaded file is retained by the server.
Start the optional server with python server.py and open this page at http://127.0.0.1:8765/index.html. The standalone page still works without the server.
Generation Context Router
Auto-detects the authoritative generation pathway before any artifact generation. Manual overrides must have the required inputs. Source provenance and unresolved conflicts block release.
Enter inputs and select Detect route.
Start the local Python server with OPENAI_API_KEY configured to enable AI generation, independent AI review, and bounded refinement. Unverified outputs remain blocked from production release.
Project & evidence intake
Evaluation hierarchy: Reference Materials → Prompt → Golden Deliverable → Model Output. Higher-authority evidence prevails. This browser-only engine does not pretend to understand DOCX/XLSX contents; upload or paste extracted text and complete evidence-based reviews.
Local autosave uses this browser only. Export JSON for backup or transfer. Files are not uploaded to a server.
Rubric import & criterion evaluation
Import a CSV with existing columns preserved, including ID, Score, Criterion, Justification, Scoring for True Statements, Scoring for False Statements and Dimension. Additional columns are retained on export.
Frontier task-design audit
Assess each principle using actual task artifacts. A PASS must be supported by evidence, not inferred from keywords or claimed difficulty.
Trap validation
Every trap must be discoverable from supplied evidence, have sufficient evidence for one defensible answer, and test reasoning rather than guessing.
Difficulty classification
Execution checklist
Set each checkpoint to PASS, FAIL or REVIEW. FAIL and REVIEW automatically appear in the follow-up register. Automated structural tests cannot substitute for evidence inspection.
Failed checks & follow-up register
Unresolved failures and reviews from all layers appear here. Assign owners, actions and due dates; mark resolved only after rechecking evidence.
Evaluation report
Version 3.0 · Standalone/offline · Human-verifiable audit trail · No unsupported automatic compliance claims