polymath-society 0.2.12 → 0.2.14

package/dist/pipeline/coding-nutshell.js

@@ -15731,7 +15731,7 @@ 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." },
@@ -15842,11 +15842,11 @@ 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
@@ -15933,6 +15933,7 @@ 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-projects.js

@@ -49,7 +49,7 @@ 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." },
@@ -205,11 +205,11 @@ 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
@@ -296,6 +296,7 @@ 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-walkthrough.js

@@ -15995,7 +15995,7 @@ 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." },
@@ -16103,11 +16103,11 @@ 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
@@ -16194,6 +16194,7 @@ 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: {