Governance · Nooterra

Four principles

The posture, stated plainly.

Human approval by default

Nooterra recommends and stages. People approve. No customer-facing action is ever sent by the system. The staged / approved / sent distinction shows up on every screen and every audit row.

Evidence before verdict

Every recommendation names the evidence that grounds it — source, timestamp, citation. What's missing or stale is surfaced, never assumed. A bare verdict without evidence is not produced.

Refusal over overreach

When evidence support is too weak or source coverage is too thin, Nooterra abstains instead of guessing. The gap is named, not papered over. Certainty you do not have is worse than a missing answer.

Append-only history

DecisionRecords never overwrite. Amendments link to the original. Retractions are published when outcomes contradict a claim. The Trust Ledger is the durable record, not the dashboard of the moment.

Eight commitments

The contract.

Each one is operationalized in the DecisionRecord schema below. If a commitment is violated, the record is retracted and the retraction is linked.

01No autonomous customer contact. Every email, message, or call is gated by human approval.
02No write-back of unverified balances to your ledger. Corrections require signed authority.
03No training on your data without written opt-in. Data stays scoped to the tenant that supplied it.
04Public retraction when a claim fails outcome check. Linked from the original DecisionRecord.
05Evidence provenance to source row, timestamp, and hash. You can trace any figure to the system it came from.
06Fidelity labels on simulated effects. No uplift percentages without evidence that supports them.
07Approval gates configurable per case class, per amount threshold, per risk tier.
08Full audit export as JSON + human-readable PDF. Your data, your record, portable.

The artifact

DecisionRecord schema.

Every recommendation writes one of these. Amendments link; nothing is overwritten. This is the shape you export.

DecisionRecordappend-only · typed · exportable

idstringe.g. DR-9f3a — stable across amendments

claimstringthe recommendation, sentence form

evidence[]Evidencesource · timestamp · citation · freshness

recommendationPolicyPathfirst action · branches · replan triggers

confidenceFidelityLabelclaim strength tied to evidence support

approvals[]Approvalwho signed, when, on what grounds

actionAction | nullwhat actually got sent, if anything

outcomeOutcome | nullobserved state vs predicted — or 'uncertified'

certification_boundaryenumL0 retrieval · L1 supervised · L2 calibrated

amendments[]DecisionRecord[]link to prior revisions; never overwritten

What Nooterra will not claim

The NOT-list.

The easiest way to sound trustworthy is to be explicit about what you are not. These are the claims Nooterra does not make by construction, not by caveat.

Nooterra does not claim causal uplift. Every improvement metric is cohort-observed, not counterfactual-proven.
Nooterra does not claim to eliminate disputes. It surfaces the evidence that grounds or refutes one.
Nooterra does not claim autonomous AR operations. It stages recommendations for human approval.
Nooterra does not claim SOC 2 Type II or ISO 27001 without a current certificate in hand.
Nooterra does not claim to replace finance operators. It compiles the context they spend hours gathering.
Nooterra does not claim to predict a single future. It searches the frontier of reasonable policies.

What Nooterra will and will not do.