[ShadowOfTime1]

Closes #52

Knowledge format. The MIP files now follow the format MIP-12 got in #56, so all five read the same way: **Proposal status:**, **Network status:**, a dated **Verified:** line, and a ## Sources section. Canonical specs are pinned to a commit (b49e903), and activation evidence is sourced separately from the spec:

MIP	Proposal status	Network status	Activation source
MIP-3	Final	Active since MONAD_NINE (mainnet March 19, 2026; testnet March 10)	v0.13.0 release notes
MIP-4	Final	Active since MONAD_NINE (same dates)	v0.13.0 release notes
MIP-7	Draft	No activation evidence in this bundle	release notes list none
MIP-8	Final	Active since MONAD_TEN (mainnet September 2, 2026; testnet August 12)	v0.16.1 / v0.16.0 release notes
MIP-12	unchanged from #56

MIP-12 is untouched, so the #56 correction and its tests stay as they are. MIP-8's old "Activation" and "Canonical specification" sections are folded into the same header and Sources section rather than kept twice. I filled in the status and evidence lines only; the TODO(author) motivation sections in MIP-3/4/7 are a content rewrite and stay out of scope.

Request behavior. src/lib/ai/instructions.ts builds the instructions server-side and always appends two things after the configured system prompt, so a custom prompt from Edge Config cannot drop them:

• an evidence rule: proposal status and network activation are separate facts; a Final spec alone does not prove a change is live; say a MIP is activated only with separately sourced evidence and name the network and date; when evidence is missing or conflicting, say the bundle cannot confirm activation and link the canonical spec; link the bundle's source and qualify facts with the file's verification date;
• the current UTC date, worded as today's date and kept apart from each file's Verified date.

No model, tool or Markdown rendering change.

Tests. src/app/api/chat/route.test.ts calls the real POST handler with botid, rate limiting and the SiloRail client mocked, captures the outgoing request and checks its system message for:

• a Final MIP with separately sourced activation (MIP-8 from the real bundle: its pinned spec URL, release-notes link and Verified date reach the provider);
• a Final MIP with no activation evidence (a fixture bundle), with the "cannot confirm activation" rule present;
• a custom admin prompt, with the rule and the date still appended after it.

Removing the rule from the helper fails all three. src/lib/ai/knowledge-sources.test.ts keeps every mip-*.md on the same status, verified-date and pinned-source format.

Manual answer check. I couldn't run the live model here (no SiloRail wallet on my side), so this is the check to run on a preview, not something a unit test asserts:

1. "Is MIP-8 live on mainnet?" should say yes, name MONAD_TEN and September 2, 2026, link the v0.16.1 release notes, and say "as verified on October 4, 2026" or similar.
2. "Is MIP-7 active?" should say it is a Draft and the bundle cannot confirm activation, and link the MIP-7 spec.
3. With a custom system prompt saved in /admin/chat, question 2 should still get the same qualification.

scripts/check-mip-upstream.mjs currently reports every MIP as "Review needed" because the upstream specs moved past the reviewed commit. That is the same on master and belongs to upstream sync, which #52 leaves out, so upstream-sources.json is untouched.

Verification: pnpm lint (0 errors), tsc --noEmit, vitest run (118 passed), next build.

[vercel]

@ShadowOfTime1 is attempting to deploy a commit to the MF Experimental Team on Vercel.

A member of the Team first needs to authorize it.

[portdeveloper]

lgtm, thanks for keeping the evidence rule present with a custom prompt