@mnemom/mnemom 0.17.13 → 0.17.15-next.0

package/dist/commands/agent-setup.d.ts

@@ -5,11 +5,13 @@
  * If a customer names no agent, `mnemom agent` launches under a per-customer
  * governed identity `agent-<handle>`. This module makes that identity REAL in
  * their Mnemom org: it births the agent through the gateway, claims it (so the
- * customer owns it), and publishes a sane default coding posture. The CARDS are
- * OFF by default (operator 2026-10-02: alignment and protection analysis are
- * billed, so nobody pays for them until they turn them on); the goal contract is
- * what nudges (re-anchor on drift). Specifically:
- *   - alignment card: autonomy_mode + integrity_mode = off,
+ * customer owns it), and publishes a sane default coding posture. Governance that
+ * costs a model call is OFF by default (operator 2026-10-02: AIP integrity and
+ * Safe House analysis are billed, so nobody pays for them until they turn them
+ * on); governance that costs none stays ON (operator 2026-10-03: autonomy is
+ * rules-only verification + policy). The goal contract is what nudges (re-anchor
+ * on drift). Specifically:
+ *   - alignment card: autonomy_mode = observe, integrity_mode = off,
  *     principal.relationship = delegated_authority, coding bounded/forbidden
  *     actions, escalate-on-irreversible;
  *   - protection card: mode = off, thresholds 0.60/0.80/0.95 and all four
@@ -17,7 +19,8 @@
  *   - goal-alignment: nudge — sealed per-conversation via the goal contract
  *     (--goal …), graded and re-anchored by the gateway.
  * `--governance [integrity,safe-house]` opts a setup in (observe-only, billed at
- * cost on top of the agent fee); a refresh of an already-claimed agent keeps the
+ * cost on top of the agent fee); `--no-governance` turns every card off, autonomy
+ * included; a refresh of an already-claimed agent keeps the
  * modes its own cards carry unless a --governance / --no-governance flag says otherwise.
  *
  * A launch never runs on an unclaimed agent (operator 2026-10-03): an unclaimed agent lands in
@@ -36,6 +39,12 @@ export type GovernanceMode = "off" | "observe";
  * the customer pays for at cost, so both are off unless asked for.
  */
 export interface GovernanceOptIn {
+    /**
+     * Autonomy verification + policy (alignment card autonomy_mode). Rules only, no
+     * model call, so it is on by default and under every `--governance` choice; only
+     * an explicit `--no-governance` turns it off.
+     */
+    autonomy: boolean;
     /** AIP integrity analysis (alignment card integrity_mode). */
     integrity: boolean;
     /** Safe House screening (protection card mode). */
@@ -45,9 +54,10 @@ export interface GovernanceOptIn {
 export declare const GOVERNANCE_FEATURES: readonly ["integrity", "safe-house"];
 /**
  * Parse the `--governance [features]` / `--no-governance` flag. undefined = not
- * given (new agents get none; a refresh keeps the agent's current modes),
- * true = every feature, false = none, a string = a comma-separated list of
- * `integrity`, `safe-house`, `all` or `none`. Throws on an unknown name.
+ * given (new agents get autonomy only; a refresh keeps the agent's current modes),
+ * true = every feature, false (`--no-governance`) = everything off, autonomy
+ * included, a string = a comma-separated list of `integrity`, `safe-house`, `all`
+ * or `none`, each keeping autonomy on. Throws on an unknown name.
  */
 export declare function parseGovernanceFlag(value: boolean | string | undefined): GovernanceOptIn | undefined;
 /**
@@ -55,14 +65,25 @@ export declare function parseGovernanceFlag(value: boolean | string | undefined)
  * whatever an existing agent's own cards already carry (kept as-is on a refresh).
  */
 export interface CardModes {
-    /** autonomy_mode — verification + policy. No model call, so not billed; kept as-is on a refresh. */
+    /** autonomy_mode — verification + policy. No model call, so not billed: observe by default, kept as-is on a refresh. */
     autonomy: string;
     integrity: string;
     protection: string;
+    /**
+     * enforcement.allow_unmapped_tools. A new observe agent gets true: its card maps no
+     * tools to capabilities, so under the deny default every Claude Code tool is an
+     * `unmapped_denied` violation (a "fail" verdict every turn, plus an awaited
+     * grace-period read). Forbidden-tool and tool_matches escalation rules still apply.
+     * A refresh keeps the agent's own value; undefined = write none.
+     */
+    allowUnmappedTools?: boolean;
 }
 /** One phrase for the posture line: "integrity off, Safe House off". */
 export declare function describeModes(m: Pick<CardModes, "integrity" | "protection">): string;
-/** The modes for a NEW agent: off unless --governance asks for a feature. */
+/**
+ * The modes for a NEW agent: autonomy observe (rules only, no model call) unless
+ * --no-governance; integrity and Safe House off unless --governance asks for them.
+ */
 export declare function newAgentModes(governance?: GovernanceOptIn): CardModes;
 /** Printed when a setup ran without --governance: how to opt in, and what it costs. */
 export declare const GOVERNANCE_OPT_IN_HINT: string;
@@ -234,5 +255,5 @@ export declare function resolvePreflightedAccount(): Promise<{
     orgLabel: string;
     entitlement: CodeEntitlement | null;
 }>;
-export declare function buildCodeAlignmentCard(agentId: string, modes?: Pick<CardModes, "autonomy" | "integrity">): Record<string, unknown>;
+export declare function buildCodeAlignmentCard(agentId: string, modes?: Pick<CardModes, "autonomy" | "integrity" | "allowUnmappedTools">): Record<string, unknown>;
 export declare function buildCodeProtectionCard(agentId: string, mode?: string): Record<string, unknown>;

package/dist/commands/agent-setup.js

@@ -5,11 +5,13 @@
  * If a customer names no agent, `mnemom agent` launches under a per-customer
  * governed identity `agent-<handle>`. This module makes that identity REAL in
  * their Mnemom org: it births the agent through the gateway, claims it (so the
- * customer owns it), and publishes a sane default coding posture. The CARDS are
- * OFF by default (operator 2026-10-02: alignment and protection analysis are
- * billed, so nobody pays for them until they turn them on); the goal contract is
- * what nudges (re-anchor on drift). Specifically:
- *   - alignment card: autonomy_mode + integrity_mode = off,
+ * customer owns it), and publishes a sane default coding posture. Governance that
+ * costs a model call is OFF by default (operator 2026-10-02: AIP integrity and
+ * Safe House analysis are billed, so nobody pays for them until they turn them
+ * on); governance that costs none stays ON (operator 2026-10-03: autonomy is
+ * rules-only verification + policy). The goal contract is what nudges (re-anchor
+ * on drift). Specifically:
+ *   - alignment card: autonomy_mode = observe, integrity_mode = off,
  *     principal.relationship = delegated_authority, coding bounded/forbidden
  *     actions, escalate-on-irreversible;
  *   - protection card: mode = off, thresholds 0.60/0.80/0.95 and all four
@@ -17,7 +19,8 @@
  *   - goal-alignment: nudge — sealed per-conversation via the goal contract
  *     (--goal …), graded and re-anchored by the gateway.
  * `--governance [integrity,safe-house]` opts a setup in (observe-only, billed at
- * cost on top of the agent fee); a refresh of an already-claimed agent keeps the
+ * cost on top of the agent fee); `--no-governance` turns every card off, autonomy
+ * included; a refresh of an already-claimed agent keeps the
  * modes its own cards carry unless a --governance / --no-governance flag says otherwise.
  *
  * A launch never runs on an unclaimed agent (operator 2026-10-03): an unclaimed agent lands in
@@ -47,22 +50,23 @@ const say = (line = "") => console.error(line);
 export const GOVERNANCE_FEATURES = ["integrity", "safe-house"];
 /**
  * Parse the `--governance [features]` / `--no-governance` flag. undefined = not
- * given (new agents get none; a refresh keeps the agent's current modes),
- * true = every feature, false = none, a string = a comma-separated list of
- * `integrity`, `safe-house`, `all` or `none`. Throws on an unknown name.
+ * given (new agents get autonomy only; a refresh keeps the agent's current modes),
+ * true = every feature, false (`--no-governance`) = everything off, autonomy
+ * included, a string = a comma-separated list of `integrity`, `safe-house`, `all`
+ * or `none`, each keeping autonomy on. Throws on an unknown name.
  */
 export function parseGovernanceFlag(value) {
     if (value === undefined)
         return undefined;
     if (typeof value === "boolean")
-        return { integrity: value, safeHouse: value };
-    const choice = { integrity: false, safeHouse: false };
+        return { autonomy: value, integrity: value, safeHouse: value };
+    const choice = { autonomy: true, integrity: false, safeHouse: false };
     const names = value
         .split(",")
         .map((s) => s.trim().toLowerCase())
         .filter(Boolean);
     if (names.length === 0)
-        return { integrity: true, safeHouse: true };
+        return { autonomy: true, integrity: true, safeHouse: true };
     for (const name of names) {
         if (name === "all") {
             choice.integrity = true;
@@ -102,15 +106,23 @@ async function existingAgentCard(kind, agentId) {
     const sources = await getCardSources(kind, agentId);
     return sources?.agent?.card_json ?? null;
 }
-/** The modes for a NEW agent: off unless --governance asks for a feature. */
+/**
+ * The modes for a NEW agent: autonomy observe (rules only, no model call) unless
+ * --no-governance; integrity and Safe House off unless --governance asks for them.
+ */
 export function newAgentModes(governance) {
-    const integrity = modeOf(governance?.integrity ?? false);
-    return { autonomy: integrity, integrity, protection: modeOf(governance?.safeHouse ?? false) };
+    const autonomy = modeOf(governance?.autonomy ?? true);
+    return {
+        autonomy,
+        integrity: modeOf(governance?.integrity ?? false),
+        protection: modeOf(governance?.safeHouse ?? false),
+        allowUnmappedTools: autonomy === "observe" ? true : undefined,
+    };
 }
 /**
  * The modes for refreshing an EXISTING agent: an explicit --governance /
  * --no-governance wins; else it keeps the modes its own cards carry. No agent-layer
- * card → off (it never had one; new setups are off). A card with no mode field →
+ * card → the new-agent defaults (it never had one). A card with no mode field →
  * observe (the prior CLI always wrote observe). A failed read → null: skip the refresh.
  */
 async function existingAgentModes(agentId, governance) {
@@ -129,11 +141,18 @@ async function existingAgentModes(agentId, governance) {
         say(fmt.dim("  Re-run `mnemom agent setup`, or pass --governance / --no-governance to set them."));
         return null;
     }
-    const integrity = align ? (modeField(align, "integrity_mode") ?? "observe") : "off";
+    const fresh = newAgentModes();
+    const ownAllow = align?.enforcement
+        ?.allow_unmapped_tools;
     return {
-        autonomy: align ? (modeField(align, "autonomy_mode") ?? "observe") : "off",
-        integrity,
-        protection: prot ? (modeField(prot, "mode") ?? "observe") : "off",
+        autonomy: align ? (modeField(align, "autonomy_mode") ?? "observe") : fresh.autonomy,
+        integrity: align ? (modeField(align, "integrity_mode") ?? "observe") : fresh.integrity,
+        protection: prot ? (modeField(prot, "mode") ?? "observe") : fresh.protection,
+        allowUnmappedTools: align
+            ? typeof ownAllow === "boolean"
+                ? ownAllow
+                : undefined
+            : fresh.allowUnmappedTools,
     };
 }
 /** Printed when a setup ran without --governance: how to opt in, and what it costs. */
@@ -335,9 +354,9 @@ async function confirmClaimed(opts, anchorKey, fp) {
     }
     say(fmt.success(`mnemom agent: claimed '${slug}' (${agentId}) in ${opts.orgLabel}.`));
     await noteSameNamedAgent(slug, agentId);
-    // Publish the default cards (off unless --governance). A card write that can't complete does
-    // NOT undo the claim — the agent is claimed and governed on whatever landed; we warn how to
-    // finish it.
+    // Publish the default cards (autonomy observe; integrity + Safe House off unless
+    // --governance). A card write that can't complete does NOT undo the claim — the agent is
+    // claimed and governed on whatever landed; we warn how to finish it.
     if (opts.setup !== false) {
         await publishDefaultCards(agentId, newAgentModes(opts.governance));
         if (!opts.governance)
@@ -714,15 +733,21 @@ export async function resolvePreflightedAccount() {
     return { orgId: active?.org_id ?? entitlement?.org_id ?? "", orgLabel, entitlement };
 }
 // ── The default coding posture (cards off; goal contracts nudge) ─────────────
-export function buildCodeAlignmentCard(agentId, modes = { autonomy: "off", integrity: "off" }) {
+export function buildCodeAlignmentCard(agentId, modes = {
+    autonomy: "off",
+    integrity: "off",
+}) {
     return {
         card_version: "unified/2026-04-26",
         agent_id: agentId,
         // Integrity off ⇒ the gateway skips the AIP checkpoint and the observer skips
-        // its trace analysis: no model call, no charge. A new agent's autonomy
-        // (verification + policy) follows it; a refresh keeps the agent's own.
+        // its trace analysis: no model call, no charge. Autonomy (verification +
+        // policy) is rules only: observe for a new agent; a refresh keeps the agent's own.
         autonomy_mode: modes.autonomy,
         integrity_mode: modes.integrity,
+        ...(modes.allowUnmappedTools === undefined
+            ? {}
+            : { enforcement: { allow_unmapped_tools: modes.allowUnmappedTools } }),
         principal: { type: "agent", identifier: agentId, relationship: "delegated_authority" },
         values: {
             declared: ["transparency", "safety", "honesty", "user_control", "harm_prevention"],

package/dist/commands/card.js

@@ -884,8 +884,9 @@ export async function cardEditCommand(agentName, options = {}) {
         cardYaml = yaml.dump({
             card_version: "unified/2026-04-26",
             agent_id: agentId,
-            autonomy_mode: "off",
+            autonomy_mode: "observe",
             integrity_mode: "off",
+            enforcement: { allow_unmapped_tools: true },
             principal: { type: "agent", identifier: agentId, relationship: "delegated_authority" },
             values: { declared: ["transparency", "safety", "honesty"] },
             autonomy: {

package/dist/commands/onboard.js

@@ -352,8 +352,8 @@ async function earnRating(agentId, steps, say) {
 }
 // ── starter alignment card ──────────────────────────────────────────────────────
 /**
- * A starter alignment card, `off` by default (alignment analysis is billed; the
- * user turns it on) — the honest, minimal declaration
+ * A starter alignment card: autonomy observe (rules only, no model call), integrity
+ * `off` (alignment analysis is billed; the user turns it on) — the honest, minimal declaration
  * of intent for a freshly-onboarded agent. Mirrors the shape `wrap` seeds so the
  * two skill-runners produce a consistent first card.
  */
@@ -361,8 +361,10 @@ function buildStarterAlignmentCard(agentId) {
     return {
         card_version: "unified/2026-04-26",
         agent_id: agentId,
-        autonomy_mode: "off",
+        autonomy_mode: "observe",
         integrity_mode: "off",
+        // No capability mappings here, so unmapped tools must pass or every turn fails.
+        enforcement: { allow_unmapped_tools: true },
         principal: { type: "agent", identifier: agentId, relationship: "delegated_authority" },
         values: { declared: ["transparency", "safety", "honesty"] },
         autonomy: {

package/dist/commands/wrap.js

@@ -256,8 +256,10 @@ function buildStarterAlignmentCard(agentId) {
     return {
         card_version: "unified/2026-04-26",
         agent_id: agentId,
-        autonomy_mode: "off",
+        autonomy_mode: "observe",
         integrity_mode: "off",
+        // No capability mappings here, so unmapped tools must pass or every turn fails.
+        enforcement: { allow_unmapped_tools: true },
         principal: { type: "agent", identifier: agentId, relationship: "delegated_authority" },
         values: { declared: ["transparency", "safety", "honesty"] },
         autonomy: {

package/dist/index.js

@@ -173,8 +173,8 @@ const agentCmd = program
     .option("--invite <token>", "Redeem a Mnemom Agent invitation (pinv_…): unlocks Mnemom Agent for you + credits your starter µ; launches too when a session name is given")
     .option("--setup", "Re-check the served agent and refresh its default cards before launch")
     .option("--no-setup", "Skip the default-card writes (the agent is still confirmed and claimed)")
-    .option("--governance [features]", "Turn governance ON for the agent's cards: AIP integrity analysis + Safe House screening, observe-only (default: off for new agents). Pick features with a comma list: integrity, safe-house. Billed at cost on top of the agent fee")
-    .option("--no-governance", "Turn governance OFF for the agent's cards (a card refresh otherwise keeps an existing agent's modes)")
+    .option("--governance [features]", "Turn governance ON for the agent's cards: AIP integrity analysis + Safe House screening, observe-only (default: off for new agents; autonomy policy checks are rules only and on either way). Pick features with a comma list: integrity, safe-house. Billed at cost on top of the agent fee")
+    .option("--no-governance", "Turn governance OFF for the agent's cards, autonomy policy checks included (by default new agents run autonomy observe, rules only; a card refresh otherwise keeps an existing agent's modes)")
     .option("--no-yolo", "Do NOT pass --dangerously-skip-permissions (yolo is on by default)")
     .option("--no-context-hint", "Do NOT seed Claude Code's context-hint feature flag before launch")
     .addOption(new Option("--slack", "EXPERIMENTAL (`mnemom experimental enable slack`): attach the Slack channel — messages addressed to this session arrive in it, and it can reply (see `mnemom agent slack`)").hideHelp(hiddenUnless("slack")))

package/dist/lib/first-run/backtest.d.ts

@@ -6,14 +6,15 @@
  * Read-only; no network. Only token counts, model ids and timestamps are read from a
  * transcript line — never message content.
  *
- * The rate is ONE flat number (operator policy 2026-10-02): the median paired total-bill
- * saving of the S46 chain45 C2n arm (context runtime, aggressive profile, no goal) against
- * stock, 23.8% over n = 13 counted rounds (mnemom-adw scratchpad/trustcode-mvp/runs/
- * s46-status.json; early exits and void arms excluded). It applies to every request whose
- * original context is at or above the aggressive profile's fold point (its 60k tail
- * budget); smaller requests never fold, so they save 0. A flat rate over-counts mid-size
- * contexts and under-counts large ones; in aggregate that evens out. Per-size rates come
- * back once there are enough measured pairs. The token saving uses the same rule on the
+ * The rate is a median paired total-bill saving of a no-goal context-runtime arm against
+ * stock, taken from the dataset whose context sizes cover the request (operator
+ * 2026-10-05: the datasets test different context windows, so they are never pooled):
+ *   - chain45 (S46 C2n, stock peaks 212-448k): 24.3% over n = 14 counted rounds
+ *     (mnemom-adw scratchpad/trustcode-mvp/runs/s46-status.json; early exits excluded),
+ *     for requests from the aggressive fold point (60k) up to 450k;
+ *   - chain240 (d240p4 gwoff, stock peaks ~965k, Claude Code's 1M compaction): 31.6% over
+ *     n = 3 matched runs, for requests at or above 450k.
+ * Smaller requests never fold, so they save 0. The token saving uses the same rule on the
  * prompt (operator 2026-10-02: one rule for dollars and tokens).
  */
 /** One billed request, deduped by message id. `t` is epoch seconds. */
@@ -69,14 +70,18 @@ export declare function parseCompactLine(line: Buffer): Compaction | null;
  * pass (no second read of the file).
  */
 export declare function loadTranscript(path: string, seen?: Set<string>, compactions?: Compaction[]): Row[];
-/** S46 C2n vs stock, median paired total-bill saving (n = 13). */
-export declare const S46_MEDIAN_SAVING = 0.238;
+/** chain45: S46 C2n vs stock, median paired total-bill saving (n = 14). */
+export declare const S46_MEDIAN_SAVING = 0.243;
+/** chain240: d240p4 gwoff vs stock, median paired total-bill saving (n = 3). */
+export declare const D240_MEDIAN_SAVING = 0.316;
 /** The aggressive profile's tail budget (cli/src/lib/agent.ts): below it nothing folds. */
 export declare const AGGRESSIVE_FOLD_FROM = 60000;
+/** Above chain45's largest measured context, chain240 is the measurement that applies. */
+export declare const CHAIN45_MAX_CONTEXT = 450000;
 export declare const ESTIMATOR_METHOD: string;
 export type Strategy = "none" | "aggressive";
 export type RateNote = "below_aggressive" | null;
-/** The per-request rate the screen uses. `raw` equals `screen`: no floor, no curve. */
+/** The per-request rate the screen uses. `raw` equals `screen`: no floor. */
 export declare function ruleRate(x: number): {
     strategy: Strategy;
     raw: number;

package/dist/lib/first-run/backtest.js

@@ -6,14 +6,15 @@
  * Read-only; no network. Only token counts, model ids and timestamps are read from a
  * transcript line — never message content.
  *
- * The rate is ONE flat number (operator policy 2026-10-02): the median paired total-bill
- * saving of the S46 chain45 C2n arm (context runtime, aggressive profile, no goal) against
- * stock, 23.8% over n = 13 counted rounds (mnemom-adw scratchpad/trustcode-mvp/runs/
- * s46-status.json; early exits and void arms excluded). It applies to every request whose
- * original context is at or above the aggressive profile's fold point (its 60k tail
- * budget); smaller requests never fold, so they save 0. A flat rate over-counts mid-size
- * contexts and under-counts large ones; in aggregate that evens out. Per-size rates come
- * back once there are enough measured pairs. The token saving uses the same rule on the
+ * The rate is a median paired total-bill saving of a no-goal context-runtime arm against
+ * stock, taken from the dataset whose context sizes cover the request (operator
+ * 2026-10-05: the datasets test different context windows, so they are never pooled):
+ *   - chain45 (S46 C2n, stock peaks 212-448k): 24.3% over n = 14 counted rounds
+ *     (mnemom-adw scratchpad/trustcode-mvp/runs/s46-status.json; early exits excluded),
+ *     for requests from the aggressive fold point (60k) up to 450k;
+ *   - chain240 (d240p4 gwoff, stock peaks ~965k, Claude Code's 1M compaction): 31.6% over
+ *     n = 3 matched runs, for requests at or above 450k.
+ * Smaller requests never fold, so they save 0. The token saving uses the same rule on the
  * prompt (operator 2026-10-02: one rule for dollars and tokens).
  */
 import { forEachLine, isoSeconds, parseJson } from "./lines.js";
@@ -159,18 +160,24 @@ export function loadTranscript(path, seen, compactions) {
     return out;
 }
 // ── the saving rate ──────────────────────────────────────────────────────────
-/** S46 C2n vs stock, median paired total-bill saving (n = 13). */
-export const S46_MEDIAN_SAVING = 0.238;
+/** chain45: S46 C2n vs stock, median paired total-bill saving (n = 14). */
+export const S46_MEDIAN_SAVING = 0.243;
+/** chain240: d240p4 gwoff vs stock, median paired total-bill saving (n = 3). */
+export const D240_MEDIAN_SAVING = 0.316;
 /** The aggressive profile's tail budget (cli/src/lib/agent.ts): below it nothing folds. */
 export const AGGRESSIVE_FOLD_FROM = 60_000;
-export const ESTIMATOR_METHOD = "flat S46 median (23.8%, C2n vs stock, n = 13) on every request whose original context is " +
-    ">= 60k (the aggressive profile's fold point), 0 below; request count and output held at stock";
-/** The per-request rate the screen uses. `raw` equals `screen`: no floor, no curve. */
+/** Above chain45's largest measured context, chain240 is the measurement that applies. */
+export const CHAIN45_MAX_CONTEXT = 450_000;
+export const ESTIMATOR_METHOD = "per request by original context: 0 below 60k (the aggressive profile's fold point); the " +
+    "chain45 median (24.3%, S46 C2n vs stock, n = 14) from 60k to 450k; the chain240 median " +
+    "(31.6%, d240p4 gwoff vs stock, n = 3) at or above 450k; request count and output held at stock";
+/** The per-request rate the screen uses. `raw` equals `screen`: no floor. */
 export function ruleRate(x) {
     if (x < AGGRESSIVE_FOLD_FROM) {
         return { strategy: "none", raw: 0, screen: 0, note: "below_aggressive" };
     }
-    return { strategy: "aggressive", raw: S46_MEDIAN_SAVING, screen: S46_MEDIAN_SAVING, note: null };
+    const r = x >= CHAIN45_MAX_CONTEXT ? D240_MEDIAN_SAVING : S46_MEDIAN_SAVING;
+    return { strategy: "aggressive", raw: r, screen: r, note: null };
 }
 /**
  * The one swap point for a session that did NOT run through Mnemom: rows → list-price

package/dist/lib/first-run/screen.d.ts

@@ -8,8 +8,8 @@
  * Sessions come in two buckets (index.ts): THROUGH Mnemom (launch ledger, a window / env
  * signal, or fold evidence), whose saving is REALIZED (replay.ts: stock Claude Code
  * replayed at its default window, can be negative), and NOT through Mnemom, whose saving
- * is the POTENTIAL one (backtest.ts: the flat S46 rule, 23.8% of every request at or
- * above 60k, for dollars and for tokens alike). Tokens and dollars are both shown
+ * is the POTENTIAL one (backtest.ts: 24.3% of every request from 60k to 450k, 31.6% at or
+ * above 450k, for dollars and for tokens alike). Tokens and dollars are both shown
  * (operator 2026-10-02). Percentages drop their decimals (rounded DOWN).
  *
  * Variations; every number has one source and no two lines make the same claim:
@@ -51,12 +51,12 @@ export declare const STRONG_MIN_USD = 5;
 export declare const STRONG_MIN_PCT = 5;
 export declare const DAYS = 30;
 /**
- * The forward line: chain45 S46 C2n (context runtime, aggressive profile) total bill vs
- * stock, -20.9% [-25.6, -15.9], Opus 5.5, n=13 (r15 readout; operator 2026-10-02: quote
- * S46, not S43), 95% CI rounded to fives, 15-25. The d240p1 ~45% figure is NOT used
- * (its re-bill undercounts output).
+ * The forward line (operator 2026-10-05: the wide range): the low end is the lowest
+ * median across the governed arms, chain45 S46 sealed (C2g) 20.5% (n=16); the high end is
+ * the highest, chain240 d240p4 sealed 35.7% (n=3). The no-goal medians, chain45 24.3% and
+ * chain240 31.6%, sit inside it. All Opus 5.5, total bill vs stock.
  */
-export declare const FORWARD_LINE = "Long sessions are where Mnemom saves the most: about 15 to 25% in our datasets.";
+export declare const FORWARD_LINE = "Long sessions are where Mnemom saves the most: from about 20% to over 35% in our datasets.";
 /** Prices, promotions and plan allocations move; the saving is in usage (operator 2026-10-01). */
 export declare const PRICE_LINE = "Prices and plan limits change. Mnemom cuts what you use, so you stay ahead.";
 /** What `mnemom agent` does today: routes through the gateway, records, keeps context lean. */

package/dist/lib/first-run/screen.js

@@ -8,8 +8,8 @@
  * Sessions come in two buckets (index.ts): THROUGH Mnemom (launch ledger, a window / env
  * signal, or fold evidence), whose saving is REALIZED (replay.ts: stock Claude Code
  * replayed at its default window, can be negative), and NOT through Mnemom, whose saving
- * is the POTENTIAL one (backtest.ts: the flat S46 rule, 23.8% of every request at or
- * above 60k, for dollars and for tokens alike). Tokens and dollars are both shown
+ * is the POTENTIAL one (backtest.ts: 24.3% of every request from 60k to 450k, 31.6% at or
+ * above 450k, for dollars and for tokens alike). Tokens and dollars are both shown
  * (operator 2026-10-02). Percentages drop their decimals (rounded DOWN).
  *
  * Variations; every number has one source and no two lines make the same claim:
@@ -52,14 +52,15 @@ export const STRONG_MIN_USD = 5;
 export const STRONG_MIN_PCT = 5;
 export const DAYS = 30;
 /**
- * The forward line: chain45 S46 C2n (context runtime, aggressive profile) total bill vs
- * stock, -20.9% [-25.6, -15.9], Opus 5.5, n=13 (r15 readout; operator 2026-10-02: quote
- * S46, not S43), 95% CI rounded to fives, 15-25. The d240p1 ~45% figure is NOT used
- * (its re-bill undercounts output).
+ * The forward line (operator 2026-10-05: the wide range): the low end is the lowest
+ * median across the governed arms, chain45 S46 sealed (C2g) 20.5% (n=16); the high end is
+ * the highest, chain240 d240p4 sealed 35.7% (n=3). The no-goal medians, chain45 24.3% and
+ * chain240 31.6%, sit inside it. All Opus 5.5, total bill vs stock.
  */
-export const FORWARD_LINE = "Long sessions are where Mnemom saves the most: about 15 to 25% in our datasets.";
-const FORWARD_SOURCE = "operator 2026-10-02: S46 C2n total bill -20.9% [-25.6, -15.9] (Opus 5.5, n=13), 95% CI " +
-    "rounded to fives, 15 to 25; measured vs stock";
+export const FORWARD_LINE = "Long sessions are where Mnemom saves the most: from about 20% to over 35% in our datasets.";
+const FORWARD_SOURCE = "operator 2026-10-05: lowest to highest governed-arm median, total bill vs stock: chain45 " +
+    "S46 C2g 20.5% (n=16) to chain240 d240p4 sealed 35.7% (n=3); no-goal medians 24.3% and " +
+    "31.6% inside";
 /** Prices, promotions and plan allocations move; the saving is in usage (operator 2026-10-01). */
 export const PRICE_LINE = "Prices and plan limits change. Mnemom cuts what you use, so you stay ahead.";
 /** What `mnemom agent` does today: routes through the gateway, records, keeps context lean. */
@@ -387,8 +388,8 @@ export function buildScreen(inp) {
         gap();
         L.push(...wrap(FORWARD_LINE));
         A("forward_range_pct", {
-            shown: "15 to 25",
-            raw: [-25.6, -20.9, -15.9],
+            shown: "20 to over 35",
+            raw: { chain45: { c2n: -24.3, c2g: -20.5 }, chain240: { gwoff: -31.6, sealed: -35.7 } },
             method: FORWARD_SOURCE,
         });
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mnemom/mnemom",
-  "version": "0.17.13",
+  "version": "0.17.15-next.0",
   "description": "Transparent AI agent tracing",
   "type": "module",
   "bin": {