polymath-society 0.2.12 → 0.2.14

package/dist/pipeline/coding-coaching.js

@@ -15672,7 +15672,7 @@ var CODING_CRITERIA = [
     label: "Abstraction \u2014 seeing the load-bearing thing, and being right",
     graded: true,
     definition: "Under genuine ambiguity (an obvious path sitting right there, no answer key): did the USER see that the obvious path was subtly wrong \u2014 or that something decisive was invisible \u2014 and do the non-obvious thing that made it right, AND were they right? The graded unit is the whole move: the CATCH (noticing the fork) -> the QUESTION (formulating it) -> the CALL (resolving it). The move surfaces three interchangeable ways, scored identically: a REFRAME (collapse a mess to its real axis), a QUESTION (pose the hard thing nobody posed), or an INSTRUMENT (build the code/chart/structure that makes the truth legible, then read it). Tag the SUBSTANCE, not the medium \u2014 a chart built (technical act) to make a human-intensity call legible is a JUDGMENT catch.",
-    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / 'don't hurt Natalie, not long-term compatible' (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
+    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / a human-stakes call where the person overrode the analytically-clean answer on knowledge only they could have about someone involved, and was right (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
     rungs: [
       { level: 2, marker: "A fork existed; defers the call to the AI, or takes the obvious branch; vague." },
       { level: 4, marker: "Notices something, but the question is vague or obvious \u2014 anyone in the seat asks it." },
@@ -16290,11 +16290,11 @@ function computeWorkstyle(inp) {
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-14.2",
+  version: "2026-07-16.2",
   // = 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).
-  rubricVersion: "f51b3a93a3",
+  rubricVersion: "7983c01d53",
   // Canonical scale = the one the product owner set for the public report
   // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
   // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
@@ -16381,6 +16381,7 @@ var DEFAULT_CALIBRATION = {
   codingTiers: {
     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." },
+    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
   },
   codingBenchmarks: {
@@ -16752,6 +16753,8 @@ async function compileCodingProfile() {
   const avgWords = aggregate?.throughput?.avgWordsPerDay ?? null;
   const pushPctl = lognPct(avgWords, tiers?.push);
   const focusPctl = lognPct(flowPctDay, tiers?.focus);
+  const workDayHrs = aggregate?.flow?.flow?.workDayHours ? Math.round(aggregate.flow.flow.workDayHours * 10) / 10 : null;
+  const hoursPctl = lognPct(workDayHrs, tiers?.hours);
   const hoursAt = (min) => Object.entries(pHist).reduce((a, [k, v]) => Number(k) >= min ? a + v : a, 0) / 60;
   const orchPctl = hoursAt(5) >= 5 || hoursAt(3) >= 10 && parallelism.queueOps >= 1e3 ? 99.5 : hoursAt(3) >= 10 ? 99 : hoursAt(2) >= 0.1 * (soloMin + parMin) / 60 ? 97 : hoursAt(2) >= 2 ? 90 : 70;
   const stackUp = [
@@ -16761,14 +16764,25 @@ async function compileCodingProfile() {
       `you peak at ~${peakWords.toLocaleString()} words of pure direction in a single day \u2014 the very top sustain ~${BEST.throughput.topAvgWordsPerDay.toLocaleString()} every day, week after week; the wildest single days on record push ~${BEST.throughput.topPeakWordsPerDay.toLocaleString()}`
     ),
     sRow(
-      "How well you use AI",
+      "Hours you put in",
+      null,
+      workDayHrs != null ? `~${workDayHrs}h of active building on a typical work day \u2014 engaged builders average ~3h, and sustaining 7h+ day after day, week after week, is top-1% territory` : "not enough dated activity yet to measure a typical work day"
+    ),
+    sRow(
+      "Your AI parallelism",
       ceilOf("delegation"),
       soloPct != null ? `~${soloPct}% of your hours you drive a single agent (${Math.round(soloMin / 60)}h solo vs ${Math.round(parMin / 60)}h with 2+ live) \u2014 the Claude Code lead keeps 10\u201315 going at once, so this is the single biggest speed-up still on the table` : "you brief the AI sharply, but mostly drive one agent at a time \u2014 the top run 10\u201315 in parallel, a several-fold speed-up"
     ),
+    // The evidence line cites the RANKING metric (flow as a share of the work
+    // day), never absolute hours — a friend with MORE flow hours over a longer
+    // day landed a WORSE percentile and the hours-based line read as a bug
+    // (2026-07-16: "he has more hours but less percentile??"). The label says
+    // "when you work" ON PURPOSE: this chip ranks quality of the hours, and
+    // the volume of hours is its own chip above.
     sRow(
-      "How well you focus",
+      "Focus quality when you work",
       focusScore,
-      flowHrsDay != null ? `only ~${flowHrsDay}h of your day is true deep flow \u2014 the best hold ~4h in one unbroken block; yours fragments across parallel sessions, so the gain is directing your attention at one thread for longer` : "your attention scatters across parallel sessions \u2014 directing it at one thread for longer is the unlock"
+      flowPctDay != null ? `~${Math.round(flowPctDay * 100)}% of your work day is true deep flow${flowHrsDay != null ? ` (~${flowHrsDay}h)` : ""} \u2014 the very best hold ~${Math.round(BEST.flow.topPctOfDay * 100)}% of theirs in one unbroken block; yours fragments across parallel sessions, so the gain is directing your attention at one thread for longer` : "your attention scatters across parallel sessions \u2014 directing it at one thread for longer is the unlock"
     ),
     // The four estimates LEAD with a concrete moment from their own sessions, then an
     // HONEST caveat — how many graded sessions back the read. No invented weakness, no
@@ -16801,12 +16815,17 @@ async function compileCodingProfile() {
       r.metric = avgWords;
       r.curve = "push";
     }
-    if (r.label === "How well you focus" && focusPctl != null) {
+    if (r.label === "Hours you put in" && hoursPctl != null) {
+      r.percentile = hoursPctl;
+      r.metric = workDayHrs;
+      r.curve = "hours";
+    }
+    if (r.label === "Focus quality when you work" && focusPctl != null) {
       r.percentile = focusPctl;
       r.metric = flowPctDay;
       r.curve = "focus";
     }
-    if (r.label === "How well you use AI") r.percentile = orchPctl;
+    if (r.label === "Your AI parallelism") r.percentile = orchPctl;
   }
   return { generatedAt: (/* @__PURE__ */ new Date()).toISOString(), gradedSessions: grades.length, agglomeration, outcomeCadence, dayDigest, dayChart, timeline, walkthroughs, rubric, wordsByDay, dayThroughput, throughputCeiling: tCeiling, stackUp, criteria, parallelism, throughput, workstyle, flow, focus, expertiseMap, frontierArsenal, frontierDetail, coaching, gap, aggregate, delegationRollup, gapDist, projects };
 }

package/dist/pipeline/coding-day-digest.js

@@ -15922,7 +15922,7 @@ var CODING_CRITERIA = [
     label: "Abstraction \u2014 seeing the load-bearing thing, and being right",
     graded: true,
     definition: "Under genuine ambiguity (an obvious path sitting right there, no answer key): did the USER see that the obvious path was subtly wrong \u2014 or that something decisive was invisible \u2014 and do the non-obvious thing that made it right, AND were they right? The graded unit is the whole move: the CATCH (noticing the fork) -> the QUESTION (formulating it) -> the CALL (resolving it). The move surfaces three interchangeable ways, scored identically: a REFRAME (collapse a mess to its real axis), a QUESTION (pose the hard thing nobody posed), or an INSTRUMENT (build the code/chart/structure that makes the truth legible, then read it). Tag the SUBSTANCE, not the medium \u2014 a chart built (technical act) to make a human-intensity call legible is a JUDGMENT catch.",
-    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / 'don't hurt Natalie, not long-term compatible' (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
+    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / a human-stakes call where the person overrode the analytically-clean answer on knowledge only they could have about someone involved, and was right (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
     rungs: [
       { level: 2, marker: "A fork existed; defers the call to the AI, or takes the obvious branch; vague." },
       { level: 4, marker: "Notices something, but the question is vague or obvious \u2014 anyone in the seat asks it." },
@@ -16033,11 +16033,11 @@ var TRANSCRIPT_LINE_RE = new RegExp(`^\\s*(${NOTE})((,| and) ${NOTE})*\\s*$|^\\s
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-14.2",
+  version: "2026-07-16.2",
   // = 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).
-  rubricVersion: "f51b3a93a3",
+  rubricVersion: "7983c01d53",
   // Canonical scale = the one the product owner set for the public report
   // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
   // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
@@ -16124,6 +16124,7 @@ var DEFAULT_CALIBRATION = {
   codingTiers: {
     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." },
+    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
   },
   codingBenchmarks: {

package/dist/pipeline/coding-delegation.js

@@ -16329,7 +16329,7 @@ var CODING_CRITERIA = [
     label: "Abstraction \u2014 seeing the load-bearing thing, and being right",
     graded: true,
     definition: "Under genuine ambiguity (an obvious path sitting right there, no answer key): did the USER see that the obvious path was subtly wrong \u2014 or that something decisive was invisible \u2014 and do the non-obvious thing that made it right, AND were they right? The graded unit is the whole move: the CATCH (noticing the fork) -> the QUESTION (formulating it) -> the CALL (resolving it). The move surfaces three interchangeable ways, scored identically: a REFRAME (collapse a mess to its real axis), a QUESTION (pose the hard thing nobody posed), or an INSTRUMENT (build the code/chart/structure that makes the truth legible, then read it). Tag the SUBSTANCE, not the medium \u2014 a chart built (technical act) to make a human-intensity call legible is a JUDGMENT catch.",
-    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / 'don't hurt Natalie, not long-term compatible' (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
+    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / a human-stakes call where the person overrode the analytically-clean answer on knowledge only they could have about someone involved, and was right (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
     rungs: [
       { level: 2, marker: "A fork existed; defers the call to the AI, or takes the obvious branch; vague." },
       { level: 4, marker: "Notices something, but the question is vague or obvious \u2014 anyone in the seat asks it." },
@@ -16440,11 +16440,11 @@ var TRANSCRIPT_LINE_RE = new RegExp(`^\\s*(${NOTE})((,| and) ${NOTE})*\\s*$|^\\s
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-14.2",
+  version: "2026-07-16.2",
   // = 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).
-  rubricVersion: "f51b3a93a3",
+  rubricVersion: "7983c01d53",
   // Canonical scale = the one the product owner set for the public report
   // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
   // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
@@ -16531,6 +16531,7 @@ var DEFAULT_CALIBRATION = {
   codingTiers: {
     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." },
+    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
   },
   codingBenchmarks: {

package/dist/pipeline/coding-expertise.js

@@ -15939,7 +15939,7 @@ var CODING_CRITERIA = [
     label: "Abstraction \u2014 seeing the load-bearing thing, and being right",
     graded: true,
     definition: "Under genuine ambiguity (an obvious path sitting right there, no answer key): did the USER see that the obvious path was subtly wrong \u2014 or that something decisive was invisible \u2014 and do the non-obvious thing that made it right, AND were they right? The graded unit is the whole move: the CATCH (noticing the fork) -> the QUESTION (formulating it) -> the CALL (resolving it). The move surfaces three interchangeable ways, scored identically: a REFRAME (collapse a mess to its real axis), a QUESTION (pose the hard thing nobody posed), or an INSTRUMENT (build the code/chart/structure that makes the truth legible, then read it). Tag the SUBSTANCE, not the medium \u2014 a chart built (technical act) to make a human-intensity call legible is a JUDGMENT catch.",
-    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / 'don't hurt Natalie, not long-term compatible' (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
+    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / a human-stakes call where the person overrode the analytically-clean answer on knowledge only they could have about someone involved, and was right (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
     rungs: [
       { level: 2, marker: "A fork existed; defers the call to the AI, or takes the obvious branch; vague." },
       { level: 4, marker: "Notices something, but the question is vague or obvious \u2014 anyone in the seat asks it." },
@@ -16050,11 +16050,11 @@ var TRANSCRIPT_LINE_RE = new RegExp(`^\\s*(${NOTE})((,| and) ${NOTE})*\\s*$|^\\s
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-14.2",
+  version: "2026-07-16.2",
   // = 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).
-  rubricVersion: "f51b3a93a3",
+  rubricVersion: "7983c01d53",
   // Canonical scale = the one the product owner set for the public report
   // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
   // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
@@ -16141,6 +16141,7 @@ var DEFAULT_CALIBRATION = {
   codingTiers: {
     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." },
+    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
   },
   codingBenchmarks: {

package/dist/pipeline/coding-flow.js

@@ -287,7 +287,7 @@ var CODING_CRITERIA = [
     label: "Abstraction \u2014 seeing the load-bearing thing, and being right",
     graded: true,
     definition: "Under genuine ambiguity (an obvious path sitting right there, no answer key): did the USER see that the obvious path was subtly wrong \u2014 or that something decisive was invisible \u2014 and do the non-obvious thing that made it right, AND were they right? The graded unit is the whole move: the CATCH (noticing the fork) -> the QUESTION (formulating it) -> the CALL (resolving it). The move surfaces three interchangeable ways, scored identically: a REFRAME (collapse a mess to its real axis), a QUESTION (pose the hard thing nobody posed), or an INSTRUMENT (build the code/chart/structure that makes the truth legible, then read it). Tag the SUBSTANCE, not the medium \u2014 a chart built (technical act) to make a human-intensity call legible is a JUDGMENT catch.",
-    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / 'don't hurt Natalie, not long-term compatible' (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
+    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / a human-stakes call where the person overrode the analytically-clean answer on knowledge only they could have about someone involved, and was right (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
     rungs: [
       { level: 2, marker: "A fork existed; defers the call to the AI, or takes the obvious branch; vague." },
       { level: 4, marker: "Notices something, but the question is vague or obvious \u2014 anyone in the seat asks it." },
@@ -394,11 +394,11 @@ var IDLE_CAP_MS = 12 * 6e4;
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-14.2",
+  version: "2026-07-16.2",
   // = 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).
-  rubricVersion: "f51b3a93a3",
+  rubricVersion: "7983c01d53",
   // Canonical scale = the one the product owner set for the public report
   // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
   // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
@@ -485,6 +485,7 @@ var DEFAULT_CALIBRATION = {
   codingTiers: {
     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." },
+    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
   },
   codingBenchmarks: {

package/dist/pipeline/coding-frontier-detail.js

@@ -15905,7 +15905,7 @@ var CODING_CRITERIA = [
     label: "Abstraction \u2014 seeing the load-bearing thing, and being right",
     graded: true,
     definition: "Under genuine ambiguity (an obvious path sitting right there, no answer key): did the USER see that the obvious path was subtly wrong \u2014 or that something decisive was invisible \u2014 and do the non-obvious thing that made it right, AND were they right? The graded unit is the whole move: the CATCH (noticing the fork) -> the QUESTION (formulating it) -> the CALL (resolving it). The move surfaces three interchangeable ways, scored identically: a REFRAME (collapse a mess to its real axis), a QUESTION (pose the hard thing nobody posed), or an INSTRUMENT (build the code/chart/structure that makes the truth legible, then read it). Tag the SUBSTANCE, not the medium \u2014 a chart built (technical act) to make a human-intensity call legible is a JUDGMENT catch.",
-    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / 'don't hurt Natalie, not long-term compatible' (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
+    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / a human-stakes call where the person overrode the analytically-clean answer on knowledge only they could have about someone involved, and was right (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
     rungs: [
       { level: 2, marker: "A fork existed; defers the call to the AI, or takes the obvious branch; vague." },
       { level: 4, marker: "Notices something, but the question is vague or obvious \u2014 anyone in the seat asks it." },
@@ -16016,11 +16016,11 @@ var TRANSCRIPT_LINE_RE = new RegExp(`^\\s*(${NOTE})((,| and) ${NOTE})*\\s*$|^\\s
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-14.2",
+  version: "2026-07-16.2",
   // = 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).
-  rubricVersion: "f51b3a93a3",
+  rubricVersion: "7983c01d53",
   // Canonical scale = the one the product owner set for the public report
   // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
   // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
@@ -16107,6 +16107,7 @@ var DEFAULT_CALIBRATION = {
   codingTiers: {
     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." },
+    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
   },
   codingBenchmarks: {

package/dist/pipeline/coding-frontier.js

@@ -47,7 +47,7 @@ var CODING_CRITERIA = [
     label: "Abstraction \u2014 seeing the load-bearing thing, and being right",
     graded: true,
     definition: "Under genuine ambiguity (an obvious path sitting right there, no answer key): did the USER see that the obvious path was subtly wrong \u2014 or that something decisive was invisible \u2014 and do the non-obvious thing that made it right, AND were they right? The graded unit is the whole move: the CATCH (noticing the fork) -> the QUESTION (formulating it) -> the CALL (resolving it). The move surfaces three interchangeable ways, scored identically: a REFRAME (collapse a mess to its real axis), a QUESTION (pose the hard thing nobody posed), or an INSTRUMENT (build the code/chart/structure that makes the truth legible, then read it). Tag the SUBSTANCE, not the medium \u2014 a chart built (technical act) to make a human-intensity call legible is a JUDGMENT catch.",
-    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / 'don't hurt Natalie, not long-term compatible' (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
+    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / a human-stakes call where the person overrode the analytically-clean answer on knowledge only they could have about someone involved, and was right (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
     rungs: [
       { level: 2, marker: "A fork existed; defers the call to the AI, or takes the obvious branch; vague." },
       { level: 4, marker: "Notices something, but the question is vague or obvious \u2014 anyone in the seat asks it." },
@@ -203,11 +203,11 @@ var TRANSCRIPT_LINE_RE = new RegExp(`^\\s*(${NOTE})((,| and) ${NOTE})*\\s*$|^\\s
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-14.2",
+  version: "2026-07-16.2",
   // = 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).
-  rubricVersion: "f51b3a93a3",
+  rubricVersion: "7983c01d53",
   // Canonical scale = the one the product owner set for the public report
   // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
   // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
@@ -294,6 +294,7 @@ var DEFAULT_CALIBRATION = {
   codingTiers: {
     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." },
+    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
   },
   codingBenchmarks: {

package/dist/pipeline/coding-gap-dist.js

@@ -287,7 +287,7 @@ var CODING_CRITERIA = [
     label: "Abstraction \u2014 seeing the load-bearing thing, and being right",
     graded: true,
     definition: "Under genuine ambiguity (an obvious path sitting right there, no answer key): did the USER see that the obvious path was subtly wrong \u2014 or that something decisive was invisible \u2014 and do the non-obvious thing that made it right, AND were they right? The graded unit is the whole move: the CATCH (noticing the fork) -> the QUESTION (formulating it) -> the CALL (resolving it). The move surfaces three interchangeable ways, scored identically: a REFRAME (collapse a mess to its real axis), a QUESTION (pose the hard thing nobody posed), or an INSTRUMENT (build the code/chart/structure that makes the truth legible, then read it). Tag the SUBSTANCE, not the medium \u2014 a chart built (technical act) to make a human-intensity call legible is a JUDGMENT catch.",
-    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / 'don't hurt Natalie, not long-term compatible' (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
+    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / a human-stakes call where the person overrode the analytically-clean answer on knowledge only they could have about someone involved, and was right (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
     rungs: [
       { level: 2, marker: "A fork existed; defers the call to the AI, or takes the obvious branch; vague." },
       { level: 4, marker: "Notices something, but the question is vague or obvious \u2014 anyone in the seat asks it." },
@@ -394,11 +394,11 @@ var IDLE_CAP_MS = 12 * 6e4;
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-14.2",
+  version: "2026-07-16.2",
   // = 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).
-  rubricVersion: "f51b3a93a3",
+  rubricVersion: "7983c01d53",
   // Canonical scale = the one the product owner set for the public report
   // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
   // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
@@ -485,6 +485,7 @@ var DEFAULT_CALIBRATION = {
   codingTiers: {
     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." },
+    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
   },
   codingBenchmarks: {

package/dist/pipeline/coding-gap.js

@@ -15673,7 +15673,7 @@ var CODING_CRITERIA = [
     label: "Abstraction \u2014 seeing the load-bearing thing, and being right",
     graded: true,
     definition: "Under genuine ambiguity (an obvious path sitting right there, no answer key): did the USER see that the obvious path was subtly wrong \u2014 or that something decisive was invisible \u2014 and do the non-obvious thing that made it right, AND were they right? The graded unit is the whole move: the CATCH (noticing the fork) -> the QUESTION (formulating it) -> the CALL (resolving it). The move surfaces three interchangeable ways, scored identically: a REFRAME (collapse a mess to its real axis), a QUESTION (pose the hard thing nobody posed), or an INSTRUMENT (build the code/chart/structure that makes the truth legible, then read it). Tag the SUBSTANCE, not the medium \u2014 a chart built (technical act) to make a human-intensity call legible is a JUDGMENT catch.",
-    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / 'don't hurt Natalie, not long-term compatible' (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
+    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / a human-stakes call where the person overrode the analytically-clean answer on knowledge only they could have about someone involved, and was right (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
     rungs: [
       { level: 2, marker: "A fork existed; defers the call to the AI, or takes the obvious branch; vague." },
       { level: 4, marker: "Notices something, but the question is vague or obvious \u2014 anyone in the seat asks it." },
@@ -16291,11 +16291,11 @@ function computeWorkstyle(inp) {
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-14.2",
+  version: "2026-07-16.2",
   // = 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).
-  rubricVersion: "f51b3a93a3",
+  rubricVersion: "7983c01d53",
   // Canonical scale = the one the product owner set for the public report
   // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
   // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
@@ -16382,6 +16382,7 @@ var DEFAULT_CALIBRATION = {
   codingTiers: {
     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." },
+    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
   },
   codingBenchmarks: {
@@ -16753,6 +16754,8 @@ async function compileCodingProfile() {
   const avgWords = aggregate?.throughput?.avgWordsPerDay ?? null;
   const pushPctl = lognPct(avgWords, tiers?.push);
   const focusPctl = lognPct(flowPctDay, tiers?.focus);
+  const workDayHrs = aggregate?.flow?.flow?.workDayHours ? Math.round(aggregate.flow.flow.workDayHours * 10) / 10 : null;
+  const hoursPctl = lognPct(workDayHrs, tiers?.hours);
   const hoursAt = (min) => Object.entries(pHist).reduce((a, [k, v]) => Number(k) >= min ? a + v : a, 0) / 60;
   const orchPctl = hoursAt(5) >= 5 || hoursAt(3) >= 10 && parallelism.queueOps >= 1e3 ? 99.5 : hoursAt(3) >= 10 ? 99 : hoursAt(2) >= 0.1 * (soloMin + parMin) / 60 ? 97 : hoursAt(2) >= 2 ? 90 : 70;
   const stackUp = [
@@ -16762,14 +16765,25 @@ async function compileCodingProfile() {
       `you peak at ~${peakWords.toLocaleString()} words of pure direction in a single day \u2014 the very top sustain ~${BEST.throughput.topAvgWordsPerDay.toLocaleString()} every day, week after week; the wildest single days on record push ~${BEST.throughput.topPeakWordsPerDay.toLocaleString()}`
     ),
     sRow(
-      "How well you use AI",
+      "Hours you put in",
+      null,
+      workDayHrs != null ? `~${workDayHrs}h of active building on a typical work day \u2014 engaged builders average ~3h, and sustaining 7h+ day after day, week after week, is top-1% territory` : "not enough dated activity yet to measure a typical work day"
+    ),
+    sRow(
+      "Your AI parallelism",
       ceilOf("delegation"),
       soloPct != null ? `~${soloPct}% of your hours you drive a single agent (${Math.round(soloMin / 60)}h solo vs ${Math.round(parMin / 60)}h with 2+ live) \u2014 the Claude Code lead keeps 10\u201315 going at once, so this is the single biggest speed-up still on the table` : "you brief the AI sharply, but mostly drive one agent at a time \u2014 the top run 10\u201315 in parallel, a several-fold speed-up"
     ),
+    // The evidence line cites the RANKING metric (flow as a share of the work
+    // day), never absolute hours — a friend with MORE flow hours over a longer
+    // day landed a WORSE percentile and the hours-based line read as a bug
+    // (2026-07-16: "he has more hours but less percentile??"). The label says
+    // "when you work" ON PURPOSE: this chip ranks quality of the hours, and
+    // the volume of hours is its own chip above.
     sRow(
-      "How well you focus",
+      "Focus quality when you work",
       focusScore,
-      flowHrsDay != null ? `only ~${flowHrsDay}h of your day is true deep flow \u2014 the best hold ~4h in one unbroken block; yours fragments across parallel sessions, so the gain is directing your attention at one thread for longer` : "your attention scatters across parallel sessions \u2014 directing it at one thread for longer is the unlock"
+      flowPctDay != null ? `~${Math.round(flowPctDay * 100)}% of your work day is true deep flow${flowHrsDay != null ? ` (~${flowHrsDay}h)` : ""} \u2014 the very best hold ~${Math.round(BEST.flow.topPctOfDay * 100)}% of theirs in one unbroken block; yours fragments across parallel sessions, so the gain is directing your attention at one thread for longer` : "your attention scatters across parallel sessions \u2014 directing it at one thread for longer is the unlock"
     ),
     // The four estimates LEAD with a concrete moment from their own sessions, then an
     // HONEST caveat — how many graded sessions back the read. No invented weakness, no
@@ -16802,12 +16816,17 @@ async function compileCodingProfile() {
       r.metric = avgWords;
       r.curve = "push";
     }
-    if (r.label === "How well you focus" && focusPctl != null) {
+    if (r.label === "Hours you put in" && hoursPctl != null) {
+      r.percentile = hoursPctl;
+      r.metric = workDayHrs;
+      r.curve = "hours";
+    }
+    if (r.label === "Focus quality when you work" && focusPctl != null) {
       r.percentile = focusPctl;
       r.metric = flowPctDay;
       r.curve = "focus";
     }
-    if (r.label === "How well you use AI") r.percentile = orchPctl;
+    if (r.label === "Your AI parallelism") r.percentile = orchPctl;
   }
   return { generatedAt: (/* @__PURE__ */ new Date()).toISOString(), gradedSessions: grades.length, agglomeration, outcomeCadence, dayDigest, dayChart, timeline, walkthroughs, rubric, wordsByDay, dayThroughput, throughputCeiling: tCeiling, stackUp, criteria, parallelism, throughput, workstyle, flow, focus, expertiseMap, frontierArsenal, frontierDetail, coaching, gap, aggregate, delegationRollup, gapDist, projects };
 }

package/dist/pipeline/coding-grade.js

@@ -15753,7 +15753,7 @@ var CODING_CRITERIA = [
     label: "Abstraction \u2014 seeing the load-bearing thing, and being right",
     graded: true,
     definition: "Under genuine ambiguity (an obvious path sitting right there, no answer key): did the USER see that the obvious path was subtly wrong \u2014 or that something decisive was invisible \u2014 and do the non-obvious thing that made it right, AND were they right? The graded unit is the whole move: the CATCH (noticing the fork) -> the QUESTION (formulating it) -> the CALL (resolving it). The move surfaces three interchangeable ways, scored identically: a REFRAME (collapse a mess to its real axis), a QUESTION (pose the hard thing nobody posed), or an INSTRUMENT (build the code/chart/structure that makes the truth legible, then read it). Tag the SUBSTANCE, not the medium \u2014 a chart built (technical act) to make a human-intensity call legible is a JUDGMENT catch.",
-    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / 'don't hurt Natalie, not long-term compatible' (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
+    read: "FOUR LEVERS set the height: (1) INVISIBILITY \u2014 how non-obvious there was anything to catch; a sharp operator in the flow drives past it (the DOMINANT lever); (2) DIFFICULTY of forming the catch; (3) STAKES \u2014 'without this, things silently go wrong' is the 9-10 condition; (4) SOUNDNESS \u2014 did the call hold. THREE OVERRIDING GATES: (a) OBVIOUSNESS KILLS IT \u2014 a correct, even load-bearing catch a sharp operator reaches anyway scores LOW ('you're evaluating the person not the AI' is correct but obvious -> ~5; 'compute in the OAuth dead-time' is real but fairly obvious -> ~6). (b) LEVERAGE BEATS INSIGHT \u2014 a move whose payoff is a HIGHER RATE OF FUTURE INSIGHT (building the tool that lets him set his own boundaries; timestamp-by-timestamp structure to instruct the grader precisely) raises the bandwidth of the loop that produces every later catch, and can outrank a clean one-off reframe. (c) SOUNDNESS GATE \u2014 a confident WRONG call to a noticed fork is an ANTI-SIGNAL; an honest 'not sure, here's a range' is NOT penalized. HUNT THE FORK BLIND \u2014 reconstruct what continuing-without-noticing looked like, then check whether/why he diverged; do NOT hand yourself the fork pre-formed or you score everyone's problem-finding a 10. TAG every instance: kind = systems | judgment | both (+ optional register: human/product/measurement). Reasoning and instrument-building are the SAME criterion. ANCHORS: 10 = 'OCEAN won't change, measure effectiveness of actions' (reframe that redefines the measure) / a human-stakes call where the person overrode the analytically-clean answer on knowledge only they could have about someone involved, and was right (lived judgment, requires being the person) / 'plot it sorted, let me set the boundaries' (leverage \u2014 builds the bandwidth tool); 9 = 'word count IS the proxy \u2014 sitting there is unfakeable focus' / 'uncertainty collapses in that direction'; 8 = 'a better metric is the difficulty of content + original synthesis' (hard opening question); 7 = 'this candidate seems too strong for his judgement to be real' (real non-obvious catch, modest stakes); 6 = 'compute in the OAuth dead-time' / 'why overcomplicate it, single agent' (real but obvious); 5 = 'you're evaluating the person not the AI' (correct but obvious).",
     rungs: [
       { level: 2, marker: "A fork existed; defers the call to the AI, or takes the obvious branch; vague." },
       { level: 4, marker: "Notices something, but the question is vague or obvious \u2014 anyone in the seat asks it." },
@@ -16209,11 +16209,11 @@ var TRANSCRIPT_LINE_RE = new RegExp(`^\\s*(${NOTE})((,| and) ${NOTE})*\\s*$|^\\s
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-14.2",
+  version: "2026-07-16.2",
   // = 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).
-  rubricVersion: "f51b3a93a3",
+  rubricVersion: "7983c01d53",
   // Canonical scale = the one the product owner set for the public report
   // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
   // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
@@ -16300,6 +16300,7 @@ var DEFAULT_CALIBRATION = {
   codingTiers: {
     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." },
+    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
   },
   codingBenchmarks: {