[sa-anyan]

On the same M&E, the reactor came out at 15.6 m³, 125 m³ and 156 m³ on different runs. The AI re-estimated its sizing inputs on later runs (a pricing repair, a model change, a cleared row) and nothing compared the new size with the earlier one. Monte Carlo also never varied AI guesses, because they were saved without a range.

Guardrail (backend/blocks/equipment/estimate_guard.py)

Each item's first AI-derived size is anchored together with a digest of its M&E data (the item and its connected streams). Every later AI size is measured against that anchor:

Case	Result
Same M&E, change ≤ 25%	Kept (stable). Compared with the original anchor, so drift never creeps.
Same M&E, change > 25%	Held. The anchored estimate is priced; the new one is saved with its size, its drift and the assumptions behind it.
M&E changed, or size unit changed	Re-anchored. That is a legitimate change.
User entered any sizing input	The user owns it: re-anchored, never held.
Manual price, or no size	Left alone.

• Equipment page: an "AI estimate check" lists held items, for example "Used: 15.6 m3 · New AI estimate: 156 m3 (+900%)", with Keep and Use buttons. Use makes the new estimate the anchor.
• What's covered: only equipment sizing. It is the one AI path that re-estimates on later runs. OPEX AI values are never replaced once written (OPEX proposes cited reference defaults without the AI #61/OPEX AI researches in the background from Entry approval #62), and CAPEX analyst positions change only when the analyst's inputs change (Fast background analyst: every stage researched in parallel from upload #64).

Monte Carlo varies AI predictions

• AI ranges: each AI sizing assumption now returns low/high: the engineering range for that service, not a fixed percentage. Units are converted, and the range is dropped if the value falls outside it.
• Prompt version: bumped to …-27-assumption-ranges, so older cached answers aren't reused.
• No change to Monte Carlo itself: discovery already varies any input with a saved range.
• Disagreement becomes uncertainty: when two AI runs disagree, the shared assumption's range widens to cover both, for example productivity 1–2 kg/m³/h.
• One small fix: discovery no longer crashes when a price basis was saved with a null currency.

Live check (real AI, the user's workbook, a copy of the database)

Check	Result
AI equipment run	Every AI-sized item anchored. Ranges returned, for example residence time 24 h (4–168 h), working fraction 0.85 (0.7–0.95)
Second AI run	Every item stable, 0% drift
Monte Carlo discovery	13 of 17 AI sizing inputs varied. The 4 fixed are densities from M&E measurements.

Tests

• Backend pytest (Python 3.12.10): 4,675 passed, 3 skipped, 0 failures. The Monte Carlo and guard tests were rerun after the null-currency fix: 70 passed.
• New test_estimate_guard.py (8 tests) covers:
  • anchoring, holding a jump (with the restored inputs and the widened range), drift within tolerance, re-anchoring when the M&E or user changes, items left alone;
  • range unit conversion and validity;
  • Monte Carlo varying an AI assumption that has a range;
  • wiring: the equipment run prices the anchored size and recalculates once.
• Frontend: tsc OK, node tests 64/64 (held list, keep and use), next build OK.

Not covered

• Existing studies: they have no anchor yet. Their next AI equipment run sets one. The 15.6, 125 and 156 m³ values already saved in the user's studies aren't corrected; the user should check the reactor basis once.
• Cross-study anchors: each study anchors on its own. Re-uploading the same workbook as a new study starts fresh, apart from the AI cache.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
flowvalue-ai	Ready	Preview	Nov 14, 2026 11:52pm UTC