polymath-society 0.2.5 → 0.2.7

package/dist/pipeline/coding-projects.js

@@ -205,7 +205,7 @@ var TRANSCRIPT_LINE_RE = new RegExp(`^\\s*(${NOTE})((,| and) ${NOTE})*\\s*$|^\\s
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-12.2",
+  version: "2026-07-14.2",
   // = RUBRIC_FINGERPRINT of the shipped npm rubric (packages/coding-analyzer/
   // src/grade/criteria.ts). A unit test fails loud when the rubric changes
   // without this being updated (and re-pushed to the server).
@@ -294,7 +294,7 @@ var DEFAULT_CALIBRATION = {
     }
   ],
   codingTiers: {
-    push: { median: 300, sigma: 1.121, tag: "estimated \u2014 log-normal fit", source: "Anthropic 2026 Claude Code study (400k sessions): median engaged user ~8 prompts/day at ~35 words \u2014 ~300 directed words/day; tail shape from its actions-per-prompt distribution. AVERAGE over substantial days, cumulative not spiky." },
+    push: { median: 300, sigma: 0.91, tag: "estimated \u2014 log-normal fit", source: "Anthropic 2026 Claude Code study (400k sessions): median engaged user ~8 prompts/day at ~35 words \u2014 ~300 directed words/day; tail: a 1-in-1,000 user AVERAGES ~5k directed words/day (near-daily heavy dictation, clearly below the ~10k single-day record). AVERAGE over substantial days, cumulative not spiky." },
     focus: { median: 0.08, sigma: 0.559, tag: "estimated \u2014 proxy-anchored", source: "median: Anthropic 20h/week engaged-user figure implies ~1h/day of true rapid exchange against an 8h working day; ceiling anchored to the ~4h/day deep-work limit (45% share = 1-in-1,000). AVERAGE share, cumulative not spiky." },
     orchestration: { tag: "estimated \u2014 behavior bands", source: "no published concurrency distribution exists; ceiling anchored to the Claude Code lead's documented 10-15 parallel sessions" }
   },
@@ -327,12 +327,18 @@ var DEFAULT_CALIBRATION = {
       line: "the Claude Code lead runs 10\u201315 sessions at once; his #1 tip is 3\u20135 git worktrees in parallel"
     },
     throughput: {
-      // Top-0.1% words/day reference lines, anchored to the throughput ladder
-      // (THROUGHPUT_LADDER in throughput.ts): score 10 "world-class sustained" =
-      // 9k/day, score 11 "superhuman ceiling" = 12k/day. Calibrated to the person's
-      // own anchors, not a wild guess — but still a target, not a measured population.
-      topAvgWordsPerDay: 9e3,
-      topPeakWordsPerDay: 12e3,
+      // Two DIFFERENT kinds of line (2026-07-14.2): topAvg is a PERCENTILE
+      // claim — what a 1-in-1,000 user AVERAGES over substantial days (5k/day:
+      // near-daily heavy dictation; a measured top dictating power user
+      // averages ~3.9k). topPeak is NOT a percentile — it is the CEILING
+      // reference, the wildest single days on record (~10k: a 14-16h
+      // launch-day marathon), rendered as "wildest single days", never
+      // "top X% peak". An avg-percentile and a peak-percentile can't both be
+      // shown at a same-ish value (avg≈peak reads broken — user call), and a
+      // peak PERCENTILE line is unknowable anyway; the record framing is
+      // honest and keeps the visual gap.
+      topAvgWordsPerDay: 5e3,
+      topPeakWordsPerDay: 1e4,
       benchLabel: "top 0.1%",
       // No published "words typed/day" for heavy users — the real story is OUTPUT.
       loopsHeadline: ">1,000,000 lines of Rust at 99.8% tests passing",
@@ -398,13 +404,38 @@ async function main() {
     }
     r.mins += s.activeMin || 0;
   }
-  const resolveName = (s) => {
+  const pathEvidence = async (s, desc) => {
+    try {
+      const txt = await fs.readFile(s.file, "utf8");
+      let best = null, bestN = 0;
+      for (const r of desc) {
+        const n = txt.split(r.cwd + "/").length - 1;
+        if (n > bestN) {
+          best = r;
+          bestN = n;
+        }
+      }
+      return bestN > 0 ? best : null;
+    } catch {
+      return null;
+    }
+  };
+  const titleEvidence = (s, desc) => {
+    const title = (s.title || "").toLowerCase();
+    const hits = desc.filter((r) => title.includes(base(r.cwd).replace(/[-_]/g, " ")) || title.includes(base(r.cwd)));
+    return hits.length === 1 ? hits[0] : null;
+  };
+  const resolveName = async (s) => {
     const cwd = s.project || "";
     const own = gitRemote(cwd);
     if (own) return pretty(repoName(own));
     const pref = cwd.endsWith("/") ? cwd : cwd + "/";
     const desc = repos.filter((r) => r.cwd.startsWith(pref));
     if (!desc.length) return pretty(base(cwd));
+    const byPath = await pathEvidence(s, desc);
+    if (byPath) return byPath.name;
+    const byTitle = titleEvidence(s, desc);
+    if (byTitle) return byTitle.name;
     const t = Date.parse(s.start || "") || 0;
     const dist = (r) => t < r.min ? r.min - t : t > r.max ? t - r.max : 0;
     return desc.slice().sort((a, b) => dist(a) - dist(b) || b.mins - a.mins)[0].name;
@@ -413,7 +444,7 @@ async function main() {
   for (const s of ss) {
     const d = day(s.start);
     if (!d) continue;
-    const p = resolveName(s);
+    const p = await resolveName(s);
     (dayProj[d] = dayProj[d] || {})[p] = (dayProj[d][p] || 0) + (s.activeMin || 0);
   }
   const byDay = {};

package/dist/pipeline/coding-walkthrough.js

@@ -16043,7 +16043,7 @@ var TRANSCRIPT_LINE_RE = new RegExp(`^\\s*(${NOTE})((,| and) ${NOTE})*\\s*$|^\\s
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-12.2",
+  version: "2026-07-14.2",
   // = RUBRIC_FINGERPRINT of the shipped npm rubric (packages/coding-analyzer/
   // src/grade/criteria.ts). A unit test fails loud when the rubric changes
   // without this being updated (and re-pushed to the server).
@@ -16132,7 +16132,7 @@ var DEFAULT_CALIBRATION = {
     }
   ],
   codingTiers: {
-    push: { median: 300, sigma: 1.121, tag: "estimated \u2014 log-normal fit", source: "Anthropic 2026 Claude Code study (400k sessions): median engaged user ~8 prompts/day at ~35 words \u2014 ~300 directed words/day; tail shape from its actions-per-prompt distribution. AVERAGE over substantial days, cumulative not spiky." },
+    push: { median: 300, sigma: 0.91, tag: "estimated \u2014 log-normal fit", source: "Anthropic 2026 Claude Code study (400k sessions): median engaged user ~8 prompts/day at ~35 words \u2014 ~300 directed words/day; tail: a 1-in-1,000 user AVERAGES ~5k directed words/day (near-daily heavy dictation, clearly below the ~10k single-day record). AVERAGE over substantial days, cumulative not spiky." },
     focus: { median: 0.08, sigma: 0.559, tag: "estimated \u2014 proxy-anchored", source: "median: Anthropic 20h/week engaged-user figure implies ~1h/day of true rapid exchange against an 8h working day; ceiling anchored to the ~4h/day deep-work limit (45% share = 1-in-1,000). AVERAGE share, cumulative not spiky." },
     orchestration: { tag: "estimated \u2014 behavior bands", source: "no published concurrency distribution exists; ceiling anchored to the Claude Code lead's documented 10-15 parallel sessions" }
   },
@@ -16165,12 +16165,18 @@ var DEFAULT_CALIBRATION = {
       line: "the Claude Code lead runs 10\u201315 sessions at once; his #1 tip is 3\u20135 git worktrees in parallel"
     },
     throughput: {
-      // Top-0.1% words/day reference lines, anchored to the throughput ladder
-      // (THROUGHPUT_LADDER in throughput.ts): score 10 "world-class sustained" =
-      // 9k/day, score 11 "superhuman ceiling" = 12k/day. Calibrated to the person's
-      // own anchors, not a wild guess — but still a target, not a measured population.
-      topAvgWordsPerDay: 9e3,
-      topPeakWordsPerDay: 12e3,
+      // Two DIFFERENT kinds of line (2026-07-14.2): topAvg is a PERCENTILE
+      // claim — what a 1-in-1,000 user AVERAGES over substantial days (5k/day:
+      // near-daily heavy dictation; a measured top dictating power user
+      // averages ~3.9k). topPeak is NOT a percentile — it is the CEILING
+      // reference, the wildest single days on record (~10k: a 14-16h
+      // launch-day marathon), rendered as "wildest single days", never
+      // "top X% peak". An avg-percentile and a peak-percentile can't both be
+      // shown at a same-ish value (avg≈peak reads broken — user call), and a
+      // peak PERCENTILE line is unknowable anyway; the record framing is
+      // honest and keeps the visual gap.
+      topAvgWordsPerDay: 5e3,
+      topPeakWordsPerDay: 1e4,
       benchLabel: "top 0.1%",
       // No published "words typed/day" for heavy users — the real story is OUTPUT.
       loopsHeadline: ">1,000,000 lines of Rust at 99.8% tests passing",
@@ -16217,7 +16223,7 @@ function limiter(max) {
 }
 async function main() {
   const sessions = JSON.parse(await fs6.readFile(path11.join(CODING, "sessions.json"), "utf8"));
-  const todo = sessions.filter((s) => !isTrivial(s) && !s.duplicateOf).sort((a, b) => b.humanChars - a.humanChars);
+  const todo = sessions.filter((s) => s.klass === "interactive" && !isTrivial(s) && !s.duplicateOf).sort((a, b) => b.humanChars - a.humanChars);
   await fs6.mkdir(DIR, { recursive: true });
   const pending = [];
   for (const s of todo) {