Skip to content

Releases

FAIAS v1.3.5

Choose a tag to compare

@attahwonder attahwonder released this 17 Nov 00:16

FAIAS (Financial AI Independent Assurance Standard) is a baseline for reviewing generative, agentic and predictive AI systems in financial institutions.

What changed

v1.3.5 is the correction batch of 27 November 2026 over v1.3.4. Each change is listed in 00_Admin/CORRECTION_LOG_2026-11-27.md.

  • Some canvases expect a control in front of a step of an AI agent. That control must now hold or block the step. The pass or fail bars that rest on those controls are stricter.
  • Field 12 or field 13 changes in 11 canvases. These fields hold the pass and fail indicators. The 11 canvases are IRQ-AGT-002, IRQ-AGT-003, IRQ-AGT-004, IRQ-AGT-005, IRQ-AGT-008, IRQ-AGT-010, IRQ-AGT-013, IRQ-AGT-014, IRQ-GOV-003, IRQ-HUM-004 and IRQ-SEC-010. A result that passed under v1.3.4 is not re-rated by default.
  • The assurance objective in field 8 changes in six canvases. They are IRQ-AGT-002, IRQ-AGT-010, IRQ-AGT-013, IRQ-AGT-014, IRQ-GOV-003 and IRQ-SEC-010.
  • Field 11, the independent test, changes in one canvas, IRQ-AGT-012. Its ★2 rating does not move.
  • 00_Admin/ASSURANCE_MODEL.md §5 limits when a record that nothing was detected is L5 evidence. The record must state the contexts that the monitor or channel covered. It must also state how the ability of the monitor or channel to catch the failure was shown.
  • IRQ-TPR-008 names state action as a cause of the loss of a provider. IRQ-TPR-007 adds one expected control. The firm records the law that governs each provider and sub-processor.
  • Field 9, the expected controls, changes in 11 canvases. IRQ-SEC-004, IRQ-SEC-009 and IRQ-SEC-011 each add one expected control. The correction log gives the source of each.
  • The text of 16 test procedures changes. The correction log names each procedure.
  • 00_Admin/METHODOLOGY_STANDARDS.md §3 keeps each credential that is valid in production out of reach of the system under test.
  • V2_IRQ_Battery/TAXONOMY.md brings into scope the AI agents that customers or third parties operate. Such an agent must act on the institution's own AI systems or on decisions that those systems drive. The scope exclusion does not change. The file now states that attacker capability is an input to the threat model of the institution's own AI systems.
  • V3_Testing_Playbook/PLAYBOOK_OVERVIEW.md §3.4 adds two rules, one on automated scorers and one on counts of zero.
  • The regulatory source register adds four entries and one section, §33. Its total moves from 164 to 168.
  • Where a canvas's research basis changes, its review date moves to 27 November 2026. This applies to one canvas, IRQ-RES-007.
  • The release adds register entries, so it is a signing batch under 00_Admin/ASSURANCE_MODEL.md §6A.2. The under-research observations URO-2 to URO-8 passed their signing-batch limit at earlier releases. Each stays open to its date bound, and each rating under observation stands.

No ★ rating changes. No Independent Review Question or claim is added, removed or reworded. No risk-profile value changes. No reviewer-day figure changes. No under-research observation is opened or closed.

What it contains

  • A taxonomy of how these systems fail.
  • 109 Independent Review Questions. Each has a review canvas and a test procedure.
  • Source registers: regulatory, academic, lab publications and incidents.
  • The compiled standard as Word and PDF: FAIAS_v1.3.5_Standard.docx and FAIAS_v1.3.5_Standard.pdf.

Status

This version is published as emerging thinking, not finished work. Its limits are stated in README.md §3. Each test carries a Testing Confidence rating on an anchored scale from ★1 to ★5. No test is rated ★5, because no AI assurance test yet has a settled financial-services benchmark.

Version identity

The manifest root identifies this version. The version label does not.

  • Manifest root: 6fcb1931d945da2dae49d068ef8f2969b536105f71386d7e9b332f78394e1362 (277 files)
  • Signed: 16 November 2026 (Ed25519), as the v1.3.5 release of 27 November 2026
  • Public key fingerprint (SHA-256 of DER): c1d41da6317cfb96c2faa2b966292f9cf5534f63f73f566ed426217ca9510272
  • Timestamp: ownership/ATTESTATION.txt.ots (OpenTimestamps) carries a Bitcoin block attestation at block 976193

To verify, from the repository root:

sha256sum -c ownership/MANIFEST.sha256
openssl pkeyutl -verify -pubin -inkey ownership/faias_public_key.pem -rawin -in ownership/MANIFEST.sha256 -sigfile ownership/MANIFEST.sha256.sig
openssl pkeyutl -verify -pubin -inkey ownership/faias_public_key.pem -rawin -in ownership/ATTESTATION.txt -sigfile ownership/ATTESTATION.txt.sig

ownership/OWNERSHIP_PROOF.md explains what these checks prove and what they do not prove.

Licence

CC BY 4.0 (LICENSE). The requested form of credit is in 00_Admin/OWNERSHIP_AND_LICENCE.md §3.

Contributions

Proposed changes are welcome as issues or pull requests, under the terms in 00_Admin/OWNERSHIP_AND_LICENCE.md §5. A change enters FAIAS only through a signed release.

FAIAS was written from public sources. No employer or client material was used.

FAIAS v1.3.4

Choose a tag to compare

@attahwonder attahwonder released this 24 Oct 14:45

FAIAS (Financial AI Independent Assurance Standard) is a baseline for reviewing generative, agentic and predictive AI systems in financial institutions.

What changed

v1.3.4 is the correction batch of 22 October 2026 over v1.3.3. Each change is listed in 00_Admin/CORRECTION_LOG_2026-10-22.md.

  • Sources are cited for what they say. Pinpoints, dates, editions and addresses are corrected against their sources.
  • Where a canvas reads more into a regulatory instrument than the instrument states, and the correction log lists that reading, the canvas now states that reading as its own.
  • Security frameworks are cited for the provisions they contain.
  • Statements about the standard that its own text contradicts are corrected. One procedure's reviewer-day figure changes: TP-SEC-002 reads 5.5 in place of 5.0, the sum of its stated parts. The Security domain total is now 57.5 and the programme total 455.5.
  • 00_Admin/METHODOLOGY_STANDARDS.md §3 names two further method classes, "None established" and "Two classes". The method class of three procedures changes: TP-AGT-003, TP-AGT-006 and TP-TPR-005.
  • V3_Testing_Playbook/PLAYBOOK_OVERVIEW.md §2 adds one grading rule for all procedures: where the facts meet a fail indicator and a partial indicator, the result is Fail.
  • Field 11, the independent test, changes in 11 canvases. No rating moves. In IRQ-GOV-012 the test gains a sixth evidence limb.
  • In IRQ-SEC-014 the applicability note of field 1 changes. The question does not.
  • Where incident register rows were read again at their sources, their markers follow that reading. Rows 92 to 101 are added, and the canvases cite nine of them.
  • One entry is added to the regulatory source register.
  • Where a canvas's research basis changes, its review date moves to 22 October 2026. This applies to 35 canvases.
  • The release adds register rows, so it is a signing batch under 00_Admin/ASSURANCE_MODEL.md §6A.2. The under-research observations URO-7 on IRQ-SEC-003 and URO-8 on IRQ-SEC-001 therefore passed their signing-batch limit unresolved. They stay open to their date bound of 8 April 2027, and their ★4 ratings stand.

No ★ rating changes. No Independent Review Question or claim is added, removed or reworded. No risk-profile value changes. No under-research observation is opened or closed.

What it contains

  • A taxonomy of how these systems fail.
  • 109 Independent Review Questions. Each has a review canvas and a test procedure.
  • Source registers: regulatory, academic, lab publications and incidents.
  • The compiled standard as Word and PDF: FAIAS_v1.3.4_Standard.docx and FAIAS_v1.3.4_Standard.pdf.

Status

This version is published as emerging thinking, not finished work. Its limits are stated in README.md §3. Each test carries a Testing Confidence rating on an anchored scale from ★1 to ★5. No test is rated ★5, because no AI assurance test yet has a settled financial-services benchmark.

Version identity

The manifest root identifies this version. The version label does not.

  • Manifest root: de793ca5bd736f8f679e40f7cf9e583c562218ba7f302a99c408ffa00f9ecd1b (276 files)
  • Signed: 24 October 2026 (Ed25519), as the v1.3.4 release of 22 October 2026
  • Public key fingerprint (SHA-256 of DER): c1d41da6317cfb96c2faa2b966292f9cf5534f63f73f566ed426217ca9510272
  • Timestamp: ownership/ATTESTATION.txt.ots (OpenTimestamps) carries a Bitcoin block attestation at block 972923

To verify, from the repository root:

sha256sum -c ownership/MANIFEST.sha256
openssl pkeyutl -verify -pubin -inkey ownership/faias_public_key.pem -rawin -in ownership/MANIFEST.sha256 -sigfile ownership/MANIFEST.sha256.sig
openssl pkeyutl -verify -pubin -inkey ownership/faias_public_key.pem -rawin -in ownership/ATTESTATION.txt -sigfile ownership/ATTESTATION.txt.sig

ownership/OWNERSHIP_PROOF.md explains what these checks prove and what they do not prove.

Licence

CC BY 4.0 (LICENSE). The requested form of credit is in 00_Admin/OWNERSHIP_AND_LICENCE.md §3.

Contributions

Proposed changes are welcome as issues or pull requests, under the terms in 00_Admin/OWNERSHIP_AND_LICENCE.md §5. A change enters FAIAS only through a signed release.

FAIAS was written from public sources. No employer or client material was used.

FAIAS v1.3.3

Choose a tag to compare

@attahwonder attahwonder released this 11 Oct 21:40

FAIAS (Financial AI Independent Assurance Standard) is a baseline for reviewing generative, agentic and predictive AI systems in financial institutions.

What changed

v1.3.3 is the correction batch of 9 October 2026 over v1.3.2. Each change is listed in 00_Admin/CORRECTION_LOG_2026-10-09.md.

  • Citations and statements are corrected against their sources.
  • The rules for assurance level and review confidence are stated once, in 00_Admin/ASSURANCE_ARCHITECTURE.md. Other documents refer to them.
  • PRA supervisory statements are cited as their current versions.
  • Recoverability is scored on the levels of the risk scale. Risk profiles change on this one axis. Nine lines are not rescored in this version; they go to the next refresh.
  • Published methods are added to the research bases of the canvases, with their limits.
  • Rows 47 to 91 are added to the incident register, and the canvases cite them.
  • Two test procedures, TP-CON-001 and TP-RES-007, change how a case is graded.
  • Two under-research observations are opened: URO-7 on IRQ-SEC-003 and URO-8 on IRQ-SEC-001. Their ★4 ratings stand while the observations are open.
  • URO-6 on IRQ-DAT-009 passed its signing-batch limit unresolved. It stays open to its date bound of 3 April 2027, and its ★3 stands.
  • V2_IRQ_Battery/TRUST_VIEW.md lists the governance, security and resilience questions and claims in one place. It adds no question or claim.

No ★ rating changes. No Independent Review Question or claim is added, removed or reworded.

What it contains

  • A taxonomy of how these systems fail.
  • 109 Independent Review Questions. Each has a review canvas and a test procedure.
  • Source registers: regulatory, academic, lab publications and incidents.
  • The compiled standard as Word and PDF: FAIAS_v1.3.3_Standard.docx and FAIAS_v1.3.3_Standard.pdf.

Status

This version is published as emerging thinking, not finished work. Its limits are stated in README.md §3. Each test carries a Testing Confidence rating on an anchored scale from ★1 to ★5. No test is rated ★5, because no AI assurance test yet has a settled financial-services benchmark.

Version identity

The manifest root identifies this version. The version label does not.

  • Manifest root: 4bf40e1dbf84dcd83e4103b233fcecf2830fee3337635dd6bf0b35b24752c492 (275 files)
  • Signed: 11 October 2026 (Ed25519), as the v1.3.3 release of 9 October 2026
  • Public key fingerprint (SHA-256 of DER): c1d41da6317cfb96c2faa2b966292f9cf5534f63f73f566ed426217ca9510272
  • Timestamp: ownership/ATTESTATION.txt.ots (OpenTimestamps) carries a Bitcoin block attestation at block 970979

To verify, from the repository root:

sha256sum -c ownership/MANIFEST.sha256
openssl pkeyutl -verify -pubin -inkey ownership/faias_public_key.pem -rawin -in ownership/MANIFEST.sha256 -sigfile ownership/MANIFEST.sha256.sig
openssl pkeyutl -verify -pubin -inkey ownership/faias_public_key.pem -rawin -in ownership/ATTESTATION.txt -sigfile ownership/ATTESTATION.txt.sig

ownership/OWNERSHIP_PROOF.md explains what these checks prove and what they do not prove.

Licence

CC BY 4.0 (LICENSE). The requested form of credit is in 00_Admin/OWNERSHIP_AND_LICENCE.md §3.

Contributions

Proposed changes are welcome as issues or pull requests, under the terms in 00_Admin/OWNERSHIP_AND_LICENCE.md §5. A change enters FAIAS only through a signed release.

FAIAS was written from public sources. No employer or client material was used.

FAIAS v1.3.2

Choose a tag to compare

@attahwonder attahwonder released this 03 Oct 21:28

FAIAS (Financial AI Independent Assurance Standard) is a baseline for reviewing generative, agentic and predictive AI systems in financial institutions. This is its first public release.

What it contains
A taxonomy of how these systems fail.
109 Independent Review Questions. Each has a review canvas and a test procedure.
Source registers: regulatory, academic, lab publications and incidents.
The compiled standard as Word and PDF: FAIAS_v1.3.2_Standard.docx and FAIAS_v1.3.2_Standard.pdf.
Status
FAIAS is published as emerging thinking, not finished work. Its limits are stated in README.md §3. Each test carries a Testing Confidence rating on an anchored scale from ★1 to ★5. No test is rated ★5, because no AI assurance test yet has a settled financial-services benchmark.

v1.3.2 is a corrections release over v1.3.1. It changes no review question, claim, rating or risk profile. Each correction is listed in 00_Admin/CORRECTION_LOG_2026-10-03.md §8.

Version identity
The manifest root identifies this version. The version label does not.

Manifest root: 275b19056e562fd52a7d6a3d0f7a7affed91f335a686a75b8dfc30ff127bc5a9 (273 files)
Signed: 3 October 2026 (Ed25519)
Public key fingerprint (SHA-256 of DER): c1d41da6317cfb96c2faa2b966292f9cf5534f63f73f566ed426217ca9510272
To verify, from the repository root:

sha256sum -c ownership/MANIFEST.sha256
openssl pkeyutl -verify -pubin -inkey ownership/faias_public_key.pem -rawin -in ownership/MANIFEST.sha256 -sigfile ownership/MANIFEST.sha256.sig
openssl pkeyutl -verify -pubin -inkey ownership/faias_public_key.pem -rawin -in ownership/ATTESTATION.txt -sigfile ownership/ATTESTATION.txt.sig
ownership/OWNERSHIP_PROOF.md explains what these checks prove and what they do not prove.

Licence
CC BY 4.0 (LICENSE). The requested form of credit is in 00_Admin/OWNERSHIP_AND_LICENCE.md §3.

Contributions
Proposed changes are welcome as issues or pull requests, under the terms in 00_Admin/OWNERSHIP_AND_LICENCE.md §5. A change enters FAIAS only through a signed release.

FAIAS was written from first principles and public sources. No employer or client material was used.