| Internet-Draft | Pre-Silicon Evidence Profile | November 2026 |
| Watts | Expires 20 May 2027 | [Page] |
Hardware security designs are increasingly assessed before fabrication: by gate-level leakage simulation, fault-injection simulation, formal proofs and equivalence checking. The results of these assessments are routinely copied into reports, slide decks and certification packages with no binding to the design that was assessed, the stimulus that was used, the conditions under which the numbers hold, or the platform that produced them. This document defines an Entity Attestation Token (EAT) profile for carrying such pre-silicon assessment results as signed Evidence. Every claim is classified as identifying the assessed artifact, describing a measured behavior, or describing the measurement conditions. A provenance claim, which states whether the signing key is rooted in hardware or software, is mandatory, and a simulated Trusted Platform Module is always reported as a software root. The profile is explicitly not a certification format: it records what was measured and under which envelope, so that a Relying Party can decide what the numbers are evidence of.¶
This note is to be removed before publishing as an RFC.¶
A reference implementation (schema "aiid-evidence/v1", JSON encoding, TPM 2.0 or software-key signing) is in the ai-id-engine repository, files hwid/src/evidence.rs and docs/EVIDENCE.md. This draft describes that implementation and proposes how it maps onto EAT. It is experimental.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 20 May 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document.¶
The RATS architecture [RFC9334] describes how an Attester produces Evidence about itself, how a Verifier appraises that Evidence, and how a Relying Party consumes the resulting Attestation Results. The Entity Attestation Token [RFC9711] provides a claims-based encoding for Evidence and allows profiles that constrain and extend it for a particular use.¶
This document addresses a use that existing profiles do not cover: Evidence about a hardware design, produced before that design exists as silicon. Typical sources are:¶
Such results are only meaningful relative to the exact netlist, cell library, power model, stimulus set and statistical parameters that produced them. A first-order |t| value of 3.15 says nothing on its own. Together with "4000 traces per group, 4000 samples, alpha 0.05, Bonferroni threshold 4.37, sky130 typical corner, zero-delay cell models, noiseless", it becomes a reproducible statement with a known minimum detectable effect. This profile binds the numbers to those conditions under a signature, and records the trust root of that signature honestly.¶
This profile does not define how to perform any of these assessments, and it does not turn their results into a certification. A report that conforms to this profile can be evidence for a certification process; it is never itself a certificate.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
A report is a body of named fields plus an attestation over the body. Each field carries exactly one class:¶
The classification is part of the signed body. A Verifier MUST reject a report in which:¶
The classes let a Relying Party apply policy mechanically. For example, it can accept behavior claims only when the identity claims match a netlist digest it has independently obtained, and only when the neither claims show a trace count at or above its own minimum.¶
| Name | Class | Content |
|---|---|---|
| provenance | identity | Signing root; see Section 3. |
| hwid | identity | Salted identifier of the assessment host. |
| model_digest | identity | Map from artifact path to SHA-256 [FIPS180-4] digest. It covers the RTL, netlists, power-model netlists and capacitance files, cell models and the cell library. |
| stimulus_set | neither | Digests of the stimulus generator, harness and configuration, plus seeds, modes and trace counts per group. |
| tvla_parameters | neither | Profile name, assessed design, traces per group, samples per trace (M), alpha, the first- and second-order critical values, and the number of second-order tests. |
| measurement_envelope | neither | See Section 2.3. |
| tvla_result | behavior | Null-control and positive-control peaks, first- and second-order peaks, minimum detectable effect, verdict and gate exit code. |
Additional fields MAY be present (for example area, latency, fault-injection counts or equivalence status). Each additional field MUST carry a class.¶
A leakage statistic without controls cannot be interpreted. The null control (random versus random inputs on the assessed design) bounds the false-positive behavior of the harness. The positive control (fixed versus random on an unprotected reference) shows that the harness can see a leak. A conforming tvla_result MUST include both control results. A Verifier MUST NOT treat a report as evidence of absence of leakage if the null control exceeded its threshold or the positive control did not.¶
The measurement_envelope states what the behavior claims are evidence of. It MUST contain:¶
The list of effects not modelled is REQUIRED because it is the part of a pre-silicon report most often lost when results are quoted elsewhere.¶
The provenance field records the root of the key that signs the report. Its value contains:¶
The following rules apply:¶
The body is serialized canonically: compact JSON [RFC8259] with fields in the order written. The report nonce is:¶
nonce = SHA-256( "AIID-EVIDENCE-V1" || canonical(body) )¶
The nonce is used as the freshness value of the platform attestation.¶
The report is the triple (body, body hash, attestation). A Verifier MUST:¶
This section describes how the model maps onto an EAT [RFC9711] profile, so that reports can travel in RATS conceptual messages.¶
A Verifier producing Attestation Results from such Evidence SHOULD:¶
A Relying Party SHOULD compare model_digest against artifacts it obtained independently, for example from a reproducible build whose digests are themselves published.¶
The following abbreviated body comes from the reference implementation. It describes a 2-share masked Montgomery multiplier at 4000 traces per group, which fails second-order TVLA as expected. Digests are truncated.¶
{"schema":"aiid-evidence/v1","created":"2026-11-16T16:10:00Z",
"fields":[
{"name":"provenance","class":"identity","value":{
"attestation_kind":"software-fallback",
"source":"none: no TPM, software-fallback key file",
"tpm_manufacturer":"","simulated_tpm":false,"root":"software"}},
{"name":"hwid","class":"identity","value":"9f2c...e1"},
{"name":"model_digest","class":"identity","value":{
"build/sky130_masked_v2.v":"4b1e...07",
"build/tvla_masked_v2.caps":"c3a9...5d"}},
{"name":"stimulus_set","class":"neither","value":{
"files":{"config/tvla.toml":"77d0...2a"},
"description":{"fixed_vs_random":{"seed":17,
"traces_per_group":4000}}}},
{"name":"tvla_parameters","class":"neither","value":{
"profile":"ci","traces_per_group":4000,"samples":4000,
"alpha":0.05,"crit_first_order":4.37,"crit_second_order":4.46}},
{"name":"measurement_envelope","class":"neither","value":{
"kind":"simulated gate-level power traces; not silicon",
"library":"sky130_fd_sc_hd__tt_025C_1v80",
"not_modelled":["glitches","coupling and IR drop","EM"]}},
{"name":"tvla_result","class":"behavior","value":{
"null_control":{"peak":3.24,"crit":4.37,"passed":true},
"positive_control_plain":{"peak":82.96,"detected":true},
"first_order_peak":3.15,"second_order_peak":26.50,
"mde_sigma":0.0977,"exit":2,
"verdict":"second-order leak in 2-share design"}}]}
¶
This profile improves the binding and honesty of pre-silicon results. It does not make them stronger than the assessment that produced them.¶
The hwid field is a salted identifier of the assessment host. Different salts yield unlinkable identifiers. Implementations SHOULD use a per-deployment salt and MUST NOT include raw hardware identifiers. Artifact paths in model_digest can reveal project structure, so publishers MAY replace them with opaque labels if the digests remain stable.¶
This document uses the profile identifier "tag:ai-id-engine,2026:presilicon-evidence/v1". No IANA action is requested for the identifier itself.¶
[RFC Editor: remove before publication.]¶
The ai-id-engine project implements this profile in JSON form:¶
The EAT and COSE encoding of Section 5 is not yet implemented.¶