ARF Correction & Retraction Policy

If your organization is scored by the Agentic Readiness Framework

(ARF) and you believe the score is incorrect, this page describes how

to request correction or retraction.

What you can request

1. Factual correction: a scanner result or judge verdict is

demonstrably wrong (e.g., we report you have no OpenAPI spec when

your spec is publicly served at /openapi.json).

1. ARL stage dispute (V8 only): you disagree with the ARL 0-9

stage assigned to a specific (org, check) pair. We will re-grade

that cell against the criterion file (backend/research/v8/standards/arl9/checks/{check_id}.md)

and the evidence we captured. If our grade was wrong, we re-score

that check. If the criterion file itself is wrong (e.g., the ARL 8

row cites an artifact that doesn't exist), we route the issue to

the criterion-file review process — see §V8-A below.

1. Methodology dispute: you disagree with the definition of a

check or the weighting of a dimension. We will record the dispute

and respond, but methodology changes go through our published

versioning process and may not result in re-scoring.

1. Retraction: you do not consent to your organization being

publicly scored or ranked.

How to request

Email corrections@getjackfruit.ai with:

• Your organization's primary domain
• The scanrunid you are disputing (visible at

https://getjackfruit.ai/<your-domain> in the footer)

• The specific check_id(s) and what evidence you believe was missed
• For retraction requests: a statement that you are an authorized

representative

You may also submit a dispute via GitHub issue at

https://github.com/levyido12/ARF/issues with the label

correction-request.

What happens next

For factual corrections we cannot reproduce or that depend on

non-public evidence (workspace artifacts, internal docs), we will

note the dispute on the public record and pause score publication

until we receive supporting evidence under NDA.

What we will do

• Confirmed factual error: we re-score the org with the corrected

evidence and publish the updated score. The prior score is moved

to a "corrected_at" record visible in the public history.

• Methodology dispute we accept: we record the dispute publicly

and consider it in the next framework revision.

• Retraction: we remove your organization from public benchmark

pages within 21 business days. Industry-aggregate statistics may

retain anonymized contribution unless you explicitly request full

deletion.

What we will not do

• Modify a score without underlying evidence change
• Suppress methodology disputes that are simply uncomfortable
• Re-score an organization more than 4 times per year except for

confirmed factual errors

Legal

• GDPR: if you are a controller or processor with GDPR exposure,

this policy also serves as the Article 16 (right to rectification)

pathway. EU-domiciled organizations may also email

gdpr@getjackfruit.ai for accelerated handling.

• Defamation / misrepresentation: claims of defamatory misuse of

the score should be sent to legal@getjackfruit.ai with the

specific public statement at issue.

V8-A: ARL stage disputes (added 2026-05-23)

V8 introduces a 10-stage ARL (Agentic Readiness Level) ladder from

ARL 0 to ARL 9. Two new dispute categories specific to V8:

Stage-grade dispute: you believe our grade is wrong for a

specific (org, check). The artifact we used to grade is the per-check

criterion file at backend/research/v8/standards/arl9/checks/{check_id}.md

plus the scanner / judge evidence captured in check_results. To

dispute:

• Cite the check_id and the ARL stage you believe is correct
• Point to the row in the criterion file's ## ARL stage criteria

table whose description you believe matches your evidence

• Attach the evidence (URL, screenshot, repo path, or workspace

artifact under NDA)

We will re-grade against the cited row and respond. The criterion

file is treated as the rubric of record; we do not re-grade against

a different rubric without a methodology dispute.

Criterion-file dispute: you believe the criterion file itself is

wrong — most commonly because an ARL 8 or ARL 9 row cites an

artifact that doesn't exist (e.g., a SDK package name the LLM panel

invented). The 205 LLM-authored criterion files carry a banner

warning that ARL 8/9 artifact names may be placeholders; this is

the dispute path for those cases.

If we accept a criterion-file dispute:

• That ARL stage row is marked pending-review in the criterion file
• All published V8 grades that depended on the disputed row are

flagged on the public record; the affected (org, check) cells

display "criterion under review" until the file is corrected

• We do not silently retroactively change V8 grades — every change

appears in the score history

A V8 grade is published only if Chief produces `chiefverdictv8 ∈

{publish}`. The Chief's integrity guards (criterion-SHA fingerprint,

per-judge anchor calibration, dual-schema persistence) gate the V8

verdict separately from the V7 verdict. If chiefverdictv8 is

criteriondrift or judgeunreliable, no V8 grade is published

for that scan_run — the V7 MPS is shown alone with an explicit "V8

not yet validated for this scan" badge. This is the safety net

behind §V8-A.

Pre-pilot status: V8 is in pilot as of this writing. We do not

publish V8 grades on the public benchmark pages. The above sections

apply when V8 grades become public.

Versioning

This policy is versioned. The current version is dated at the top of

the page. Substantive changes (other than typos or contact updates)

are noted in docs/CHANGELOG.md and announced via the same channels

as scoring methodology changes.