@vincemakes/kiso-evals 0.42.1 → 0.42.3

package/README.md

@@ -2,6 +2,6 @@
 
 The faux provider (a scripted model — zero keys, deterministic), the
 incident fixture library (every fixture is a real production accident
-from uooki 2026), and the cross-provider contract tests.
+from a production host, 2026), and the cross-provider contract tests.
 
 See the repository README for the framework overview.

package/dist/faux.d.ts

@@ -1,5 +1,5 @@
 /**
- * L7 — the faux provider: iteration's foundation (mauri ADR-0010).
+ * L7 — the faux provider: iteration's foundation (the predecessor's ADR-0010).
  *
  * A provider that needs no API key and no network. A `FauxScript` is a
  * turn-by-turn playbook: what the "model" emits. The kernel cannot tell a

package/dist/faux.js

@@ -1,5 +1,5 @@
 /**
- * L7 — the faux provider: iteration's foundation (mauri ADR-0010).
+ * L7 — the faux provider: iteration's foundation (the predecessor's ADR-0010).
  *
  * A provider that needs no API key and no network. A `FauxScript` is a
  * turn-by-turn playbook: what the "model" emits. The kernel cannot tell a

package/dist/fixtures/compaction-regrowth.d.ts

@@ -2,7 +2,7 @@
  * Fixture: COMPACTION REGROWTH — repeated compaction re-archives already
  * cleared messages, O(N²), silently overwriting original content.
  *
- * Incident (uooki production, 2026, video pipeline): measured 579 store_calls
+ * Incident (a production host, 2026, video pipeline): measured 579 store_calls
  * / 480 archive entries against a naive expectation of 44 — ~10× at 50 turns,
  * extrapolating to 64× at 200 turns and 227× at 1000. Root cause: compaction
  * re-archived messages whose content was already the clear marker, writing

package/dist/fixtures/compaction-regrowth.js

@@ -2,7 +2,7 @@
  * Fixture: COMPACTION REGROWTH — repeated compaction re-archives already
  * cleared messages, O(N²), silently overwriting original content.
  *
- * Incident (uooki production, 2026, video pipeline): measured 579 store_calls
+ * Incident (a production host, 2026, video pipeline): measured 579 store_calls
  * / 480 archive entries against a naive expectation of 44 — ~10× at 50 turns,
  * extrapolating to 64× at 200 turns and 227× at 1000. Root cause: compaction
  * re-archived messages whose content was already the clear marker, writing
@@ -21,7 +21,7 @@
 import { projectMessages } from "@vincemakes/kiso-core";
 export const compactionRegrowth = {
     name: "compaction-regrowth",
-    incident: "uooki video pipeline: 579 store_calls vs 44 expected at 50 turns — re-archiving cleared markers overwrote originals (2026)",
+    incident: "a video pipeline: 579 store_calls vs 44 expected at 50 turns — re-archiving cleared markers overwrote originals (2026)",
     script: [
         {
             events: [

package/dist/fixtures/permission-negotiation.d.ts

@@ -2,7 +2,7 @@
  * Fixture: PERMISSION NEGOTIATION — the same call shape can be denied once
  * and allowed later; the model must see both outcomes distinctly.
  *
- * Incident (mauri ADR-0002): a permission system that is a yes/no gate has
+ * Incident (the predecessor's ADR-0002): a permission system that is a yes/no gate has
  * no memory and no upgrade path. The reference implementation's ask model
  * allows deny-with-reason,
  * and the reason feeds back to the model so it can adjust. The kernel's

package/dist/fixtures/permission-negotiation.js

@@ -2,7 +2,7 @@
  * Fixture: PERMISSION NEGOTIATION — the same call shape can be denied once
  * and allowed later; the model must see both outcomes distinctly.
  *
- * Incident (mauri ADR-0002): a permission system that is a yes/no gate has
+ * Incident (the predecessor's ADR-0002): a permission system that is a yes/no gate has
  * no memory and no upgrade path. The reference implementation's ask model
  * allows deny-with-reason,
  * and the reason feeds back to the model so it can adjust. The kernel's
@@ -11,7 +11,7 @@
  */
 export const permissionNegotiation = {
     name: "permission-negotiation",
-    incident: "uooki GatePipeline was a message-prefix interceptor with no session memory — a denied call could not be re-attempted after user approval (mauri ADR-0002 audit)",
+    incident: "a gate pipeline was a message-prefix interceptor with no session memory — a denied call could not be re-attempted after user approval (the predecessor's ADR-0002 audit)",
     script: [
         {
             events: [

package/dist/fixtures/silent-tool-failure.d.ts

@@ -2,7 +2,7 @@
  * Fixture: SILENT TOOL FAILURE — an error result is swallowed and the model
  * keeps narrating success.
  *
- * Incident (uooki production, 2026): a tool call failed with is_error=true,
+ * Incident (a production host, 2026): a tool call failed with is_error=true,
  * the model ignored it and continued with "results look great, done" —
  * the turn ended "completed" with zero usable work. Root cause class:
  * error results ride the same channel as successes, and nothing forced the

package/dist/fixtures/silent-tool-failure.js

@@ -2,7 +2,7 @@
  * Fixture: SILENT TOOL FAILURE — an error result is swallowed and the model
  * keeps narrating success.
  *
- * Incident (uooki production, 2026): a tool call failed with is_error=true,
+ * Incident (a production host, 2026): a tool call failed with is_error=true,
  * the model ignored it and continued with "results look great, done" —
  * the turn ended "completed" with zero usable work. Root cause class:
  * error results ride the same channel as successes, and nothing forced the
@@ -19,7 +19,7 @@
  */
 export const silentToolFailure = {
     name: "silent-tool-failure",
-    incident: "uooki turn: search tool returned is_error, model ignored it and reported success (2026-06)",
+    incident: "a production turn: search tool returned is_error, model ignored it and reported success (2026-06)",
     script: [
         {
             events: [

package/dist/fixtures/terminal-lies.d.ts

@@ -1,7 +1,7 @@
 /**
  * Fixture: TERMINAL LIES — the model reports done work that never landed.
  *
- * Incident (uooki production, 2026): a long-document turn called
+ * Incident (a production host, 2026): a long-document turn called
  * `document_create` twice, the chat asserted "created the document artifact", and the
  * Canvas stayed empty — the tool was an inline Markdown table that never
  * reaches the canvas, and all three delivery guards were blind to it

package/dist/fixtures/terminal-lies.js

@@ -1,7 +1,7 @@
 /**
  * Fixture: TERMINAL LIES — the model reports done work that never landed.
  *
- * Incident (uooki production, 2026): a long-document turn called
+ * Incident (a production host, 2026): a long-document turn called
  * `document_create` twice, the chat asserted "created the document artifact", and the
  * Canvas stayed empty — the tool was an inline Markdown table that never
  * reaches the canvas, and all three delivery guards were blind to it
@@ -17,7 +17,7 @@
 import { analyzeDelivery } from "../governance/delivery.js";
 export const terminalLies = {
     name: "terminal-lies",
-    incident: "uooki long-document turn: model claimed 'created the document artifact', canvas stayed empty, turn reported completed (document_create trap, 2026-07-31)",
+    incident: "a long-document turn: model claimed 'created the document artifact', canvas stayed empty, turn reported completed (document_create trap, 2026-07-31)",
     delivery: {
         required: true,
         producers: new Set(["create_artifact", "docx_create"]),

package/dist/fixtures/types.d.ts

@@ -3,7 +3,7 @@
  * assertion the kernel must satisfy.
  *
  * This is kiso's answer to "prove the loop is honest": every fixture is a
- * real incident abstracted (uooki production, 2026). A fixture the loop
+ * real incident abstracted (a production host, 2026). A fixture the loop
  * cannot pass is a regression the loop must not ship. Other frameworks
  * advertise features; kiso advertises "verifyable against known failure
  * modes" — because the reference implementation's loudest production

package/dist/fixtures/types.js

@@ -3,7 +3,7 @@
  * assertion the kernel must satisfy.
  *
  * This is kiso's answer to "prove the loop is honest": every fixture is a
- * real incident abstracted (uooki production, 2026). A fixture the loop
+ * real incident abstracted (a production host, 2026). A fixture the loop
  * cannot pass is a regression the loop must not ship. Other frameworks
  * advertise features; kiso advertises "verifyable against known failure
  * modes" — because the reference implementation's loudest production

package/dist/fixtures/unknown-tool.d.ts

@@ -1,7 +1,7 @@
 /**
  * Fixture: UNKNOWN TOOL — the model calls a tool that is not registered.
  *
- * Incident class (uooki + the reference implementation's community): a tool
+ * Incident class (a production host + the reference implementation's community): a tool
  * name drifts (renamed,
  * ghost entry, typo'd enum) and the harness either silently drops the call
  * (agent loops forever trying to find the tool) or executes something

package/dist/fixtures/unknown-tool.js

@@ -1,7 +1,7 @@
 /**
  * Fixture: UNKNOWN TOOL — the model calls a tool that is not registered.
  *
- * Incident class (uooki + the reference implementation's community): a tool
+ * Incident class (a production host + the reference implementation's community): a tool
  * name drifts (renamed,
  * ghost entry, typo'd enum) and the harness either silently drops the call
  * (agent loops forever trying to find the tool) or executes something
@@ -11,7 +11,7 @@
  */
 export const unknownTool = {
     name: "unknown-tool",
-    incident: "uooki ghost-tool class: manifest listed a toolkit class name instead of the callable function name; activate_tools silently dropped it and the agent looped (2026-07)",
+    incident: "a ghost-tool class: manifest listed a toolkit class name instead of the callable function name; activate_tools silently dropped it and the agent looped (2026-07)",
     script: [
         {
             events: [

package/dist/fixtures/user-abort.d.ts

@@ -2,7 +2,7 @@
  * Fixture: USER ABORT — the user hits stop mid-run; the loop must land on an
  * honest `aborted` terminal, not a fake `completed`.
  *
- * Incident (uooki production, 2026): long artifact turns had a dead
+ * Incident (a production host, 2026): long artifact turns had a dead
  * interrupt window — a stop arriving before the executor registered its
  * steering channel returned 409 and the turn kept running to "completed".
  * The structural fix: the loop checks the signal at every phase boundary

package/dist/fixtures/user-abort.js

@@ -2,7 +2,7 @@
  * Fixture: USER ABORT — the user hits stop mid-run; the loop must land on an
  * honest `aborted` terminal, not a fake `completed`.
  *
- * Incident (uooki production, 2026): long artifact turns had a dead
+ * Incident (a production host, 2026): long artifact turns had a dead
  * interrupt window — a stop arriving before the executor registered its
  * steering channel returned 409 and the turn kept running to "completed".
  * The structural fix: the loop checks the signal at every phase boundary
@@ -31,7 +31,7 @@ export function makeAbortSignal(flipAfter) {
 }
 export const userAbort = {
     name: "user-abort",
-    incident: "uooki long artifact turn: stop during the pre-executor window returned 409, turn ran to 'completed' (2026-07-26)",
+    incident: "a long artifact turn: stop during the pre-executor window returned 409, turn ran to 'completed' (2026-07-26)",
     script: [
         {
             events: [

package/dist/governance/delivery.d.ts

@@ -1,5 +1,5 @@
 /**
- * Delivery truth — "done" means the ledger says so, not the model (uooki
+ * Delivery truth — "done" means the ledger says so, not the model (a production host
  * done_guard, 30 incidents distilled).
  *
  * EC-1 (0.13.0): this module MOVED here from packages/core/src/governance/.

package/dist/governance/delivery.js

@@ -1,5 +1,5 @@
 /**
- * Delivery truth — "done" means the ledger says so, not the model (uooki
+ * Delivery truth — "done" means the ledger says so, not the model (a production host
  * done_guard, 30 incidents distilled).
  *
  * EC-1 (0.13.0): this module MOVED here from packages/core/src/governance/.

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@vincemakes/kiso-evals",
-	"version": "0.42.1",
+	"version": "0.42.3",
 	"description": "kiso evals \u2014 the faux provider, the incident fixture library, and the cross-provider contract tests.",
 	"type": "module",
 	"license": "MIT",
@@ -21,12 +21,12 @@
 		"test": "vitest run"
 	},
 	"dependencies": {
-		"@vincemakes/kiso-core": "0.42.1"
+		"@vincemakes/kiso-core": "0.42.3"
 	},
 	"devDependencies": {
 		"@anthropic-ai/sdk": "^0.115.0",
-		"@vincemakes/kiso-provider-anthropic": "0.42.1",
-		"@vincemakes/kiso-provider-openai": "0.42.1",
+		"@vincemakes/kiso-provider-anthropic": "0.42.3",
+		"@vincemakes/kiso-provider-openai": "0.42.3",
 		"@types/node": "^26.1.2",
 		"openai": "^7.3.0",
 		"typescript": "^5.7.2",