PREVIEW PROTOTYPE — Cloudflare Access-protected draft generation. AI outputs are NOT independently assured. Do not enter confidential or regulated information.
Authenticated Pages extraction reads supported files in memory and imports locatable evidence into the selected generation source. Extraction is not semantic verification; a human must validate claims and citations.
Select multiple reference files and choose Extract & import evidence. The current generation request limit remains in force. Extracted source text is inserted into the generation context; source claims still require human validation.
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.
Human reviewer feedback & recalibration
Paste a consolidated human review or individual criterion comments. Feedback guides a fresh generation and independent assurance review; it never overrides critical gates. Recalibration uses the paid AI backend.
Use Generate & Validate for a normal run. Recalibrate sends the entered feedback with the same task inputs.
Rubric composition targets
Rubric counts are advisory targets. Automatic repair retries validation failures up to three times. Unresolved assurance failures block release. Existing scoring and security rules are unchanged. Domain-specific dimension checks remain subject to independent assurance.
Authenticated draft generation uses the Pages Function and backend Worker. No automated semantic assurance; all generated artifacts require human verification.
Project & evidence intake
Evaluation hierarchy: Reference Materials → Prompt → Golden Deliverable → Model Output. Higher-authority evidence prevails. Authenticated Pages extraction supports PDF/DOCX/XLSX and text. Review imported evidence and complete evidence-based assurance checks.
Local autosave uses this browser. Clicking Generate & Validate sends entered Prompt, Reference Material and Golden Deliverable text to the protected backend and OpenAI for draft generation. Export JSON for backup.
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 3.1 · Single-file Pages-ready · Human-verifiable audit trail · No unsupported automatic compliance claims