Framework · NIST AI RMF
From written policy to the action it governs.
The NIST AI RMF is voluntary guidance. elan1 is strongest where policy meets a consequential write: who may act, who must approve, what is refused.
The NIST AI Risk Management Framework organises AI risk work into four functions: Govern, Map, Measure and Manage. For AI agents that change records, the gap is usually between written policy and the action itself. elan1 closes it by checking each action against versioned rules and recording the outcome in a proof record.
elan1 does not certify you or itself against NIST AI RMF.
01What it asks
Four functions. GOVERN runs across the rest.
AI RMF 1.0 (NIST AI 100-1, January 2023) is voluntary. It organises AI risk work into four functions; profiles adapt it to a use.
- GOVERNPolicies, accountability and culture. It runs across the other three functions.
- MAPContext, and what could go wrong.
- MEASUREAnalyse, assess and track the risks that were mapped.
- MANAGEPrioritise and act, including after deployment.
- Generative AI ProfileNIST AI 600-1 (July 2024) adds a Generative AI Profile.
| Date | What applies |
|---|---|
| Jan 2023 | AI RMF 1.0 published as NIST AI 100-1. |
| Jul 2024 | Generative AI Profile published as NIST AI 600-1. |
Source: NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023; NIST AI 600-1, Generative Artificial Intelligence Profile, July 2024.
02How elan1 helps
Where elan1 acts, and the evidence it leaves.
Strongest at GOVERN and MANAGE, partial at MAP and MEASURE. We mark the partial rows as partial.
| Function | elan1 mechanism | Evidence | At the moment of action? |
|---|---|---|---|
| GOVERN | Declarations name who may act, who approves, what is never allowed | Declaration, actor, autonomy | Mode A enforced |
| MAP — partial | Each declaration states the write types it covers | F05, F12 | — |
| MEASURE — partial | Decide states a confidence per answer; refusals counted | Decision digest in F14 | — |
| MANAGE | Typed refusals, guardrail, autonomy ceiling, fail toward a human | Records of allows and refusals | Mode A enforced; Mode B recorded |
“Enforced” applies only in Mode A, where elan1 holds the write credential. In Mode B every write intent is evaluated and recorded; we call that attested advisory, never blocked.
03Not covered
What elan1 helps with, and what it does not.
Stated at equal weight, so nobody reads a gap as a feature.
Helps you produce evidence for
- GOVERN: who may act, who approves and what is never allowed, applied at the write
- MANAGE: typed refusals, a guardrail, an autonomy ceiling, and failing toward a human
- MEASURE (partial): a stated confidence per answer from elan1 Decide; refusals counted
- MAP (partial): the write types each declaration covers
Does not cover
- Risks in your models
- Bias and fairness testing
- Model test, evaluation, verification and validation
- Impact on people
- Culture
- Model-supplier risk
04Regulation mapper
Clause by clause, to the rung that acts.
Select a row to see which rung of the Right Path it maps to. Switch frameworks to compare; nothing here is a certification.
NIST AI RMF: Policies and accountability (GOVERN) maps to rung 4 evaluate().
| Requirement | elan1 control | Evidence produced | Not covered | Enforced at action time? |
|---|---|---|---|---|
| Rung 4 evaluate(): declarations name who may act, who approves, what is never allowed | Declaration, actor, autonomy | Culture and accountability structures outside the write | Yes, Mode A only | |
| Rung 5 GUARDRAIL and rung 6: typed refusals, guardrail, autonomy ceiling, fail toward a human | Records of allows and refusals | Incident response and decommissioning | Mode A enforced; Mode B recorded | |
| Decide states a confidence per answer; refusals counted at rung 9 | Decision digest in F14 | Model test, evaluation, verification and validation | After the action | |
| Rung 2 WRITE DOOR: each declaration states the write types it covers | F05, F12 | Context mapping and impacts on people | Recorded at the action, both modes | |
| None | None | Not covered: what could go wrong, and for whom, is yours to map | No. Outside elan1 | |
| None | None | Risks inside your models and model suppliers | No. Outside elan1 |
Where this is weak: Mode B records the write; it cannot stop it. “Enforced” applies only in Mode A, where elan1 holds the write credential.
Clause references summarise the official text. Not legal advice. elan1 holds no certificate against any framework. Source: NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023; NIST AI 600-1, July 2024.
Read the full page: NIST AI RMF: what it asks and how elan1 helps
05Worked example
One rule, its clauses, and the record it writes.
A treasury policy says agents never change where a vendor is paid. As a declaration, the policy is applied at the write, and every attempt leaves a record.
1 · The rule, in the declaration
# decl_vendor_master · v5 (illustrative) write_type: vendor.bank_change pre_gates: [G3_PERMISSIONS] never_by: actor_kind == agent approver: role:treasury_lead approver_not_requester: true on_refusal: typed_code, record, meter 0.4 record: always # refusals included
An agent proposed changing a vendor's bank account. The declaration says agents never make this change, so the write was refused with a typed code. The record was still written, and the meter shows 0.4 of a governed action.
2 · The clauses it serves
- GOVERNThe policy names who may act and what is never allowed.
- MANAGEThe risk is acted on at the write: refused, with a typed code.
- MEASUREEvery refusal is counted, at 0.4 of a governed action, so you can track how often the policy fires.
3 · The proof record it writes
- F05vendor.bank_change
- F06decl_vendor_master
- F075
- F09agt_specimen_ap01
- F10agent
- F11A1
- F12vend_specimen_0072
- F162027-07-14T10:18:02+05:30
Illustrative specimen. Not customer data.
06What an auditor sees
One record, each field mapped to its clause.
The rule that allowed or refused each action, who approved, and a record unchanged since. Refusals show the policy was applied, not only written.
- F01tenant_id
- tnt_specimen_01
- F02seq
- 1043
- F03prev_digest
- sha256:4be1…90c7
- F04write_id
- wr_specimen_7c3
- F05write_type
- purchase_invoice.post
- F06declaration_id
- decl_gst_purchase
- F07declaration_version
- 4
- F08policy_version
- 2027.06.1
- F09actor_id
- agt_specimen_ap01
- F10actor_kind
- agent
- F11autonomy_level
- A1
- F12record_id
- pinv_specimen_0418
- F13record_version
- 0
- F14payload_hash
- sha256:c03a…11fe
- F15lexicon_hash
- sha256:e2c4…0b16
- F16recorded_at
- 2027-07-14T10:44:51+05:30
G4_GST_RATE_MISMATCH (illustrative) · proof record written · metered: 0.4 governed actionIllustrative specimen. Not customer data.
Field notes · select a note to mark its fields
07FAQ
NIST AI RMF: questions we are asked.
It applies to agents as to any AI system, through four functions: Govern sets policy and accountability, Map frames the context and risks, Measure assesses them and Manage acts on them. For agents that change records, the hard part is applying written policy at the moment of action. elan1 supports Govern and Manage there, and is partial at Map and Measure.
No. The AI RMF 1.0, published by the US National Institute of Standards and Technology in January 2023 as NIST AI 100-1, is voluntary guidance. Organisations use it to structure their AI risk work; elan1 claims no conformance to it.
Govern and Manage at the moment of action, and Map and Measure in part. Versioned declarations, named approvers and typed refusals put written policy into each action, and every outcome leaves a proof record. Testing and evaluating the risks inside your models stays outside elan1.
No. The AI RMF is voluntary guidance, and we claim no conformance to it. elan1 supports the GOVERN and MANAGE functions where policy meets a consequential write, and is partial at MAP and MEASURE.
elan1 Decide states a confidence on every answer and elan1 Govern counts refusals. Measuring the risks inside your models — testing, evaluation, verification and validation, bias and fairness — is outside elan1.
A refusal carries a typed code from a frozen register and still writes a proof record. That shows the policy was applied at the moment of action, not only written down.
No. A model may produce a write intent; it may never produce an allow. evaluate() applies versioned rules, and rungs 3 to 8 cannot reach a model client.
NIST AI 600-1 is the Generative AI Profile of the AI RMF, published in July 2024. It describes risks that are specific to, or made worse by, generative AI and suggests actions for managing them under the four functions. elan1 helps only where a generative model's output becomes a write intent: that write is checked and recorded like any other, and risks inside the model and its suppliers stay with you.
Last reviewed
Reviewed against the official text. Not legal advice; your counsel decides how the law applies to you.
| Date | Change |
|---|---|
| 6 Oct 2026 | First version. Mapped the four functions to elan1 mechanisms, with MAP and MEASURE marked partial; listed what is not covered. |
| 10 Oct 2026 | Reviewed at launch: wording updated now that elan1 Govern, Work, Decide and Run are available. Clause references unchanged. |
// NEXT STEP
Find out where you stand, in five minutes.
Twelve questions. A plain result. Nothing is stored or sent, and no email is needed.