[leon1418]

Opened by the knowledge auto-update pipeline. Draft — a human decides.

What changed upstream

Grok 4.7 is now available on Amazon Bedrock in AWS GovCloud (US)

Verdict: new_knowledge on bedrock.model_catalog.grok

• was — No Grok/SpaceXAI model listed in Bedrock model selection or mapping references; dispatch exists only for Anthropic and OpenAI (others fall to generic provider_module_pending).
• now — Grok 4.7 (SpaceXAI) now available on Amazon Bedrock in AWS GovCloud (US), positioned for coding, agentic tasks, and knowledge work.

Still true: Existing Anthropic/OpenAI per-provider dispatch modules and the Gemini→Bedrock mapping tables remain correct for the providers they cover; this is an additive provider, not a correction.

This is why the old value is not simply overwritten.

Proposed edits

Each row states its justification. No CI check can catch a badly reworded judgment, so the
"why" column is the only protection a reviewer has — please read it rather than the diff alone.

gcp-to-aws/references/phases/clarify/clarify-ai-only.md:137 — derived judgment

- | FedRAMP           | GovCloud Bedrock only (us-gov-east-1/us-gov-west-1) — materially smaller model catalog; verify target model availability before committing the migration                                                                        |
+ | FedRAMP           | GovCloud Bedrock only (us-gov-east-1/us-gov-west-1) — materially smaller model catalog, now including Grok 4.7 (SpaceXAI); verify target model availability before committing the migration                                                                        |

Why: The announcement states Grok 4.7 is now available on Bedrock specifically in AWS GovCloud (US), which changes what 'materially smaller model catalog' includes for FedRAMP/GovCloud targets.

Evidence (verbatim from the announcement): "Grok 4.7 is now available on Amazon Bedrock in AWS GovCloud (US)"

gcp-to-aws/references/phases/design/design-ai.md:97 — derived judgment

- | `fedramp`        | Target region forced to **GovCloud** (us-gov-east-1/us-gov-west-1); re-run Step 0.5 regional validation against GovCloud — the model catalog is materially smaller, and a `regional_warnings[]` entry is REQUIRED for every candidate model not available there.                                                                                                                           
+ | `fedramp`        | Target region forced to **GovCloud** (us-gov-east-1/us-gov-west-1); re-run Step 0.5 regional validation against GovCloud — the model catalog is materially smaller (though it now includes Grok 4.7 (SpaceXAI)), and a `regional_warnings[]` entry is REQUIRED for every candidate model not available there.                                                                                                                           

Why: Grok 4.7 availability in GovCloud is new information relevant to regional validation against the GovCloud model catalog referenced here.

Evidence (verbatim from the announcement): "Grok 4.7 is now available on Amazon Bedrock in AWS GovCloud (US)"

gcp-to-aws/references/phases/discover/discover-preview.md:104 — value

- Terra, Luna, and GPT-5.4; AWS GovCloud us-gov-west-1 / us-gov-east-1 for Terra and Luna, us-gov-west-1 also for
+ Terra, Luna, GPT-5.4, and Grok 4.7; AWS GovCloud us-gov-west-1 / us-gov-east-1 for Terra, Luna, and Grok 4.7, us-gov-west-1 also for

Why: This line enumerates the models available in AWS GovCloud; the announcement adds Grok 4.7 to that list.

Evidence (verbatim from the announcement): "Grok 4.7 is now available on Amazon Bedrock in AWS GovCloud (US)"

Known limits of this proposal

• Blast radius is not stable between runs. Two runs of the same input returned 9 and 13
  locations and neither was a superset of the other, so this list may be incomplete.
• Paths are relative to plugins/aws-startup-advisor/skills.
• The pipeline did not touch any file the judge did not name.

[herosjourney]

Verdict

Don't merge as-is. The announcement is real, but the edits put Grok in the wrong places. One of the three hunks is a factual error, and the PR misses where this knowledge actually belongs once you account for #454, #450, and #449.

Description format

This description doesn't follow the format other aws-startup-advisor PRs use. Those start with the problem, then move to the fix:

• #449: Problem / impact, then Change
• #454: Gap, then What this PR adds
• #469 and #470: Problem / impact, then Change

This body opens with What changed upstream and a pipeline verdict, then jumps to proposed diffs. It never states what is wrong or missing in the skill today, or what the fix is supposed to accomplish. Please rewrite it that way: the problem first (there is no Grok 4.7 GovCloud fact, and where that fact has to live), then the fix (which files change, and why those files).

Problems in the diff

1. The discover-preview.md edit is wrong and should be dropped. That sentence belongs to the OpenAI same-model mapping table. It describes where the GPT-5.x mantle endpoint runs (Terra and Luna are GPT-5.6 variants). Adding Grok there makes three claims the source doesn't support:

• Grok uses the OpenAI mantle endpoint.
• Grok is available in commercial us-west-2.
• Grok is in both GovCloud regions.

The announcement only says "AWS GovCloud (US)". It sends readers to the Bedrock docs for regions, endpoints, and inference profiles. The "why" column says this line enumerates models available in GovCloud; it doesn't. It enumerates GPT-5.x mantle regions.

2. The two FedRAMP edits add a model name that the pipeline can't act on. In design-ai.md, the FedRAMP row already says to re-run regional validation against GovCloud and to warn on every candidate model that isn't there. Saying "now includes Grok 4.7" doesn't change that check. Grok also has no entry in the shared ai-model-lifecycle.md or the pricing cache. The lifecycle rule allows only active, listed models in recommendation rows, so Design can't select Grok anyway. The new text points at a model nothing can recommend, and it will go stale the next time the GovCloud catalog changes.

3. The azure-to-aws copies were not updated. azure-to-aws has the same FedRAMP guidance (design-ai.md and clarify-ai-only.md, "FedRAMP → GovCloud (smaller catalog)"). Editing only gcp-to-aws leaves the two migration skills saying different things. The PR already notes the file list isn't stable (9 versus 13 locations across runs); this is one of the files it missed.

Structural changes this doesn't account for

#454 — the agent-advisor model catalog is a separate file tree. The title is fix(agent-advisor), and the "was" line describes agent-advisor's per-provider dispatch. The diff touches no agent-advisor files. All three edits are in gcp-to-aws phase docs. #454 established that agent-advisor recommends from skills/agent-advisor/references/models/, which today has only Anthropic and OpenAI catalogs. Updating migration phase docs doesn't reach that catalog. #454's review rounds also showed that a catalog entry has to be connected to the selector and covered by routing tests, or it can never be chosen. If the goal is "agent-advisor knows about Grok in GovCloud", that work is missing. Otherwise retitle the PR so it no longer claims to be an agent-advisor fix.

#450 — the capability catalog describes sources, not targets. Leaving ai-capability-catalog.json alone is correct. Grok here is a Bedrock target, and that catalog lists source providers and gateways. Worth stating in the description so nobody later adds a spacexai provider entry by mistake.

#449 — shared AI knowledge lives in canonical files under skills/shared/ai/. Model facts are written once in a canonical shared file and copied byte-for-byte into each skill, with parity checks. A Grok fact belongs in the shared lifecycle table (status and GovCloud availability) and the Bedrock pricing cache, then synced into gcp-to-aws and azure-to-aws. This PR writes a model name directly into three per-skill phase files that have no parity check, which is the drift #470's Gemini body-parity test was added to prevent.

The "Still true" note mentions the Gemini→Bedrock tables. After #469 and #470 those tables exist twice and are tied by a body-parity test. This PR doesn't touch either copy, so it's safe on that point.

Suggested path

1. Drop the discover-preview.md hunk.
2. Drop or rewrite the two FedRAMP hunks so they don't name a model the lifecycle table can't select. If a mention stays, phrase it so it won't go stale ("verify current GovCloud availability"), and apply the same wording to the azure-to-aws equivalents.
3. Put the fact in the shared lifecycle and pricing references, using the Bedrock docs for regions, model ID, and endpoint, then sync the vendored copies.
4. If agent-advisor should recommend Grok, do that in a follow-up shaped like #454 (catalog entry, selector wiring, routing tests), or retitle this PR.

One note on the pipeline: a derived-judgment step that inserts a model into a nearby GPT-5.x region sentence is the error its own warning ("No CI check can catch a badly reworded judgment") predicts. A source-type check would catch it — the announcement doesn't name an endpoint, so don't edit endpoint sentences.