@mnemom/mnemom 0.17.14 → 0.17.15

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": "0.17.14",
+  "version": "0.17.15",
   "description": "Transparent AI agent tracing",
   "type": "module",
   "bin": {