diff --git a/package-lock.json b/package-lock.json
index e7bb378d..fb8bec3f 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -25,7 +25,7 @@
         "@xterm/addon-fit": "^0.11.0",
         "@xterm/addon-web-links": "^0.12.0",
         "@xterm/xterm": "^6.0.0",
-        "ai-kit": "github:bitbaum/ai-kit#v0.2.1",
+        "ai-kit": "github:bitbaum/ai-kit#v0.6.2",
         "bip-kit": "^0.1.0",
         "bs58check": "^4.0.0",
         "class-variance-authority": "^0.7.1",
@@ -69,6 +69,9 @@
         "tailwindcss": "^4",
         "tsx": "^4.23.12",
         "typescript": "^5"
+      },
+      "engines": {
+        "node": ">=20"
       }
     },
     "node_modules/@alloc/quick-lru": {
@@ -6005,14 +6008,40 @@
         "node": ">= 14"
       }
     },
-    "node_modules/ai-kit": {
-      "name": "ai-ration",
-      "version": "0.2.1",
-      "resolved": "git+ssh://git@github.com/bitbaum/ai-kit.git#07c492ce27785d66d9623acf7b30874fb908da13",
-      "integrity": "sha512-FUhbTToHPXhJxiruyhr+1oyyYAKmqQ9yMONdD2CPL/o4/svhOuOV05M789Nfe2yhLgbWTjmh5Yc8k4Tu40umGw==",
+    "node_modules/ai-forms": {
+      "version": "0.1.2",
+      "resolved": "https://registry.npmjs.org/ai-forms/-/ai-forms-0.1.2.tgz",
+      "integrity": "sha512-agadca4pN0iGlZTlNy1Vo2SmnQB+0dNkrQSDE3wCJYeaVxKDs5KaxiwGj5+GL6mVYcI1xZ8pwnT9kL9WcxnBHA==",
       "license": "MIT",
       "engines": {
         "node": ">=18"
+      },
+      "peerDependencies": {
+        "react": ">=18"
+      },
+      "peerDependenciesMeta": {
+        "react": {
+          "optional": true
+        }
+      }
+    },
+    "node_modules/ai-kit": {
+      "version": "0.6.2",
+      "resolved": "git+ssh://git@github.com/bitbaum/ai-kit.git#ace11f14d817079ae2bf4cdc010b6daa9faacbec",
+      "license": "MIT",
+      "dependencies": {
+        "ai-forms": "^0.1.2"
+      },
+      "engines": {
+        "node": ">=20"
+      },
+      "peerDependencies": {
+        "react": ">=18"
+      },
+      "peerDependenciesMeta": {
+        "react": {
+          "optional": true
+        }
       }
     },
     "node_modules/ajv": {
diff --git a/package.json b/package.json
index 77cdb54f..251f1376 100644
--- a/package.json
+++ b/package.json
@@ -93,7 +93,6 @@
     "git:prune": "bash scripts/git/prune-merged.sh",
     "ship": "git push origin main",
     "prepare": "husky",
-    "sync:agent-core": "tsx scripts/sync-agent-core.ts",
     "probe:models": "tsx scripts/probe-models.ts",
     "check:models": "tsx scripts/check-model-ids.ts",
     "check:telemetry": "tsx scripts/check-telemetry.ts",
@@ -118,7 +117,7 @@
     "@xterm/addon-fit": "^0.11.0",
     "@xterm/addon-web-links": "^0.12.0",
     "@xterm/xterm": "^6.0.0",
-    "ai-kit": "github:bitbaum/ai-kit#v0.2.1",
+    "ai-kit": "github:bitbaum/ai-kit#v0.6.2",
     "bip-kit": "^0.1.0",
     "bs58check": "^4.0.0",
     "class-variance-authority": "^0.7.1",
diff --git a/scripts/sync-agent-core.ts b/scripts/sync-agent-core.ts
deleted file mode 100644
index fa70ec76..00000000
--- a/scripts/sync-agent-core.ts
+++ /dev/null
@@ -1,41 +0,0 @@
-/**
- * Push the canonical agent/core into OrangeCat's mirror.
- * Run: npx tsx scripts/sync-agent-core.ts   (npm run sync:agent-core)
- *
- * FleetCrown owns the canonical copy; OrangeCat mirrors it byte-for-byte. See
- * src/lib/agent/core/README.md for why the harness is duplicated rather than
- * packaged, and scripts/test/agent-core-drift.ts for the check that makes the
- * duplication safe.
- *
- * Deliberately a no-op when the sibling repo is absent — CI clones one repo at
- * a time, and a sync script that fails there would block every unrelated build.
- */
-import { readdirSync, readFileSync, writeFileSync, existsSync, mkdirSync } from "node:fs";
-import { join, dirname } from "node:path";
-import { fileURLToPath } from "node:url";
-
-const HERE = dirname(fileURLToPath(import.meta.url));
-const SRC = join(HERE, "..", "src", "lib", "agent", "core");
-const DEST = process.env.ORANGECAT_DIR
-  ? join(process.env.ORANGECAT_DIR, "src", "services", "agent-core")
-  : join(HERE, "..", "..", "orangecat", "src", "services", "agent-core");
-
-if (!existsSync(dirname(dirname(DEST)))) {
-  console.log(`↷ OrangeCat not found at ${DEST} — skipping mirror (set ORANGECAT_DIR to override)`);
-  process.exit(0);
-}
-
-mkdirSync(DEST, { recursive: true });
-
-let changed = 0;
-for (const file of readdirSync(SRC).sort()) {
-  const body = readFileSync(join(SRC, file), "utf8");
-  const target = join(DEST, file);
-  const current = existsSync(target) ? readFileSync(target, "utf8") : null;
-  if (current === body) continue;
-  writeFileSync(target, body);
-  console.log(`  → ${file}`);
-  changed++;
-}
-
-console.log(changed === 0 ? "✓ agent-core mirror already in sync" : `✓ mirrored ${changed} file(s) to ${DEST}`);
diff --git a/scripts/test/agent-core-drift.ts b/scripts/test/agent-core-drift.ts
deleted file mode 100644
index 79f7fb24..00000000
--- a/scripts/test/agent-core-drift.ts
+++ /dev/null
@@ -1,145 +0,0 @@
-/**
- * Drift gate for the mirrored agent/core.
- * Run: npx tsx scripts/test/agent-core-drift.ts
- *
- * The harness is duplicated into OrangeCat rather than packaged (see
- * src/lib/agent/core/README.md). Duplication is only safe if divergence is
- * impossible to commit accidentally — two copies of "what counts as grounded",
- * quietly disagreeing, is a worse failure than the one the harness was built to
- * fix, because it would make the two assistants wrong in different ways.
- *
- * So: SHA-256 per file, compared against the mirror. Any difference fails.
- *
- * SKIPS (exit 0) when OrangeCat is not checked out beside this repo, because CI
- * clones one repo at a time. That means the gate is a LOCAL and pre-push
- * guarantee, not a CI one — the honest boundary, stated rather than implied.
- * The corresponding check on OrangeCat's side is what catches a mirror edited
- * in isolation.
- *
- * READS THE MIRROR FROM OrangeCat's origin/main, not from its working tree.
- * The working tree is checked out to whatever branch someone happens to be
- * working on there, so a tree-based comparison answers a question nobody
- * asked — "does my canonical match a sibling repo's in-progress feature
- * branch?" — and goes red for reasons that have nothing to do with the commit
- * being pushed. On 2026-08-25 the mirror was resynced and MERGED to OrangeCat
- * main, and this gate still blocked every FleetCrown push on the machine,
- * because the OrangeCat checkout sat on an unrelated branch cut before it.
- *
- * A gate that stays red about code that is fine is worse than no gate: the
- * only way past it is --no-verify, which disables the checks that do work.
- * origin/main is what "the mirror" actually means — the shared branch both
- * repos deploy from — and it is also what the fix (`npm run sync:agent-core`,
- * commit, merge) actually updates.
- */
-import { createHash } from "node:crypto";
-import { execFileSync } from "node:child_process";
-import { readdirSync, readFileSync, existsSync } from "node:fs";
-import { join, dirname } from "node:path";
-import { fileURLToPath } from "node:url";
-
-const HERE = dirname(fileURLToPath(import.meta.url));
-const SRC = join(HERE, "..", "..", "src", "lib", "agent", "core");
-const OC_REPO = process.env.ORANGECAT_DIR
-  ?? join(HERE, "..", "..", "..", "orangecat");
-const MIRROR_PATH = "src/services/agent-core";
-const REF = process.env.ORANGECAT_REF ?? "origin/main";
-
-
-/**
- * Did THIS branch touch the canonical agent-core files?
- *
- * Drift can arise two ways: you changed canonical and did not sync, or somebody
- * changed the mirror in the other repository. Only the first is your diff. The
- * second is ambient — it was already true before you started, and blocking your
- * push on it is how a gate teaches people to pass --no-verify.
- *
- * So: your change → fail. Somebody else's → warn and let the push through, with
- * the fix printed. `desktop-release-drift` already works this way, which is why
- * it is the one check in this suite that has never blocked an unrelated push.
- */
-function branchTouchedCanonical(): boolean {
-  try {
-    // Resolve the default branch rather than assuming "main": three repos in
-    // this fleet use master and two have no origin/HEAD set.
-    let def = "main";
-    try {
-      def = execFileSync("git", ["symbolic-ref", "--short", "refs/remotes/origin/HEAD"], {
-        encoding: "utf8", stdio: ["ignore", "pipe", "ignore"],
-      }).trim().replace(/^origin\//, "") || "main";
-    } catch { /* fall through to main */ }
-    const base = execFileSync("git", ["merge-base", "HEAD", `origin/${def}`], {
-      encoding: "utf8", stdio: ["ignore", "pipe", "ignore"],
-    }).trim();
-    const changed = execFileSync("git", ["diff", "--name-only", `${base}...HEAD`], {
-      encoding: "utf8", stdio: ["ignore", "pipe", "ignore"],
-    });
-    return changed.split("\n").some(f => f.startsWith("src/lib/agent/core/"));
-  } catch {
-    // Cannot tell — assume it is yours. A gate that cannot establish innocence
-    // should not grant it.
-    return true;
-  }
-}
-
-const sha = (s: string) => createHash("sha256").update(s).digest("hex").slice(0, 16);
-
-if (!existsSync(join(OC_REPO, ".git"))) {
-  console.log(`↷ agent-core drift: OrangeCat not checked out at ${OC_REPO} — skipped`);
-  process.exit(0);
-}
-
-const git = (...args: string[]) =>
-  execFileSync("git", ["-C", OC_REPO, ...args], { encoding: "utf8", stdio: ["ignore", "pipe", "ignore"] });
-
-// A missing ref is an infrastructure fact (shallow clone, no fetch yet), not a
-// verdict about the code — skip rather than block, same as a missing checkout.
-let mirrorFiles: string[];
-try {
-  mirrorFiles = git("ls-tree", "--name-only", `${REF}:${MIRROR_PATH}`)
-    .split("\n")
-    .filter(Boolean)
-    .sort();
-} catch {
-  console.log(`↷ agent-core drift: ${REF} unavailable in ${OC_REPO} — skipped (run \`git fetch\` there)`);
-  process.exit(0);
-}
-
-const srcFiles = readdirSync(SRC).sort();
-const problems: string[] = [];
-
-for (const f of srcFiles) {
-  if (!mirrorFiles.includes(f)) {
-    problems.push(`missing from mirror: ${f}`);
-    continue;
-  }
-  const a = sha(readFileSync(join(SRC, f), "utf8"));
-  const b = sha(git("show", `${REF}:${MIRROR_PATH}/${f}`));
-  if (a !== b) problems.push(`content differs: ${f} (canonical ${a} vs mirror ${b})`);
-}
-for (const f of mirrorFiles) {
-  if (!srcFiles.includes(f)) problems.push(`extra file in mirror (not canonical): ${f}`);
-}
-
-if (problems.length > 0) {
-  const yours = branchTouchedCanonical();
-  if (!yours) {
-    console.warn("⚠ agent-core drift detected, but this branch did not touch src/lib/agent/core/ —");
-    console.warn("  the mirror moved in OrangeCat, not here. Not blocking your push.");
-    for (const p of problems) console.warn(`    ${p}`);
-    console.warn("  Fix separately: npm run sync:agent-core, then merge that in OrangeCat.");
-    process.exit(0);
-  }
-
-  console.error("✗ agent-core drift detected:");
-  for (const p of problems) console.error(`    ${p}`);
-  console.error(
-    "\n  Fix: edit the FleetCrown copy, run `npm run sync:agent-core`, then commit\n" +
-      `  and merge that change in OrangeCat — this compares against ${REF}, not a\n` +
-      "  working tree, so an unmerged local sync will not clear it.",
-  );
-  process.exit(1);
-}
-
-console.log(
-  `✓ agent-core drift: ${srcFiles.length} file(s) identical between this repo and OrangeCat ${REF}`,
-);
diff --git a/scripts/test/agent-grounding.ts b/scripts/test/agent-grounding.ts
index 3dcbe2b8..ffd46535 100644
--- a/scripts/test/agent-grounding.ts
+++ b/scripts/test/agent-grounding.ts
@@ -29,9 +29,9 @@ import {
   renderFacts,
   unrecordedFields,
   NOT_RECORDED,
-} from "../../src/lib/agent/core/facts";
-import { buildContract, buildGroundedContext, renderDirectives, buildAssistantRules, NO_BASIS } from "../../src/lib/agent/core/contract";
-import { verifyAnswer, buildRepairPrompt } from "../../src/lib/agent/core/verify";
+} from "ai-kit/grounding";
+import { buildContract, buildGroundedContext, renderDirectives, buildAssistantRules, NO_BASIS } from "ai-kit/grounding";
+import { verifyAnswer, buildRepairPrompt } from "ai-kit/grounding";
 
 // ── The real records, exactly as FleetCrown stores them ──────────────────────
 const FACTS = assignFactIds([
diff --git a/scripts/test/agent-tool-loop.ts b/scripts/test/agent-tool-loop.ts
index 35613673..48aaa455 100644
--- a/scripts/test/agent-tool-loop.ts
+++ b/scripts/test/agent-tool-loop.ts
@@ -20,7 +20,7 @@ import { z } from "zod";
 import { parseTextToolCalls, stripToolCallLines, type ModelTurn } from "../../src/lib/agent/llm";
 import { defineTool, renderToolCatalog, toOpenAITools, type ToolRegistry } from "../../src/lib/agent/tools/registry";
 import { runLokiTurn } from "../../src/lib/agent/loop";
-import { makeFact, assignFactIds } from "../../src/lib/agent/core/facts";
+import { makeFact, assignFactIds } from "ai-kit/grounding";
 
 const NAMES = ["search_people", "list_projects", "propose_action"];
 
diff --git a/scripts/test/fact-budget.ts b/scripts/test/fact-budget.ts
index 882f37e5..70e8000c 100644
--- a/scripts/test/fact-budget.ts
+++ b/scripts/test/fact-budget.ts
@@ -15,7 +15,7 @@ import {
   estimateTokens,
   omissionNotice,
 } from "@/lib/agent/fact-budget";
-import type { Fact } from "@/lib/agent/core/facts";
+import type { Fact } from "ai-kit/grounding";
 
 function assert(condition: boolean, message: string): void {
   if (!condition) throw new Error(message);
diff --git a/src/lib/agent/brief.ts b/src/lib/agent/brief.ts
index cbd6b1c4..c8300d64 100644
--- a/src/lib/agent/brief.ts
+++ b/src/lib/agent/brief.ts
@@ -26,7 +26,7 @@
 import { getStuckGoals, getGoalsDueSoon, listUpcomingCommitments } from "@/db/queries/today";
 import { getEventsDueSoon } from "@/db/queries/events";
 import { getTodayHabits } from "@/db/queries/habits";
-import type { Directive } from "@/lib/agent/core/contract";
+import type { Directive } from "ai-kit/grounding";
 
 /** Days ahead treated as "imminent" for the day-planning brief. */
 const IMMINENT_DAYS = 3;
diff --git a/src/lib/agent/context.ts b/src/lib/agent/context.ts
index e6f15906..1f15c9c8 100644
--- a/src/lib/agent/context.ts
+++ b/src/lib/agent/context.ts
@@ -26,8 +26,8 @@
  * records answers about the wrong one. Cheap deterministic facts (people,
  * projects) are kept whole; retrieved documents are the elastic part.
  */
-import { assignFactIds, renderFacts, type Fact } from "@/lib/agent/core/facts";
-import { buildGroundedContext, type Directive } from "@/lib/agent/core/contract";
+import { assignFactIds, renderFacts, type Fact } from "ai-kit/grounding";
+import { buildGroundedContext, type Directive } from "ai-kit/grounding";
 import {
   peopleFacts,
   projectFacts,
diff --git a/src/lib/agent/core/README.md b/src/lib/agent/core/README.md
deleted file mode 100644
index 1689bd4d..00000000
--- a/src/lib/agent/core/README.md
+++ /dev/null
@@ -1,74 +0,0 @@
-# agent/core — the grounding harness
-
-Pure, dependency-free TypeScript shared by **Loki** (FleetCrown) and **Cat**
-(OrangeCat). No DB, no network, no framework, no imports outside this directory.
-That constraint is what makes it mirrorable, and it is enforced by the drift
-check — do not relax it.
-
-## What it is for
-
-Both assistants had the same class of failure: a model asked to fill a rigid
-answer format against thin context invents the missing parts, and the invention
-is indistinguishable from the truth because both arrive as confident prose.
-
-The harness makes unsupported claims **hard to express** rather than merely
-discouraged:
-
-| Module | Mechanism |
-|---|---|
-| `facts.ts` | Records with a **declared field set**. Fields with no stored value render as an explicit `<not recorded>`, so absence is a stated negative rather than silence. Each record gets a citation id. |
-| `contract.ts` | A rules block **generated from this turn's facts** — enumerating the legal citation ids and the concrete gaps — plus `Directive`, for answers the app computed in SQL and the model may only phrase. |
-| `verify.ts` | A deterministic post-generation check. Flags citations that resolve to nothing, and proper nouns / numbers / paths with no source in the records or the user's message. No extra model call. |
-
-## Why absence must be explicit
-
-Loki once reported a contact as *"Ilya Druzhnikov (UZH)"*. The stored record had
-no organisation field, and the string `UZH` appears nowhere in the operator's
-data — it is the substring inside dr**UZH**nikov, surfaced by a keyword match and
-then narrated as an affiliation.
-
-A field the model was never shown is easy to invent. A field it was shown as
-`affiliation: <not recorded>` is a specific negative it has to actively
-contradict. That is the whole design.
-
-## Why the verifier is deterministic
-
-It runs on **every** turn, including free-tier ones on small models — which is
-exactly where fabrication is most likely. A verifier that costs a frontier call
-is one that gets disabled where it matters most.
-
-It works because fabrication is overwhelmingly *nominal*: models invent
-organisations, titles, file paths, phone numbers and dates. Those are
-mechanically recognisable and, if genuine, must appear in the retrieved records.
-
-## Mirroring — read before editing
-
-This directory is **duplicated verbatim** in two repos:
-
-```
-fleetcrown/src/lib/agent/core/      ← canonical
-orangecat/src/services/agent-core/  ← mirror
-```
-
-`scripts/test/agent-core-drift.ts` in **both** repos compares SHA-256 per file
-and fails CI on any difference. So:
-
-1. Edit the FleetCrown copy.
-2. Run `npm run sync:agent-core` (FleetCrown) to push the mirror.
-3. Commit both repos.
-
-This duplication is deliberate and temporary. The two apps have incompatible
-data layers (Drizzle/Postgres vs Supabase), so a shared package was not worth
-blocking on — but two silently-diverging copies of "what counts as grounded"
-would be worse than either. The drift check buys SSOT-in-practice now; the exit
-is extraction to `@fleet/agent-core` (the `@fleet/ai-forms` pattern), after
-which both repos import instead of mirroring.
-
-## What belongs here vs in the app
-
-**Here:** anything that defines what grounding *means*.
-**In the app:** anything that knows where data lives — the adapters that map
-rows to `Fact`s (`fleetcrown/src/lib/agent/sources.ts`,
-`orangecat/src/services/cat/sources.ts`) and the SQL behind `Directive`s.
-
-If you find yourself importing a DB client here, the code belongs in an adapter.
diff --git a/src/lib/agent/core/contract.ts b/src/lib/agent/core/contract.ts
deleted file mode 100644
index b277293a..00000000
--- a/src/lib/agent/core/contract.ts
+++ /dev/null
@@ -1,176 +0,0 @@
-/**
- * The grounding contract — the rules block that ships with every turn's facts.
- * MIRRORED MODULE (see core/README.md).
- *
- * Why this is generated rather than a hand-written constant: a standing prose
- * rule ("only use provided context") is a weak signal that models trade away
- * under format pressure. The failure that motivated this harness was exactly
- * that — a prompt demanding "1 focus, 3 tasks, 1 person, under 150 words, no
- * hedging" got four confidently-formatted answers, three of them invented,
- * against a context block that already said "if a question falls outside this
- * context, say so rather than guessing".
- *
- * The lesson: the model did not disobey a rule it forgot. It obeyed the
- * STRONGER of two conflicting instructions — fill five slots — because nothing
- * made the empty slot expressible. So this block does three things a static
- * prompt cannot:
- *
- *   1. Names the exact citation handles that exist this turn, so "cite a fact"
- *      is a closed-set choice rather than free text.
- *   2. Names the exact fields that are unrecorded THIS TURN, so the prohibition
- *      is concrete ("you have no affiliation for any person here") instead of
- *      abstract.
- *   3. Supplies the escape hatch verbatim, so refusing a slot is a cheaper
- *      token path than inventing one.
- */
-import { NOT_RECORDED, unrecordedFields, type Fact } from "./facts";
-
-/** The exact string the model must emit when a slot cannot be filled. */
-export const NO_BASIS = "Not in your data.";
-
-/**
- * Build the contract for a specific fact set. Empty fact sets get the strictest
- * form — with nothing retrieved, EVERY answer must be a refusal, and saying so
- * plainly beats hoping the model notices the context block is empty.
- */
-export function buildContract(facts: Fact[], directives: Directive[] = []): string {
-  const ids = [
-    ...facts.map((f) => `[${f.id}]`),
-    ...directives.map((_, i) => `[${directiveId(i)}]`),
-  ].join(" ");
-  const gaps = unrecordedFields(facts);
-
-  const rules = [
-    "## Grounding contract — this overrides every formatting instruction below",
-    "",
-    "You are answering from a fixed set of records. They are the ONLY things you know about the operator.",
-    "",
-    facts.length === 0 && directives.length === 0
-      ? `1. NO records were retrieved for this turn. You therefore cannot answer any question about the operator's projects, people, goals, habits, commitments or events. Reply "${NO_BASIS}" and say what you would need.`
-      : `1. Every claim about the operator MUST cite a record id. Legal citations this turn, and no others: ${ids}`,
-    `2. A field shown as \`${NOT_RECORDED}\` means you DO NOT KNOW it. Never supply a value for it — not from the record's own wording, not from a name that looks like a place or an organisation, not from general knowledge about a similarly-named person. A surname is not an employer.`,
-    `3. If any part of the request has no supporting record, answer that part with exactly "${NO_BASIS}" and continue with the parts you can support. A requested format NEVER obliges you to invent an item. Returning three of five requested items, each cited, is a correct and complete answer.`,
-    "4. Do not describe a person's role, employer, seniority, or history unless a record field states it. Do not infer an organisation from a name.",
-    "5. You have not browsed the web this turn. If asked to research someone, say you cannot and report only what the records hold.",
-    "6. If you are correcting an earlier answer, the correction is subject to every rule above — cite the record, or say the record does not exist.",
-  ];
-
-  if (gaps.length > 0) {
-    rules.push(
-      "",
-      `Unrecorded in THIS turn's records — you have no value for any of these and must not state one: ${gaps.join(", ")}`,
-    );
-  }
-
-  return rules.join("\n");
-}
-
-/**
- * The subset of the contract that needs no fact ids — for an assistant whose
- * context is still prose (Cat) rather than typed records.
- *
- * Weaker than `buildContract` by construction: without ids there is nothing to
- * cite, so rule 1 cannot exist and the verifier runs in entity-attribution
- * mode. What survives is the part that stopped the worst failure — never state
- * an attribute for someone in the user's data that their record does not carry,
- * and never imply research you did not perform.
- *
- * This is a stepping stone, not the destination. It exists so a live product
- * gets the protection now, without a same-day rewrite of its whole context
- * layer; the destination is typed records here too.
- */
-export function buildAssistantRules(opts: { subjectNoun: string }): string {
-  return [
-    "## Grounding rules — these override formatting instructions",
-    "",
-    `1. Everything you state about the user's own ${opts.subjectNoun} must come from the context above. Do not add an organisation, role, employer, history, or relationship that the context does not state.`,
-    "2. Do not infer an affiliation from a name. A word inside someone's name is not their employer or their city.",
-    "3. You have not browsed the web in this turn. If asked to research a person or company, say you cannot, and report only what the context holds.",
-    `4. If part of the request has no support in the context, answer that part with exactly "${NO_BASIS}" and continue with the parts you can support. A requested format never obliges you to invent an item.`,
-    "5. General knowledge (how Bitcoin, Lightning, or a payment method works) is fine to use and is not covered by rules 1–2. The restriction is on facts about THIS user and the people and organisations in their data.",
-    "6. A correction is a claim too. If you are correcting yourself, it must be supported by the context or stated as unknown.",
-  ].join("\n");
-}
-
-/**
- * A deterministic answer computed by the app, not the model.
- *
- * Some questions are not judgment calls at all. "Which goals are stuck at 0%
- * for 30+ days", "what is due in the next 3 days", "which habit is at risk"
- * are SQL predicates with exact answers, and asking a language model to derive
- * them from injected prose is strictly worse than computing them: it can only
- * introduce error. The model's job is to PHRASE the result, not to derive it.
- *
- * `answer` is empty when the query ran and found nothing — which is itself a
- * real, citable answer ("nothing is due"), and crucially different from the
- * query never having run.
- */
-export type Directive = {
-  /** What was asked, in the app's words: "goals stuck 30+ days". */
-  question: string;
-  /** Computed result lines. Empty array = ran, found nothing. */
-  answer: string[];
-  /** How it was computed, shown to the model so it can be honest about method. */
-  method: string;
-};
-
-/**
- * Citation handle for a computed answer, parallel to a Fact's [F1].
- *
- * Directives used to be uncitable, and the contract demands a citation for
- * every claim — so a model reporting a computed result had nothing legal to
- * point at and wrote "[no record id]" into the user's answer. That is the
- * harness leaking its own plumbing onto the screen. Give computed answers real
- * ids and the sentence cites [D1] like anything else.
- */
-export function directiveId(index: number): string {
-  return `D${index + 1}`;
-}
-
-/**
- * Render computed answers. These are stated as settled, because they are: the
- * model must not re-derive, second-guess, or "improve" them, and an empty
- * result must be reported as an empty result rather than backfilled from the
- * fact set.
- */
-export function renderDirectives(directives: Directive[]): string {
-  if (directives.length === 0) return "";
-  const blocks = directives.map((d, i) => {
-    const body =
-      d.answer.length > 0
-        ? d.answer.map((a) => `     - ${a}`).join("\n")
-        : "     (none — the query ran and matched nothing)";
-    return `  [${directiveId(i)}] ${d.question}   [${d.method}]\n${body}`;
-  });
-  return [
-    "## Computed answers — already resolved, do not re-derive",
-    "These were computed directly from the database for this turn. They are exact.",
-    "Report them as given and cite their id, exactly as you would a record.",
-    "Where the result is empty, say so plainly — do not substitute a plausible item from the records.",
-    "",
-    ...blocks,
-  ].join("\n");
-}
-
-/**
- * Assemble the full grounded context: contract, computed answers, then records.
- *
- * Order is deliberate and load-bearing. The contract comes FIRST so it frames
- * everything read afterwards, and the records come LAST so they sit closest to
- * the user's question — the position small models weight most heavily.
- */
-export function buildGroundedContext(input: {
-  facts: Fact[];
-  directives?: Directive[];
-  renderedFacts: string;
-}): string {
-  return [
-    buildContract(input.facts, input.directives ?? []),
-    renderDirectives(input.directives ?? []),
-    input.facts.length > 0
-      ? ["## Records", "", input.renderedFacts].join("\n")
-      : "## Records\n\n(none retrieved)",
-  ]
-    .filter(Boolean)
-    .join("\n\n---\n\n");
-}
diff --git a/src/lib/agent/core/facts.ts b/src/lib/agent/core/facts.ts
deleted file mode 100644
index c47e883d..00000000
--- a/src/lib/agent/core/facts.ts
+++ /dev/null
@@ -1,172 +0,0 @@
-/**
- * Facts — the unit of grounded context. MIRRORED MODULE (see core/README.md).
- *
- * The problem this solves, concretely. Loki was asked who to contact and
- * answered "Ilya Druzhnikov (UZH)". The stored record is:
- *
- *     { displayName: "Ilya Druzhnikov", channels: { whatsapp: "+1650…" } }
- *
- * There is no org field, and the string "UZH" appears nowhere in the operator's
- * data — it is the substring inside dr-UZH-nikov. A keyword match produced an
- * affiliation out of a surname, and prose context gave the model no way to tell
- * that "affiliation" was a field it had never been shown.
- *
- * The fix is representational, not a prompt instruction. A Fact is a RECORD with
- * a DECLARED field set, and every declared field is rendered — including the ones
- * with no value, which render as an explicit `<not recorded>`. A model that reads
- *
- *     affiliation: <not recorded>
- *
- * is being told a specific negative, which is far harder to overwrite than the
- * silence of a field that simply wasn't mentioned. Absence becomes evidence.
- *
- * Every fact also carries a short stable id ([F3]) so the answer can cite spans
- * and `verify.ts` can check citations mechanically rather than by vibes.
- *
- * Pure: no DB, no network, no framework. Apps map their rows into Facts via
- * their own adapters (FleetCrown: src/lib/agent/sources; OrangeCat: services/cat/sources).
- */
-
-/** A field that is declared for a record kind but has no stored value. */
-export const NOT_RECORDED = "<not recorded>";
-
-/**
- * One grounded record. `fields` must contain an entry for EVERY key in the
- * kind's declared field list — `null` where nothing is stored. Builders should
- * go through `makeFact`, which enforces that against the registry.
- */
-export type Fact = {
-  /** Short citation handle, assigned by `assignFactIds` (F1, F2, …). */
-  id: string;
-  /** Record kind — must be a key of the FACT_KINDS registry. */
-  kind: string;
-  /** Human label for the record (a name, a title). Never invented. */
-  subject: string;
-  /** Declared field → stored value, or null for "nothing stored". */
-  fields: Record<string, string | null>;
-  /** Where this came from, shown to the model: "people table", "goals table". */
-  source: string;
-  /**
-   * Relevance score when the fact came from similarity search. Absent for facts
-   * fetched deterministically (a SQL filter) — those are not ranked, they are
-   * simply true, and the distinction matters to the reader.
-   */
-  similarity?: number;
-};
-
-/**
- * The declared field set per record kind — the SSOT for "what could be known
- * about this kind of thing". Adding a field here makes it render as
- * `<not recorded>` everywhere it is missing, which is the entire anti-invention
- * mechanism: the model can only ever see fields we chose to declare.
- *
- * Deliberately includes fields we do NOT store (a person's `affiliation`,
- * `role`, `employer`). That is not an oversight — those are exactly the
- * attributes models invent, so naming them and marking them unrecorded is the
- * point. Do not "clean up" this list by deleting the empty ones.
- */
-export const FACT_KINDS: Record<string, readonly string[]> = {
-  person: ["name", "affiliation", "role", "how_we_met", "last_interaction", "notes", "channels"],
-  project: ["name", "status", "stack", "description", "latest_dev_log", "repo"],
-  goal: ["title", "project", "progress", "target_date", "last_updated"],
-  habit: ["title", "frequency", "current_streak", "last_checked"],
-  commitment: ["title", "due", "counterparty", "status"],
-  event: ["name", "type", "deadline", "url", "status"],
-  // Humans the operator delegates to, and the work handed to them. Separate
-  // from `person`/`commitment` because the questions are different: a crew
-  // member is asked what they are good FOR, an assignment is asked who has it
-  // and whether they said yes.
-  crew_member: ["name", "role", "skills", "engagement", "rate", "availability", "open_assignments"],
-  assignment: ["title", "assignee", "status", "due", "fee", "why"],
-  document: ["title", "source", "excerpt"],
-  pending_action: ["title", "type", "reasoning", "proposed_on", "id"],
-};
-
-/** Field list for a kind; unknown kinds fall back to whatever the fact carries. */
-export function declaredFields(kind: string, fallback: string[] = []): readonly string[] {
-  return FACT_KINDS[kind] ?? fallback;
-}
-
-/**
- * Build a Fact with every declared field present. Values not supplied become
- * null (→ `<not recorded>`). Undeclared keys are DROPPED rather than passed
- * through: if a field is worth showing the model it is worth declaring in
- * FACT_KINDS, otherwise the registry stops describing what the model sees.
- */
-export function makeFact(input: {
-  kind: string;
-  subject: string;
-  source: string;
-  values?: Record<string, string | null | undefined>;
-  similarity?: number;
-}): Fact {
-  const keys = declaredFields(input.kind, Object.keys(input.values ?? {}));
-  const fields: Record<string, string | null> = {};
-  for (const key of keys) {
-    const raw = input.values?.[key];
-    const trimmed = typeof raw === "string" ? raw.trim() : raw;
-    fields[key] = trimmed ? String(trimmed) : null;
-  }
-  return {
-    id: "",
-    kind: input.kind,
-    subject: input.subject,
-    source: input.source,
-    fields,
-    ...(input.similarity !== undefined ? { similarity: input.similarity } : {}),
-  };
-}
-
-/** Stamp sequential citation ids. Call once, after assembling the final set. */
-export function assignFactIds(facts: Fact[]): Fact[] {
-  return facts.map((f, i) => ({ ...f, id: `F${i + 1}` }));
-}
-
-/** Every citation handle in a fact set — the only legal citations in an answer. */
-export function factIds(facts: Fact[]): Set<string> {
-  return new Set(facts.map((f) => f.id));
-}
-
-/**
- * Render facts for the model. One block per record, every declared field on its
- * own line, unrecorded fields stated explicitly.
- *
- *   [F3] person — Elena Weber SINGA Switzerland  (people table)
- *        name: Elena Weber SINGA Switzerland
- *        affiliation: <not recorded>
- *        role: <not recorded>
- *        channels: whatsapp +41774730093
- *
- * The line-per-field shape matters for small models: a flat prose blob invites
- * summarising (and summarising is where invention creeps in), whereas a field
- * list invites lookup. Observed with 8B models — the same prompt over a blob
- * hallucinates roles, over a field list it reports `<not recorded>`.
- */
-export function renderFacts(facts: Fact[]): string {
-  if (facts.length === 0) return "";
-  return facts
-    .map((f) => {
-      const head = `[${f.id}] ${f.kind} — ${f.subject}  (${f.source})`;
-      const body = Object.entries(f.fields).map(
-        ([k, v]) => `     ${k}: ${v ?? NOT_RECORDED}`,
-      );
-      return [head, ...body].join("\n");
-    })
-    .join("\n\n");
-}
-
-/**
- * Which declared fields are unrecorded across the set, as
- * `kind.field` keys. The contract block names these explicitly so the rule
- * "do not state an affiliation" is anchored to a concrete gap in THIS turn's
- * context rather than being a standing abstraction the model may ignore.
- */
-export function unrecordedFields(facts: Fact[]): string[] {
-  const gaps = new Set<string>();
-  for (const f of facts) {
-    for (const [k, v] of Object.entries(f.fields)) {
-      if (v === null) gaps.add(`${f.kind}.${k}`);
-    }
-  }
-  return [...gaps].sort();
-}
diff --git a/src/lib/agent/core/verify.ts b/src/lib/agent/core/verify.ts
deleted file mode 100644
index 182c1751..00000000
--- a/src/lib/agent/core/verify.ts
+++ /dev/null
@@ -1,339 +0,0 @@
-/**
- * Groundedness verifier — MIRRORED MODULE (see core/README.md).
- *
- * Runs on the generated answer and reports claims the fact set does not support.
- * Deliberately deterministic: no second model call, no embedding round-trip, no
- * added cost or latency. That is a requirement, not a shortcut — this must run
- * on every turn including the free-tier ones, and a verifier that costs a
- * frontier call is one that gets disabled exactly where it is needed most.
- *
- * The insight that makes a cheap check work: fabrication is overwhelmingly
- * NOMINAL. Models invent organisations, titles, people, file paths, phone
- * numbers and dates — tokens that are mechanically recognisable and that must,
- * if genuine, have appeared in the retrieved records or in what the user said.
- * Grammar and hedging are hard to check; proper nouns and digits are easy.
- *
- * Scored against the real failure this was built from, every fabricated claim
- * is caught by the proper-noun or numeric rule:
- *
- *   "Ilya Druzhnikov (UZH)"                    → UZH: novel acronym
- *   "Accelerator & Bridge Program Manager"     → novel proper-noun run
- *   "University of Liechtenstein", "START Summit" → novel proper-noun runs
- *   "/opt/fleetcrown/runner/.env"              → novel path
- *
- * while the true parts ("Elena Weber SINGA Switzerland", "+41774730093") appear
- * verbatim in the records and pass clean.
- */
-import { NOT_RECORDED, type Fact } from "./facts";
-
-export type Violation = {
-  kind: "unknown-citation" | "novel-proper-noun" | "novel-number" | "novel-path" | "uncited-claim";
-  /** The offending text. */
-  text: string;
-  /** Why it is a problem, phrased for a repair prompt the model will read. */
-  detail: string;
-};
-
-export type VerifyResult = {
-  ok: boolean;
-  violations: Violation[];
-};
-
-/**
- * Words that are capitalised for reasons other than being a proper noun, or
- * that are part of this system's own vocabulary. Kept deliberately small —
- * every entry is a hole in the check, so add only what demonstrably causes
- * false positives, never to silence a true one.
- */
-const COMMON = new Set(
-  [
-    // Sentence/structural
-    "the", "a", "an", "and", "or", "but", "if", "then", "so", "because", "not",
-    "this", "that", "these", "those", "it", "its", "your", "you", "i", "we",
-    "there", "here", "what", "which", "who", "when", "where", "why", "how",
-    "no", "yes", "none", "nothing", "today", "tomorrow", "yesterday", "now",
-    "next", "last", "first", "one", "two", "three", "primary", "focus", "task",
-    "tasks", "outreach", "note", "notes", "summary", "status", "update",
-    // Days / months — real words, never evidence of a fabricated entity
-    "monday", "tuesday", "wednesday", "thursday", "friday", "saturday", "sunday",
-    "january", "february", "march", "april", "may", "june", "july", "august",
-    "september", "october", "november", "december",
-    // This system's own nouns
-    "loki", "cat", "fleetcrown", "orangecat", "not", "recorded",
-  ].map((w) => w.toLowerCase()),
-);
-
-/** Normalise for containment tests: casefold, collapse punctuation and space. */
-function norm(s: string): string {
-  return s.toLowerCase().replace(/[^a-z0-9+]+/g, " ").replace(/\s+/g, " ").trim();
-}
-
-/**
- * Everything the model was legitimately given this turn: record values, record
- * subjects, and the user's own message (a name the user typed is fair to
- * repeat). This is the corpus a claim must be traceable to.
- */
-function buildEvidence(facts: Fact[], userMessage: string, extra: string[]): string {
-  const parts: string[] = [userMessage, ...extra];
-  for (const f of facts) {
-    parts.push(f.subject, f.kind, f.source);
-    for (const v of Object.values(f.fields)) if (v) parts.push(v);
-  }
-  return norm(parts.join(" "));
-}
-
-/**
- * Lowercase words that legitimately sit INSIDE a proper name and must not break
- * it up: "University of Zurich", "Bank für Handel", "Institute for the Study of
- * Complexity". Without these, the run splits at the connector and the check
- * only ever sees the harmless halves ("University", "Zurich") while the actual
- * fabricated entity slips through unnamed.
- */
-const NAME_CONNECTORS = new Set(["of", "the", "for", "and", "de", "der", "des", "van", "von", "du", "da", "di", "für", "el", "al"]);
-
-/**
- * Named-entity candidates: ALL-CAPS acronyms, capitalised words, and the
- * multi-word runs they form (connectors allowed strictly between two
- * capitalised tokens, never at an edge).
- *
- * Both the run AND its individual tokens are emitted, deliberately. The run
- * catches composite inventions ("University of Zurich") that no single token
- * reveals; the individual tokens catch an invented acronym sitting next to a
- * real name ("Druzhnikov UZH"), where reporting only the run would name the
- * real person in the violation and produce a repair prompt that deletes the
- * true claim along with the false one.
- *
- * Sentence-initial single words are skipped — otherwise "Rotate the key" flags
- * "Rotate". That costs a little recall at sentence starts and removes the
- * dominant source of false positives; a fabricated name at a sentence start is
- * still caught by its remaining tokens.
- */
-function properNounRuns(text: string): string[] {
-  const out: string[] = [];
-  // Strip fenced and inline code — quoted identifiers are usually the user's
-  // own or a literal under discussion, not a claim about the world.
-  const prose = text.replace(/```[\s\S]*?```/g, " ").replace(/`[^`]*`/g, " ");
-
-  for (const sentence of prose.split(/(?<=[.!?:\n])\s+/)) {
-    const tokens = sentence.match(/[A-Za-z][A-Za-z0-9&.'’-]*/g) ?? [];
-    let run: string[] = [];
-
-    const flush = () => {
-      // Trim trailing connectors so "University of" never stands as a run.
-      while (run.length > 0 && NAME_CONNECTORS.has(run[run.length - 1].toLowerCase())) run.pop();
-      if (run.length > 1) out.push(run.join(" "));
-      run = [];
-    };
-
-    tokens.forEach((tok, i) => {
-      const bare = tok.replace(/[.'’-]+$/, "");
-      const isAcronym = /^[A-Z]{2,}$/.test(bare);
-      const isCapitalised = /^[A-Z][a-z]/.test(bare);
-      const isConnector = NAME_CONNECTORS.has(bare.toLowerCase());
-
-      if (isAcronym || (isCapitalised && i > 0)) {
-        run.push(bare);
-        out.push(bare); // individually checkable
-        return;
-      }
-      // A connector only continues a run that has already started.
-      if (isConnector && run.length > 0) {
-        run.push(bare);
-        return;
-      }
-      flush();
-    });
-    flush();
-  }
-  return out;
-}
-
-/** Digit groups worth checking: phone numbers, years, percentages, counts ≥ 2 digits. */
-function numericClaims(text: string): string[] {
-  const prose = text.replace(/```[\s\S]*?```/g, " ").replace(/`[^`]*`/g, " ");
-  return (prose.match(/\+?\d[\d\s().-]{3,}\d|\b\d{2,}%?\b/g) ?? []).map((s) => s.trim());
-}
-
-/**
- * File and path references — a favourite fabrication, and an unusually
- * damaging one because naming a file implies the model READ it.
- *
- * Covers absolute paths (`/opt/fleetcrown/runner/.env`), relative paths
- * (`data/contact-resolver.json`), and bare filenames with a data/config
- * extension. The relative form matters: when challenged on the UZH claim, the
- * model "corrected" itself by asserting what `data/contact-resolver.json`
- * contained — a file it was never given. That reads as citing a source, which
- * is precisely why an unverified correction is more corrosive than the
- * original error: it spends the credibility the user was trying to restore.
- */
-function pathClaims(text: string): string[] {
-  const patterns = [
-    /(?:^|[\s("'`])(\/[A-Za-z0-9_.\-/]{4,})/g, // absolute
-    /(?:^|[\s("'`])([A-Za-z0-9_.-]+\/[A-Za-z0-9_.\-/]*[A-Za-z0-9_-]\.[a-z]{2,5})/g, // relative w/ extension
-    /(?:^|[\s("'`])([A-Za-z0-9_-]+\.(?:json|env|ya?ml|sql|toml|ini|conf|log))\b/g, // bare config filename
-  ];
-  const out = new Set<string>();
-  for (const re of patterns) {
-    for (const m of text.matchAll(re)) if (m[1]) out.add(m[1]);
-  }
-  return [...out];
-}
-
-/**
- * How strictly to treat unattested names — the one real difference between the
- * two assistants that use this harness.
- *
- * `closed-world` (Loki): the assistant's entire job is reporting the operator's
- * records, so ANY unattested proper noun is a fabrication. One operator, one
- * data set, no legitimate outside knowledge in scope.
- *
- * `entity-attribution` (Cat): the assistant also answers general questions —
- * how Lightning works, which payment rails exist in Switzerland — where naming
- * Twint or Bitcoin is correct and required. Flagging those would make Cat
- * useless. So the check narrows to what actually goes wrong: attributes
- * attached to one of the USER'S OWN records. A sentence naming a known subject
- * is checked; a sentence of general explanation is not.
- *
- * The narrower mode is genuinely weaker, and that is a real trade, not a
- * loophole: Cat can still invent a fact about the wider world. It can no longer
- * invent an employer for someone in your contacts, which is the failure that
- * actually destroys trust in a personal assistant.
- */
-export type VerifyMode = "closed-world" | "entity-attribution";
-
-/** Does this sentence talk about one of the user's own records? */
-function mentionsSubject(sentence: string, subjects: string[]): boolean {
-  const s = norm(sentence);
-  return subjects.some((sub) => {
-    const n = norm(sub);
-    return n.length > 2 && s.includes(n);
-  });
-}
-
-/**
- * Verify an answer against the facts it was supposed to come from.
- *
- * `extraEvidence` lets a caller admit sources outside the fact set — computed
- * directive output, a tool result the model legitimately saw this turn.
- * Anything not in facts, the user's message, or extraEvidence is unsupported
- * by construction.
- *
- * `subjects` (entity-attribution mode) names the user's own records, so the
- * check can tell "your contact Elena works at X" from "Lightning is instant".
- */
-export function verifyAnswer(input: {
-  answer: string;
-  facts: Fact[];
-  userMessage: string;
-  extraEvidence?: string[];
-  mode?: VerifyMode;
-  subjects?: string[];
-  /**
-   * Extra legal citation handles beyond the fact ids — the [D1] series for
-   * computed answers. Without these a model that correctly cites a computed
-   * result gets flagged for citing something "that does not exist", which
-   * would train the repair pass to delete true statements.
-   */
-  extraCitationIds?: string[];
-}): VerifyResult {
-  const { answer, facts, userMessage } = input;
-  const mode: VerifyMode = input.mode ?? "closed-world";
-  const subjects = input.subjects ?? facts.map((f) => f.subject);
-  const evidence = buildEvidence(facts, userMessage, input.extraEvidence ?? []);
-  const legalIds = new Set([
-    ...facts.map((f) => f.id.toUpperCase()),
-    ...(input.extraCitationIds ?? []).map((id) => id.toUpperCase()),
-  ]);
-  const violations: Violation[] = [];
-
-  /**
-   * In entity-attribution mode, only sentences about the user's own records are
-   * subject to the name check. Built once so the per-token loop stays cheap.
-   */
-  const attributionScope =
-    mode === "entity-attribution"
-      ? answer
-          .split(/(?<=[.!?:\n])\s+/)
-          .filter((s) => mentionsSubject(s, subjects))
-          .join(" ")
-      : answer;
-
-  // 1. Citations must resolve. A citation to a record that does not exist is
-  //    the strongest possible signal of fabrication — it invents its own proof.
-  for (const cite of answer.match(/\[[FD]\d+\]/g) ?? []) {
-    const id = cite.slice(1, -1).toUpperCase();
-    if (!legalIds.has(id)) {
-      violations.push({
-        kind: "unknown-citation",
-        text: cite,
-        detail: `${cite} is not a record in this turn's context. Cite only ids that were provided, or say there is no record.`,
-      });
-    }
-  }
-
-  // 2. Named entities must be traceable. This is the anti-"UZH" rule.
-  const seen = new Set<string>();
-  for (const run of properNounRuns(attributionScope)) {
-    const n = norm(run);
-    if (!n || seen.has(n)) continue;
-    seen.add(n);
-    // Single common words are noise; multi-word runs always checked.
-    const words = n.split(" ");
-    if (words.length === 1 && (COMMON.has(words[0]) || words[0].length < 2)) continue;
-    if (words.every((w) => COMMON.has(w))) continue;
-    if (evidence.includes(n)) continue;
-    // A multi-word run whose every word is individually attested is fine —
-    // it is a rephrasing, not a new entity.
-    if (words.length > 1 && words.every((w) => COMMON.has(w) || evidence.includes(w))) continue;
-    violations.push({
-      kind: "novel-proper-noun",
-      text: run,
-      detail: `"${run}" does not appear in any record or in the operator's message. If it is an organisation, role, or place you associated with someone, the relevant field is ${NOT_RECORDED} — remove the claim.`,
-    });
-  }
-
-  // 3. Numbers must be traceable — invented phone numbers and dates read as
-  //    authoritative precisely because they are specific.
-  for (const num of numericClaims(answer)) {
-    const n = norm(num);
-    if (!n || n.length < 2) continue;
-    if (evidence.includes(n)) continue;
-    // Compare digits-only too: "+41 77 473 00 93" vs stored "+41774730093".
-    const digits = num.replace(/\D/g, "");
-    if (digits.length >= 4 && evidence.replace(/\D/g, "").includes(digits)) continue;
-    if (digits.length < 4) continue; // small counts ("3 tasks") are rhetorical
-    violations.push({
-      kind: "novel-number",
-      text: num,
-      detail: `The number "${num}" is not in any record. Do not state contact details, dates, or metrics that were not provided.`,
-    });
-  }
-
-  // 4. Paths — "update the key in /opt/fleetcrown/runner/.env" was invented
-  //    wholesale, and its specificity is what made it convincing.
-  for (const p of pathClaims(answer)) {
-    if (evidence.includes(norm(p))) continue;
-    violations.push({
-      kind: "novel-path",
-      text: p,
-      detail: `The path "${p}" is not in any record. Do not state file locations you were not given.`,
-    });
-  }
-
-  return { ok: violations.length === 0, violations };
-}
-
-/**
- * Turn violations into a repair instruction. One cheap retry with this appended
- * fixes most turns, because the model is not being asked to know more — only to
- * delete claims it cannot support.
- */
-export function buildRepairPrompt(violations: Violation[], noBasisPhrase: string): string {
-  return [
-    "Your previous answer contained claims not supported by the records. Rewrite it.",
-    "",
-    ...violations.map((v) => `- ${v.detail}`),
-    "",
-    `Remove every unsupported claim. Where removing one empties a requested item, write "${noBasisPhrase}" for that item instead of substituting something else. Keep everything that was supported, unchanged.`,
-  ].join("\n");
-}
diff --git a/src/lib/agent/fact-budget.ts b/src/lib/agent/fact-budget.ts
index 24daa8c5..0d8aa01b 100644
--- a/src/lib/agent/fact-budget.ts
+++ b/src/lib/agent/fact-budget.ts
@@ -1,4 +1,4 @@
-import type { Fact } from "@/lib/agent/core/facts";
+import type { Fact } from "ai-kit/grounding";
 
 /**
  * Fact-budget policy for the tool loop — pure, and load-bearing.
diff --git a/src/lib/agent/loop.ts b/src/lib/agent/loop.ts
index 2c25b0da..822005c3 100644
--- a/src/lib/agent/loop.ts
+++ b/src/lib/agent/loop.ts
@@ -24,15 +24,15 @@
  * observed failure mode, not defensive padding, and the loop always degrades to
  * "answer with what you have" rather than to an error.
  */
-import { assignFactIds, renderFacts, type Fact } from "@/lib/agent/core/facts";
+import { assignFactIds, renderFacts, type Fact } from "ai-kit/grounding";
 import {
   trimFactsToBudget,
   mergeFactsWithCap,
   fitFactsToBudget,
   omissionNotice,
 } from "@/lib/agent/fact-budget";
-import { buildGroundedContext, buildContract, directiveId, NO_BASIS, type Directive } from "@/lib/agent/core/contract";
-import { verifyAnswer, buildRepairPrompt, type Violation } from "@/lib/agent/core/verify";
+import { buildGroundedContext, buildContract, directiveId, NO_BASIS, type Directive } from "ai-kit/grounding";
+import { verifyAnswer, buildRepairPrompt, type Violation } from "ai-kit/grounding";
 import { callModelWithTools, type ChatMessage, type ToolCall } from "@/lib/agent/llm";
 import { renderToolCatalog, toOpenAITools, toolNames, type ToolRegistry } from "@/lib/agent/tools/registry";
 import { APP_NAME } from "@/config/brand";
diff --git a/src/lib/agent/sources.ts b/src/lib/agent/sources.ts
index 0a0e31d4..c6a3055e 100644
--- a/src/lib/agent/sources.ts
+++ b/src/lib/agent/sources.ts
@@ -15,7 +15,7 @@ import { getPendingActions } from "@/db/queries/actions";
 import { searchKnowledge, type KnowledgeHit } from "@/db/queries/knowledge-embeddings";
 import { embeddingsEnabled } from "@/lib/rag/embeddings";
 import { cleanDescription } from "@/lib/project-display";
-import { makeFact, type Fact } from "@/lib/agent/core/facts";
+import { makeFact, type Fact } from "ai-kit/grounding";
 import { isChannelAttrKey, stripChannelPrefix } from "@/config/channels";
 import { nameCandidates } from "@/lib/people-names";
 import type { DevLogEntry, UserProject } from "@/db/schema/user-projects";
diff --git a/src/lib/agent/tools/handlers.ts b/src/lib/agent/tools/handlers.ts
index a4bb4f25..3af321e2 100644
--- a/src/lib/agent/tools/handlers.ts
+++ b/src/lib/agent/tools/handlers.ts
@@ -22,7 +22,7 @@ import { createHumanTask, listOpenHumanTasks } from "@/db/queries/human-tasks";
 import { HUMAN_TASK_STATUS_LABEL, TASK_ACTOR, formatFee } from "@/config/crew";
 import { ACTION_TYPE, type ActionType } from "@/lib/constants/statuses";
 import { askGatewayAgent, isGatewayConfigured } from "@/lib/openclaw-gateway";
-import { makeFact, type Fact } from "@/lib/agent/core/facts";
+import { makeFact, type Fact } from "ai-kit/grounding";
 import { peopleFacts, projectFacts, documentFacts, pendingApprovalFacts, dateLabel } from "@/lib/agent/sources";
 import { enrichReachPayload, reachFromPerson, resolvePersonToReach } from "@/lib/people-resolve";
 import { defineTool } from "@/lib/agent/tools/registry";
diff --git a/src/lib/agent/tools/registry.ts b/src/lib/agent/tools/registry.ts
index a19d5e46..b109ca32 100644
--- a/src/lib/agent/tools/registry.ts
+++ b/src/lib/agent/tools/registry.ts
@@ -27,7 +27,7 @@
  *              a refactor.
  */
 import { z } from "zod";
-import type { Fact } from "@/lib/agent/core/facts";
+import type { Fact } from "ai-kit/grounding";
 
 /** Read = lookup. Propose = enqueue a draft for the operator to approve. */
 export type ToolKind = "read" | "propose";
diff --git a/src/lib/loki-core.ts b/src/lib/loki-core.ts
index 3db34f8d..37664335 100644
--- a/src/lib/loki-core.ts
+++ b/src/lib/loki-core.ts
@@ -14,11 +14,11 @@ import { callGroqText, GROQ_FAST_MODEL } from "@/lib/groq";
 import { getUserPreferences } from "@/db/queries/user-preferences";
 import { buildGroundedTurn, directiveEvidence } from "@/lib/agent/context";
 import { runLokiTurn } from "@/lib/agent/loop";
-import { verifyAnswer, buildRepairPrompt, type Violation } from "@/lib/agent/core/verify";
-import { NO_BASIS } from "@/lib/agent/core/contract";
+import { verifyAnswer, buildRepairPrompt, type Violation } from "ai-kit/grounding";
+import { NO_BASIS } from "ai-kit/grounding";
 import { rateLimitMessage } from "@/lib/agent/groq-error";
 import { checkAiBudget, recordAiSpend } from "@/lib/ai-budget/gate";
-import type { Fact } from "@/lib/agent/core/facts";
+import type { Fact } from "ai-kit/grounding";
 import { APP_NAME } from "@/config/brand";
 import { HTTP_TIMEOUT_LONG_MS } from "@/lib/constants/time";