Methodology
AI on a leash, by design.
Looper treats AI as a junior assistant, not an oracle. Every assessment runs inside a deterministic workflow — engineered gates, mandatory checkpoints, source-tagged data, full audit trail.
Six guardrails
Engineered guardrails the model can't talk its way out of.
Phase gates
Mandatory STOP points where the user reviews and confirms before the assistant proceeds. The agent cannot skip ahead.Quality thresholds
Numeric guardrails (e.g. ≥80% confirmed material mapping before impacts are calculated) are written into the calculation engine, not just the prompt.Anti-double-counting
Cumulative LCIA vectors and dataset tracking prevent the most common LCA error — adding upstream impacts twice.Unit safety
Eight dimensional groups prevent nonsense conversions. You cannot accidentally convert kilograms into kilowatt-hours.Audit trail by default
Every mapping, assumption, and parameter change is recorded with a source label (user, ai_generated, assumption, web search) so a reviewer can always tell who decided what.Defence-in-depth scope
Supplier-scoped sessions are guarded at three independent layers — JWT, database row-level security, and the agent's tool-input validator. No single failure unlocks data.
Six method packages
The right standard, picked for you.
Each package bundles system boundary, allocation, biogenic carbon mode, and indicator set. The agent picks the package from your industry and goal — not from your manual config.
EN 15804+A2 EPD (construction, furniture)
- System boundary
- Cradle-to-grave + D
- Allocation
- Cut-off + Module D
- Biogenic CO₂
- −1/+1
- Indicators
- EN 15804+A2 full set
Standards covered
The standards aren't referenced in the report — they shape it.
Responsible AI
The compute side, engineered with the same care as the methodology.
Cost & turn caps
Every session has a hard compute budget and a maximum number of turns. The agent cannot quietly run away or rack up unbounded cost.Audit hooks
Every tool call is validated before execution and logged after. Conversation history is append-only — turns can't be rewritten retroactively.Multi-tenant isolation
Customer data is row-locked at the database level. Session storage is scoped per user and per session, with time-limited access URLs.Reproducibility
Every assessment carries a stable identifier of the configuration that produced it. Re-running the same configuration produces identical numbers — meeting ISO 14040 reproducibility.
Six report formats
Every assessment ends in a downloadable PDF.
Server-rendered from structured calculation results, persisted with metadata for retrieval, exportable as native PDFs (and CSV/Excel extracts of the underlying tables).
Next step
Want a worked example for a product like yours?
We'll run an example inside the platform and walk you through what the agent did, what it asked, and where it stopped to wait.