@mnemom/mnemom 0.17.6-next.1 → 0.17.6-next.2

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

@@ -6,17 +6,14 @@
  * Read-only; no network. Only token counts, model ids and timestamps are read from a
  * transcript line — never message content.
  *
- * The curve (lower 95% bootstrap bound of the size-matched per-request saving, rounded
- * DOWN to 0.1 pt) comes from the S43 / S44 chain45 and d240p1 marathon paired arms:
- * mnemom-adw scratchpad/wow-first-run/CURVE-REPORT.md, curve/curve.json. The strategy is
- * chosen PER REQUEST from that request's original context (operator policy 2026-09-30):
- *
- *   < 200k        none (no fold)  → 0 on screen (the measured balanced rate there is
- *                                   slightly negative; floored, raw kept for the audit)
- *   200k – 700k   aggressive      → the measured lower bound of its bin
- *   700k – 1M     aggressive      → the 400–700k bound, EXTRAPOLATED (no measured pair)
- *   ≥ 1M          balanced        → 0 (live sessions only; a stock history cannot get
- *                                   there because Claude Code compacts first — counted)
+ * 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.
  */
 /** One billed request, deduped by message id. `t` is epoch seconds. */
 export interface Row {
@@ -59,28 +56,14 @@ export declare function parseUsageLine(line: Buffer, subagent: boolean): {
  * session copies messages) is skipped, so each request is billed once, to its origin.
  */
 export declare function loadTranscript(path: string, seen?: Set<string>): Row[];
-/** Original-context bin edges (tokens). */
-export declare const CURVE_EDGES: readonly [0, 100000, 200000, 300000, 400000, 700000, 1000000];
-/**
- * Lower 95% bound per bin, rounded down to 0.1 pt (curve/curve.json `strategies`).
- * aggressive: no measured pair below 100k or at/above 700k.
- */
-export declare const CURVE_LOWER: Readonly<Record<"aggressive" | "balanced", Record<number, number>>>;
-export declare const NO_FOLD_BELOW = 200000;
-export declare const EXTRAPOLATE_FROM = 700000;
-export declare const BALANCED_FROM = 1000000;
-export declare const POLICY = "none <200k / aggressive 200k-1M / balanced 1M+ (1M+ applies only to live sessions)";
+/** S46 C2n vs stock, median paired total-bill saving (n = 13). */
+export declare const S46_MEDIAN_SAVING = 0.238;
+/** The aggressive profile's tail budget (cli/src/lib/agent.ts): below it nothing folds. */
+export declare const AGGRESSIVE_FOLD_FROM = 60000;
 export declare const ESTIMATOR_METHOD: string;
-export type Strategy = "none" | "aggressive" | "balanced_1M+";
-export type RateNote = "floored_at_0" | "extrapolated" | "exception_1M_plus" | null;
-/** The curve bin of an original context (0 = below 100k; the top bin is open-ended). */
-export declare function binOf(x: number): number;
-/** The screen never claims a negative saving: a negative measured rate shows as 0. */
-export declare function floorRate(raw: number): {
-    screen: number;
-    note: "floored_at_0" | null;
-};
-/** The per-request rate the screen uses (`screen`) and the honest measured one (`raw`). */
+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. */
 export declare function ruleRate(x: number): {
     strategy: Strategy;
     raw: number;
@@ -90,7 +73,7 @@ export declare function ruleRate(x: number): {
 export interface Estimate {
     actual: number;
     saved: number;
-    /** The saving with the < 200k band NOT floored (audit only; never shown). */
+    /** Same as `saved` now that no rate is floored (kept for the audit record). */
     rawSaved: number;
     peak: number;
     requests: number;

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

@@ -6,17 +6,14 @@
  * Read-only; no network. Only token counts, model ids and timestamps are read from a
  * transcript line — never message content.
  *
- * The curve (lower 95% bootstrap bound of the size-matched per-request saving, rounded
- * DOWN to 0.1 pt) comes from the S43 / S44 chain45 and d240p1 marathon paired arms:
- * mnemom-adw scratchpad/wow-first-run/CURVE-REPORT.md, curve/curve.json. The strategy is
- * chosen PER REQUEST from that request's original context (operator policy 2026-09-30):
- *
- *   < 200k        none (no fold)  → 0 on screen (the measured balanced rate there is
- *                                   slightly negative; floored, raw kept for the audit)
- *   200k – 700k   aggressive      → the measured lower bound of its bin
- *   700k – 1M     aggressive      → the 400–700k bound, EXTRAPOLATED (no measured pair)
- *   ≥ 1M          balanced        → 0 (live sessions only; a stock history cannot get
- *                                   there because Claude Code compacts first — counted)
+ * 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.
  */
 import { forEachLine, isoSeconds, parseJson } from "./lines.js";
 /**
@@ -137,51 +134,19 @@ export function loadTranscript(path, seen) {
     }
     return out;
 }
-// ── the empirical curve ──────────────────────────────────────────────────────
-/** Original-context bin edges (tokens). */
-export const CURVE_EDGES = [0, 100_000, 200_000, 300_000, 400_000, 700_000, 1_000_000];
-/**
- * Lower 95% bound per bin, rounded down to 0.1 pt (curve/curve.json `strategies`).
- * aggressive: no measured pair below 100k or at/above 700k.
- */
-export const CURVE_LOWER = {
-    // bin 1 (100-200k) is measured at -0.059 but never used: < 200k runs no fold.
-    aggressive: { 2: 0.137, 3: 0.282, 4: 0.342 },
-    balanced: { 0: -0.222, 1: -0.028, 2: 0.055, 3: 0.081, 4: 0.163, 5: 0.43 },
-};
-export const NO_FOLD_BELOW = 200_000;
-export const EXTRAPOLATE_FROM = 700_000;
-export const BALANCED_FROM = 1_000_000;
-export const POLICY = "none <200k / aggressive 200k-1M / balanced 1M+ (1M+ applies only to live sessions)";
-export const ESTIMATOR_METHOD = "empirical curve (lower 95% bound, rounded down), strategy per request by its original context, " +
-    `policy ${POLICY}: < 200k the balanced arm's measured rate floored at 0, 200k-1M aggressive by bin ` +
-    "(>= 700k: the 400-700k rate, extrapolated); request count and output held at stock";
-/** The curve bin of an original context (0 = below 100k; the top bin is open-ended). */
-export function binOf(x) {
-    let i = 0;
-    while (i < CURVE_EDGES.length - 2 && x >= CURVE_EDGES[i + 1])
-        i++;
-    return i;
-}
-/** The screen never claims a negative saving: a negative measured rate shows as 0. */
-export function floorRate(raw) {
-    return raw < 0 ? { screen: 0, note: "floored_at_0" } : { screen: raw, note: null };
-}
-/** The per-request rate the screen uses (`screen`) and the honest measured one (`raw`). */
+// ── the saving rate ──────────────────────────────────────────────────────────
+/** S46 C2n vs stock, median paired total-bill saving (n = 13). */
+export const S46_MEDIAN_SAVING = 0.238;
+/** 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. */
 export function ruleRate(x) {
-    if (x >= BALANCED_FROM) {
-        return { strategy: "balanced_1M+", raw: 0, screen: 0, note: "exception_1M_plus" };
-    }
-    if (x < NO_FOLD_BELOW) {
-        const raw = CURVE_LOWER.balanced[binOf(x)];
-        return { strategy: "none", raw, ...floorRate(raw) };
-    }
-    if (x >= EXTRAPOLATE_FROM) {
-        const r = CURVE_LOWER.aggressive[binOf(400_000)];
-        return { strategy: "aggressive", raw: r, screen: r, note: "extrapolated" };
+    if (x < AGGRESSIVE_FOLD_FROM) {
+        return { strategy: "none", raw: 0, screen: 0, note: "below_aggressive" };
     }
-    const r = CURVE_LOWER.aggressive[binOf(x)];
-    return { strategy: "aggressive", raw: r, screen: r, note: null };
+    return { strategy: "aggressive", raw: S46_MEDIAN_SAVING, screen: S46_MEDIAN_SAVING, note: null };
 }
 /** The one swap point: rows → list-price actual and the conservative saving. */
 export function estimate(rows) {

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

@@ -209,7 +209,7 @@ export const LEDGER_SIGNALS = new Set(["registry_cc_id", "registry_conv_id"]);
 /** The screen's hard budget: a regression in buildScreen's line dropping must fail loudly. */
 export function assertScreenBudget(lines) {
     if (lines.length > MAX_LINES || lines.some((l) => l.length > WIDTH)) {
-        throw new Error("first-run screen exceeded its 16 x 78 budget");
+        throw new Error(`first-run screen exceeded its ${MAX_LINES} x ${WIDTH} budget`);
     }
 }
 /** A transcript's mtime (ms), 0 when it vanished since discovery (it then sorts first). */

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

@@ -1,5 +1,5 @@
 /**
- * The first-run screen: one plain terminal screen (at most 16 lines of at most 78
+ * The first-run screen: one plain terminal screen (at most 22 lines of at most 78
  * columns, ASCII) built from the backtest numbers. Pure — no I/O — so every layout is
  * unit-tested. Every number on it is recorded in the audit with its raw value and
  * method.
@@ -25,7 +25,7 @@
  *   - No live counter is promised: `mnemom agent` has no status bar today.
  */
 export declare const WIDTH = 78;
-export declare const MAX_LINES = 16;
+export declare const MAX_LINES = 22;
 /** Less history than this (API prices, in the window) is too little to backtest. */
 export declare const MIN_HISTORY_USD = 5;
 /** Variation A iff the shown saving is at least this many dollars and this many percent. */
@@ -33,12 +33,12 @@ export declare const STRONG_MIN_USD = 5;
 export declare const STRONG_MIN_PCT = 5;
 export declare const DAYS = 30;
 /**
- * The forward line: the best finished chain45 run with its error margin (operator
- * 2026-10-01, until something better): S43 C2 total bill -19.5% [-23.9, -14.9], Opus 5.5,
- * n=23, 95% CI rounded to 15-25. The d240p1 ~45% figure is NOT used (its re-bill
- * undercounts output).
+ * 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).
  */
-export declare const FORWARD_LINE = "Long sessions are where Mnemom saves the most: about 15 to 25% in our tests.";
+export declare const FORWARD_LINE = "Long sessions are where Mnemom saves the most: about 15 to 25% 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. */
@@ -52,8 +52,8 @@ export declare function sig3(x: number): number;
  * screen. The implied saving is the shown saving, which never exceeds the raw saving.
  */
 export declare function pair(total: number, saved: number): [number, number, number];
-/** part / whole as a percentage rounded DOWN to a multiple of 5. */
-export declare function pct5(part: number, whole: number): number;
+/** part / whole as a whole-number percentage, decimals dropped (rounded DOWN). */
+export declare function pctDown(part: number, whole: number): number;
 export declare function day(t: number): string;
 export declare function usd(n: number): string;
 export declare function clean(s: string): string;

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

@@ -1,5 +1,5 @@
 /**
- * The first-run screen: one plain terminal screen (at most 16 lines of at most 78
+ * The first-run screen: one plain terminal screen (at most 22 lines of at most 78
  * columns, ASCII) built from the backtest numbers. Pure — no I/O — so every layout is
  * unit-tested. Every number on it is recorded in the audit with its raw value and
  * method.
@@ -26,7 +26,7 @@
  */
 import { ESTIMATOR_METHOD } from "./backtest.js";
 export const WIDTH = 78;
-export const MAX_LINES = 16;
+export const MAX_LINES = 22;
 /** Less history than this (API prices, in the window) is too little to backtest. */
 export const MIN_HISTORY_USD = 5;
 /** Variation A iff the shown saving is at least this many dollars and this many percent. */
@@ -34,14 +34,14 @@ export const STRONG_MIN_USD = 5;
 export const STRONG_MIN_PCT = 5;
 export const DAYS = 30;
 /**
- * The forward line: the best finished chain45 run with its error margin (operator
- * 2026-10-01, until something better): S43 C2 total bill -19.5% [-23.9, -14.9], Opus 5.5,
- * n=23, 95% CI rounded to 15-25. The d240p1 ~45% figure is NOT used (its re-bill
- * undercounts output).
+ * 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).
  */
-export const FORWARD_LINE = "Long sessions are where Mnemom saves the most: about 15 to 25% in our tests.";
-const FORWARD_SOURCE = "operator 2026-10-01: S43 C2 total bill -19.5% [-23.9, -14.9] (Opus 5.5, n=23), 95% CI " +
-    "rounded to 15 to 25; measured vs stock at the 1M default";
+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";
 /** 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. */
@@ -73,11 +73,11 @@ export function pair(total, saved) {
     const tot = sig3(total);
     return [tot, Math.min(sv, tot), Math.max(tot - sv, 0)];
 }
-/** part / whole as a percentage rounded DOWN to a multiple of 5. */
-export function pct5(part, whole) {
+/** part / whole as a whole-number percentage, decimals dropped (rounded DOWN). */
+export function pctDown(part, whole) {
     if (whole <= 0 || part <= 0)
         return 0;
-    return Math.floor((100 * part) / whole / 5 + 1e-9) * 5;
+    return Math.floor((100 * part) / whole + 1e-9);
 }
 const MONTHS = ["Jan", "Feb", "Mar", "Apr", "May", "Jun", "Jul", "Aug", "Sep", "Oct", "Nov", "Dec"];
 export function day(t) {
@@ -126,17 +126,17 @@ const SAVED_METHOD = `backtest.estimate(): ${ESTIMATOR_METHOD}; 2 sig figs DOWN`
 function subLines(sub, A) {
     if (!sub)
         return [];
-    const p = pct5(sub.saved, sub.actual);
+    const p = pctDown(sub.saved, sub.actual);
     const raw = sub.actual > 0 ? (100 * sub.saved) / sub.actual : 0;
     if (p < 5) {
-        A("sub_line_omitted", { shown: true, raw, method: "usage cut rounds DOWN below 5%" });
+        A("sub_line_omitted", { shown: true, raw, method: "usage cut below 5% (decimals dropped)" });
         return [];
     }
     A("sub_usage_pct", {
         shown: p,
         raw,
         method: "saving / API-price cost over the subscription sessions in the window (all sessions if " +
-            "none is tagged); cost-weighted, rounded DOWN to 5s",
+            "none is tagged); cost-weighted, decimals dropped",
     });
     return wrap(`On your subscription that is about ${p}% less usage: fewer limit hits, and less ` +
         "overage if you use extra usage.");
@@ -151,7 +151,7 @@ export function buildScreen(inp) {
     const actual = ses.reduce((t, r) => t + r.actual, 0);
     const saved = ses.reduce((t, r) => t + r.saved, 0);
     const [a, s, e] = pair(actual, saved);
-    const p = pct5(saved, actual);
+    const p = pctDown(saved, actual);
     const enough = actual >= MIN_HISTORY_USD;
     const variation = enough && s >= STRONG_MIN_USD && p >= STRONG_MIN_PCT ? "A" : enough ? "B" : "C";
     A("variation", { shown: variation, raw: { actual, saved }, method: "see screen.ts header" });
@@ -162,7 +162,11 @@ export function buildScreen(inp) {
     L.push("");
     if (variation === "A") {
         A("spent_usd", { shown: a, raw: actual, method: SPENT_METHOD });
-        A("saved_pct", { shown: p, raw: (100 * saved) / actual, method: "rounded DOWN to 5s" });
+        A("saved_pct", {
+            shown: p,
+            raw: (100 * saved) / actual,
+            method: "decimals dropped (rounded DOWN)",
+        });
         const viaRaw = ses.reduce((t, r) => t + (r.viaMnemom ? r.saved : 0), 0);
         const via = sig2(viaRaw);
         const rest = sig2(saved - viaRaw);
@@ -170,12 +174,14 @@ export function buildScreen(inp) {
             A("saved_usd", { shown: s, raw: saved, method: SAVED_METHOD });
             A("with_mnemom_usd", { shown: e, raw: actual - saved, method: "spent shown - saved shown" });
             L.push(...wrap(`At API prices that usage cost about ${usd(a)}. With Mnemom: about ${usd(e)}.`));
+            L.push("");
             L.push(...wrap(`That's about ${usd(s)} less (${p}%), and you don't change how you work.`));
         }
         else {
             const m = `${SAVED_METHOD}; sessions the launch ledger shows ran through Mnemom`;
             A("saved_via_mnemom_usd", { shown: via, raw: viaRaw, method: m });
             L.push(...wrap(`At API prices that usage comes to about ${usd(a)} without Mnemom.`));
+            L.push("");
             if (rest < 1) {
                 L.push(...wrap(`Mnemom saved you about ${usd(via)} of it (${p}%), and you didn't change how you work.`));
             }
@@ -202,11 +208,14 @@ export function buildScreen(inp) {
             raw: big.actual - big.saved,
             method: "its cost shown - its saving (2 sig figs DOWN)",
         });
+        L.push("");
         const where = big.label ? `${big.label} on ${day(big.end)}` : `on ${day(big.end)}`;
         L.push(...wrap(big.viaMnemom
             ? `Your biggest session, ${where}: ${usd(ba)} without Mnemom, about ${usd(be)} with it.`
             : `Your biggest session, ${where}: ${usd(ba)}, with Mnemom about ${usd(be)}.`));
-        L.push(...subLines(inp.sub, A));
+        const subL = subLines(inp.sub, A);
+        if (subL.length)
+            L.push("", ...subL);
     }
     else {
         if (variation === "B") {
@@ -219,10 +228,12 @@ export function buildScreen(inp) {
             L.push(...wrap(`Your sessions used about ${usd(a)} at API prices. They stayed small, so Mnemom ` +
                 "wouldn't have changed that much."));
         }
-        L.push(FORWARD_LINE);
+        if (variation === "B")
+            L.push("");
+        L.push(...wrap(FORWARD_LINE));
         A("forward_range_pct", {
             shown: "15 to 25",
-            raw: [-23.9, -19.5, -14.9],
+            raw: [-25.6, -20.9, -15.9],
             method: FORWARD_SOURCE,
         });
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mnemom/mnemom",
-  "version": "0.17.6-next.1",
+  "version": "0.17.6-next.2",
   "description": "Transparent AI agent tracing",
   "type": "module",
   "bin": {