fix(C01,C02,C03,C10,AppC): improve testability and remove ambiguous l…

[vtknightmare]

…anguage

Preface rule: 'conformance can be objectively validated through testing, inspection, or audit.'

• AppC AC.3.6: remove 'tokenized' (NLP tokenization is not the same as data-security tokenization; the word was technically incorrect)
• AppC AC.4.3: define 'critical security finding' as CVSS >= 9.0 or organizational severity policy; without a threshold the control is not auditable
• AppC AC.6.2: scope fine-tuning to orgs controlling model training infra; SaaS tool users cannot implement fine-tuning
• C01 1.2.1: replace 'assessed and addressed' with minimum documented deliverable (impact assessment + chosen mitigation)
• C01 1.4.5: add 'documented set of adversarially relevant input subgroups' so evaluation scope is verifiable
• C01 1.4.6: require chosen defense and tuning rationale to be recorded alongside model artifact; 'appropriate' was not auditable
• C02 2.1.6: replace 'regional legal constraints' (200+ jurisdictions) with 'regional content-policy classification' (configurable attribute)
• C02 2.3.1: remove 'illegal requests' (jurisdiction-dependent); replace with 'requests violating organizational content policy'
• C03 3.2.5: add minimum test criteria (prompt injection rejection and tool output sanitization) so 'covers' is verifiable
• C10 10.3.3: clarify Origin and Host are validated independently; was ambiguous about whether failing one header was sufficient

[RicoKomenda]

Same as the other PRs in this batch: the diff has 5 of the 10 changes listed (the Appendix C edits, C01 1.2.1, and C02 2.1.6 aren't here). Were those meant to be pushed?

The present ones are mostly good. C01 1.4.6, C03 3.2.5, and C10 10.3.3 are clear improvements. Two notes:

• C02 2.3.1: dropping "illegal requests" makes sense, but the appended "Requests violating organizational content policy are also covered" is vague and overlaps 2.3.3, which already covers policy-violating inputs. Might be cleaner to drop the sentence.
• C01 1.4.5: "retained per the audit log requirements" doesn't have a clear anchor in C01 (no general audit-log requirement there). Also worth keeping the old threat examples (mimicking trusted style/demographic), they made the intent concrete.

[vtknightmare]

Same as the other PRs in this batch: the diff has 5 of the 10 changes listed (the Appendix C edits, C01 1.2.1, and C02 2.1.6 aren't here). Were those meant to be pushed?

The present ones are mostly good. C01 1.4.6, C03 3.2.5, and C10 10.3.3 are clear improvements. Two notes:

* C02 2.3.1: dropping "illegal requests" makes sense, but the appended "Requests violating organizational content policy are also covered" is vague and overlaps 2.3.3, which already covers policy-violating inputs. Might be cleaner to drop the sentence.

* C01 1.4.5: "retained per the audit log requirements" doesn't have a clear anchor in C01 (no general audit-log requirement there). Also worth keeping the old threat examples (mimicking trusted style/demographic), they made the intent concrete.

Same partial-diff fix: AppC edits, C01 1.2.1, and C02 2.1.6 are now included.
On 2.3.1: dropped the appended sentence, it overlaps 2.3.3 as you said. Just removes "illegal requests":
Verify that every inbound prompt is scored by a content classifier for violence, self-harm, hate, and sexual content against configurable thresholds, and that prompts exceeding those thresholds are rejected or sanitized before reaching model context.
On 1.4.5: kept the threat examples, dropped the audit-log phrase (no anchor in C01), added testable timing:
Verify that models used in security-relevant decisions (e.g., abuse detection, fraud scoring, automated trust decisions) are evaluated, prior to deployment and after any significant model update, for systematic bias patterns that an adversary could exploit to evade controls (e.g., mimicking a trusted language style or demographic pattern to bypass detection).
Force-pushed.

[RicoKomenda]

All ten changes are in now, and both earlier points are addressed: 1.4.5 keeps the threat examples and drops the audit-log reference, and 2.3.1 drops the overlapping sentence. The new ones are solid too. AC.3.6 fixing the "tokenized" misnomer, AC.4.3's CVSS >= 9.0 threshold, and C02 2.1.6 moving off "regional legal constraints" are all good calls. No remaining concerns from me.