--- slug: serial-position passed: 2026-10-18 by: the assay line (research/assay-line/queue-1018/), Codex fixer directed by claude-assaying-boyle-1014; source record by Codex, checked by Codex checked: 130 confirmed: 119 wrong: 11 unverifiable: 0 observed: 0 stale: 0 --- This was the queue-1018 Tier A decide-and-fix pass for serial-position: all 11 triage items were decided and the JSON validated before any layer edit, with the evidence re-read for every item. The record keeps seven WRONG and four MINOR verdicts; six WRONG copies of the same grid-coverage error and all four MINOR items were fixed, while the WRONG shared placard classification remains under board note d24c18. The 119 additional CONFIRMED lines preserve the checker's source readings with attribution, not a claim that the fixer independently fetched all of them. The requested batch date is 2026-10-18; the inherited N1 and N2 records carry 2026-10-19 and source revision 20ca47ceb1. Those records were left intact, and the layer files matched the requested 96caa00974 read before editing. ## Claims - WRONG | Try one private 15-word report, inspect the exact two-integer payload, and watch an anytime-valid living curve grow. | git show 20ca47ceb1:public/strata/serial-position/page-data.json; supplemented: git show 20ca47ceb1:public/_kit/collective.js (line 136); https://arxiv.org/pdf/1810.08240; recreated: node cs-counterexample.mjs | [X1; finite-grid-coverage; prose-01-2] Re-read the exact claim in index.html:7,22,86 and src/content/strata/serial-position.md:4. Re-read public/_kit/collective.js:126-137, whose upper endpoint is "hi: Math.min(1, grid[hi] + step / 2)", and public/strata/serial-position/analysis.mjs:254-266, which calls it with "alpha: positionAlpha, gridSize: 401". Recreated the shipped kit and the complete summarizeLiving path before editing: at n=3500 zeros both return [0, 0.00125]. Under iid Bernoulli mean 0.0013, this excluding event has probability (1-0.0013)^3500 = 0.010535971014164891, exceeding alpha=0.05/15 = 0.0033333333333333335. The exact-mean e-process never crosses 300 (peak 15.21291708305156); the 4001-point control includes the mean at upper bound 0.002375, but is not a proved repair. Howard et al., https://arxiv.org/abs/1810.08240, abstract, defines a confidence sequence as "uniformly valid over an unbounded time horizon". The implemented grid intervals do not earn the advertised guarantee. Fix all layer-authored coverage promises and interval labels, retaining the numerical calculation as a finite-grid approximation with coverage explicitly not guaranteed; amend the verifier so agreement with grid arithmetic is not called coverage. Preserve the original seal as history and label the revised seal by its revision date. Do not edit the shared kit. These six copies share one error and one Corrected line. - WRONG | Position recall with simultaneous anytime-valid confidence sequences | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs; supplemented: git show 20ca47ceb1:public/_kit/collective.js (line 136); https://arxiv.org/pdf/1810.08240; recreated: node cs-counterexample.mjs | [X2; finite-grid-coverage; prose-16-1] Re-read the table caption in index.html:137. Re-read public/_kit/collective.js:126-137, whose upper endpoint is "hi: Math.min(1, grid[hi] + step / 2)", and public/strata/serial-position/analysis.mjs:254-266, which calls it with "alpha: positionAlpha, gridSize: 401". Recreated the shipped kit and the complete summarizeLiving path before editing: at n=3500 zeros both return [0, 0.00125]. Under iid Bernoulli mean 0.0013, this excluding event has probability (1-0.0013)^3500 = 0.010535971014164891, exceeding alpha=0.05/15 = 0.0033333333333333335. The exact-mean e-process never crosses 300 (peak 15.21291708305156); the 4001-point control includes the mean at upper bound 0.002375, but is not a proved repair. Howard et al., https://arxiv.org/abs/1810.08240, abstract, defines a confidence sequence as "uniformly valid over an unbounded time horizon". The implemented grid intervals do not earn the advertised guarantee. Fix all layer-authored coverage promises and interval labels, retaining the numerical calculation as a finite-grid approximation with coverage explicitly not guaranteed; amend the verifier so agreement with grid arithmetic is not called coverage. Preserve the original seal as history and label the revised seal by its revision date. Do not edit the shared kit. These six copies share one error and one Corrected line. - MINOR | A fresh model context receives the same 15 nouns as visible text and is asked to report them, once for each of the 60 counterbalanced orders the living arm deals, so word identity and position are unconfounded for machines too. | git show 20ca47ceb1:research/living-experiments/model-arm/requests/serial-position.json; supplemented: git show 20ca47ceb1:public/strata/serial-position/model-arm.json; git show 20ca47ceb1:public/strata/serial-position/analysis.mjs (line 364) | [X3; completed-model-order-balance; prose-27-1] Re-read the 60 protocol prompts in research/living-experiments/model-arm/requests/serial-position.json, the raw public/strata/serial-position/model-arm.json, and analysis.mjs:364: "const completions = total - counts.error". Recounted 203 completed draws and 157 errors. The requested design has four occurrences per word-position cell, but the completed cells are unbalanced. My recount finds 0 to 3 for nemotron-nano-9b-v2, 1 to 3 for two models, and 2 to 3 for three models, so the checker's claimed lower bound of 1 in three models is not exact. The substantive finding holds. Cheap Tier A wording fix: distinguish balanced attempted orders from the unbalanced completed sample, in the page and its research README. Keep MINOR; no dated page line. - MINOR | Practice would inflate recall and reshape the curve. | https://pdfcoffee.com/murdock-pdf-pdf-free.html; supplemented: https://pdfcoffee.com/murdock-pdf-pdf-free.html | [X4; practice-reshapes-curve; prose-44-3] Re-read Murdock (1962), printed p. 483, Results, at https://pdfcoffee.com/murdock-pdf-pdf-free.html (text lines 67-68): "its effect on the serial position curves was negligible." The preceding paragraph reports significant improvement in four of six groups across sessions. This different task has no measured practice effect here. Re-read index.html:204 and app.js:190, which assert guaranteed reshaping. Cheap Tier A fix: replace that causal prediction with the directly observed design fact that later attempts repeat the same vocabulary, retaining the first-attempt rule. This makes no new empirical generalization and needs no additional practice study. X4 and X9 share this error; no dated page line. - WRONG | This layer's check lifts this page's own source out of the HTML and runs it. | git show 20ca47ceb1:research/serial-position/verify-serial-position.mjs | [X5; placard-html-extraction; prose-57-1] Re-read research/serial-position/verify-serial-position.mjs:10,19-27: analysisPath names the separate analysis.mjs, and "const analysis = await import(pathToFileURL(analysisPath).href);" executes that module. The HTML is read for assertions, not mined for executable analysis. Re-read scripts/check-placard.mjs:92-97: its MECH.lift template writes the exact claimed extraction sentence. This is a WRONG shared placard classification, not a MINOR wording issue. Per the shared-class rule, leave the generated placard unchanged and record [open note:d24c18], the census/classification board note (coordination/tending.tsv:432). Fixing the shared classifier or generator is outside this layer pass. - WRONG | Try one private 15-word report, inspect the exact two-integer payload, and watch an anytime-valid living curve grow. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs; supplemented: git show 20ca47ceb1:public/_kit/collective.js (line 136); https://arxiv.org/pdf/1810.08240; recreated: node cs-counterexample.mjs | [X6; finite-grid-coverage; dek-payload-anytime] Re-read the duplicate dek and metadata claim in index.html:7,22,86 and src/content/strata/serial-position.md:4. Re-read public/_kit/collective.js:126-137, whose upper endpoint is "hi: Math.min(1, grid[hi] + step / 2)", and public/strata/serial-position/analysis.mjs:254-266, which calls it with "alpha: positionAlpha, gridSize: 401". Recreated the shipped kit and the complete summarizeLiving path before editing: at n=3500 zeros both return [0, 0.00125]. Under iid Bernoulli mean 0.0013, this excluding event has probability (1-0.0013)^3500 = 0.010535971014164891, exceeding alpha=0.05/15 = 0.0033333333333333335. The exact-mean e-process never crosses 300 (peak 15.21291708305156); the 4001-point control includes the mean at upper bound 0.002375, but is not a proved repair. Howard et al., https://arxiv.org/abs/1810.08240, abstract, defines a confidence sequence as "uniformly valid over an unbounded time horizon". The implemented grid intervals do not earn the advertised guarantee. Fix all layer-authored coverage promises and interval labels, retaining the numerical calculation as a finite-grid approximation with coverage explicitly not guaranteed; amend the verifier so agreement with grid arithmetic is not called coverage. Preserve the original seal as history and label the revised seal by its revision date. Do not edit the shared kit. These six copies share one error and one Corrected line. - MINOR | arrivals, never a finished sample. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs; supplemented: git show 20ca47ceb1:worker.js (line 4820) | [X7; never-final-sample; never-finished] Re-read analysis.mjs:150-167: "The store cap is reached: this arm is closed to new contributions and the record below is final." Re-read worker.js:4820: "maxRows: 25000" for serial-position, and app.js, which renders the closed state and disables sends. Cheap Tier A fix: reduce the headline to its arrivals count, which remains true before and after closure. Retain MINOR as the checker did; no dated page line. - WRONG | 99.667% confidence sequence | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs; supplemented: git show 20ca47ceb1:public/_kit/collective.js (line 136); https://arxiv.org/pdf/1810.08240; recreated: node cs-counterexample.mjs | [X8; finite-grid-coverage; position-confidence-label] Re-read the column heading "99.667% confidence sequence" in index.html:138. Re-read public/_kit/collective.js:126-137, whose upper endpoint is "hi: Math.min(1, grid[hi] + step / 2)", and public/strata/serial-position/analysis.mjs:254-266, which calls it with "alpha: positionAlpha, gridSize: 401". Recreated the shipped kit and the complete summarizeLiving path before editing: at n=3500 zeros both return [0, 0.00125]. Under iid Bernoulli mean 0.0013, this excluding event has probability (1-0.0013)^3500 = 0.010535971014164891, exceeding alpha=0.05/15 = 0.0033333333333333335. The exact-mean e-process never crosses 300 (peak 15.21291708305156); the 4001-point control includes the mean at upper bound 0.002375, but is not a proved repair. Howard et al., https://arxiv.org/abs/1810.08240, abstract, defines a confidence sequence as "uniformly valid over an unbounded time horizon". The implemented grid intervals do not earn the advertised guarantee. Fix all layer-authored coverage promises and interval labels, retaining the numerical calculation as a finite-grid approximation with coverage explicitly not guaranteed; amend the verifier so agreement with grid arithmetic is not called coverage. Preserve the original seal as history and label the revised seal by its revision date. Do not edit the shared kit. These six copies share one error and one Corrected line. - MINOR | practice shapes recall curves, and only a first attempt may enter the data. | git show 20ca47ceb1:public/strata/serial-position/app.js; supplemented: https://pdfcoffee.com/murdock-pdf-pdf-free.html | [X9; practice-reshapes-curve; conditional-9] Re-read Murdock (1962), printed p. 483, Results, at https://pdfcoffee.com/murdock-pdf-pdf-free.html (text lines 67-68): "its effect on the serial position curves was negligible." The preceding paragraph reports significant improvement in four of six groups across sessions. This different task has no measured practice effect here. Re-read index.html:204 and app.js:190, which assert guaranteed reshaping. Cheap Tier A fix: replace that causal prediction with the directly observed design fact that later attempts repeat the same vocabulary, retaining the first-attempt rule. This makes no new empirical generalization and needs no additional practice study. X4 and X9 share this error; no dated page line. - WRONG | (added) Add one private 15-word report to an anytime-valid living experiment. | git show 20ca47ceb1:public/strata/serial-position/index.html (line 14); supplemented: git show 20ca47ceb1:public/_kit/collective.js (line 136); recreated: node cs-counterexample.mjs | [X10; finite-grid-coverage; added-social-anytime] Re-read the social description in index.html:14. Re-read public/_kit/collective.js:126-137, whose upper endpoint is "hi: Math.min(1, grid[hi] + step / 2)", and public/strata/serial-position/analysis.mjs:254-266, which calls it with "alpha: positionAlpha, gridSize: 401". Recreated the shipped kit and the complete summarizeLiving path before editing: at n=3500 zeros both return [0, 0.00125]. Under iid Bernoulli mean 0.0013, this excluding event has probability (1-0.0013)^3500 = 0.010535971014164891, exceeding alpha=0.05/15 = 0.0033333333333333335. The exact-mean e-process never crosses 300 (peak 15.21291708305156); the 4001-point control includes the mean at upper bound 0.002375, but is not a proved repair. Howard et al., https://arxiv.org/abs/1810.08240, abstract, defines a confidence sequence as "uniformly valid over an unbounded time horizon". The implemented grid intervals do not earn the advertised guarantee. Fix all layer-authored coverage promises and interval labels, retaining the numerical calculation as a finite-grid approximation with coverage explicitly not guaranteed; amend the verifier so agreement with grid arithmetic is not called coverage. Preserve the original seal as history and label the revised seal by its revision date. Do not edit the shared kit. These six copies share one error and one Corrected line. - WRONG | (added) 3500 contributions, with anytime-valid position intervals. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs (line 166); supplemented: git show 20ca47ceb1:public/_kit/collective.js (line 136); recreated: node cs-counterexample.mjs | [X11; finite-grid-coverage; added-ready-intervals] Re-read armState in analysis.mjs:166, which promises "with anytime-valid position intervals" in the ready branch. Re-read public/_kit/collective.js:126-137, whose upper endpoint is "hi: Math.min(1, grid[hi] + step / 2)", and public/strata/serial-position/analysis.mjs:254-266, which calls it with "alpha: positionAlpha, gridSize: 401". Recreated the shipped kit and the complete summarizeLiving path before editing: at n=3500 zeros both return [0, 0.00125]. Under iid Bernoulli mean 0.0013, this excluding event has probability (1-0.0013)^3500 = 0.010535971014164891, exceeding alpha=0.05/15 = 0.0033333333333333335. The exact-mean e-process never crosses 300 (peak 15.21291708305156); the 4001-point control includes the mean at upper bound 0.002375, but is not a proved repair. Howard et al., https://arxiv.org/abs/1810.08240, abstract, defines a confidence sequence as "uniformly valid over an unbounded time horizon". The implemented grid intervals do not earn the advertised guarantee. Fix all layer-authored coverage promises and interval labels, retaining the numerical calculation as a finite-grid approximation with coverage explicitly not guaranteed; amend the verifier so agreement with grid arithmetic is not called coverage. Preserve the original seal as history and label the revised seal by its revision date. Do not edit the shared kit. These six copies share one error and one Corrected line. - CONFIRMED | Six published aggregate rows sum to 103 participants, but the full-sample curve cannot be reproduced from the licensed material available here. | git show 20ca47ceb1:public/strata/serial-position/page-data.json; supplemented: https://pdfcoffee.com/murdock-pdf-pdf-free.html; git show 20ca47ceb1:research/serial-position/README.md | [prose-01-1] The paper prints Ns "18 16 19 19 15 16", which sum to 103. The retained build record says "The archive is therefore an unexplained subset". The downloaded archive independently recounted to 1,200 usable rows per condition, and the page ships aggregate cells only. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | A fixed vocabulary appears in one of 60 sealed orders, one word every 600 milliseconds. | git show 20ca47ceb1:public/strata/serial-position/page-data.json; supplemented: git show 20ca47ceb1:public/strata/serial-position/app.js | [prose-04-2] The data say "word_duration_ms": 600 and "list_orders": 60; app.js awaits sleep(pageData.web_protocol.word_duration_ms) for each of the 15 words. This is the configured foreground presentation schedule, subject to browser scheduling. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Scoring is exact and local. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-04-4] scoreTypedRecall applies "normalize('NFKC').toLowerCase()", then uses "normalizedWords.indexOf(token)" and counts each matched word once. The local three-word fixture scored "3 of 15 exact matches." Misspellings have no rescue route. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Your typed words never leave this browser. | git show 20ca47ceb1:public/_kit/collective.js; supplemented: git show 20ca47ceb1:public/strata/serial-position/app.js; git show 20ca47ceb1:worker.js | [prose-04-5] The actual call site sets "preparedValues = { recall_mask: score.mask, list_id: activeListId }". The kit sends JSON.stringify({ values }); the server rejects extra fields. No raw typed strings or separate total-score field are in that experiment payload. This is a payload claim, not a claim that HTTP lacks IP or user-agent metadata. For prose-07-2 the supplied permutation excerpt is irrelevant; these additional sources supply the evidence. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The 60 orders are counterbalanced: four Latin squares, so every word serves every position in exactly four of them, and word difficulty cannot pose as a position effect. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-05-1] Independent enumeration of all 60 design orders gives exactly four occurrences of every word at every position. The code divides ids into "Math.floor(listId / POSITION_COUNT)" and rotations. The design removes systematic word-position assignment; actual volunteers are randomly dealt orders, not quota-balanced. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Only the first list you complete in a visit can be sent; later lists are practice and stay on this device. | git show 20ca47ceb1:public/strata/serial-position/app.js | [prose-05-2] app.js increments "completedAttempts += 1" and disables sends when "completedAttempts > 1". resetTrial does not reset that counter. The statement concerns the ordinary page interface, with the reload limitation separately disclosed. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | recall_mask has bit i set when position i+1 was recalled. list_id reconstructs the order. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs; supplemented: git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-07-1] The scorer uses "mask |= (1 << index)" and wordsForListId reconstructs the presentation order from square and rotation. The code and payload fixture agree. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | No words, time, device facts, identity, or score are sent. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs; supplemented: git show 20ca47ceb1:public/strata/serial-position/app.js; git show 20ca47ceb1:worker.js | [prose-07-2] The actual call site sets "preparedValues = { recall_mask: score.mask, list_id: activeListId }". The kit sends JSON.stringify({ values }); the server rejects extra fields. No raw typed strings or separate total-score field are in that experiment payload. This is a payload claim, not a claim that HTTP lacks IP or user-agent metadata. For prose-07-2 the supplied permutation excerpt is irrelevant; these additional sources supply the evidence. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Nothing has been sent. | git show 20ca47ceb1:public/strata/serial-position/app.js | [prose-08-1] resetTrial sets the initial status to "Nothing has been sent." No contribution call runs during initialization; the only POST route is the send handler. Read as the current trial status. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Descriptive mean across these one-trial volunteers. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-11-1] The summary computes "totalRecallMean: mean(decoded.map((row) => row.recalled))". It is a descriptive mean of accepted rows; the page separately discloses the soft first-attempt lock. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | first 3 minus middle 5 | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-12-1] participantContrasts uses "positions.slice(0, 3)" minus "positions.slice(5, 10)": positions 1 to 3 minus 6 to 10. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Bankroll against no contrast: 1.00x now, 1.00x at its peak | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs; supplemented: git show 20ca47ceb1:public/strata/serial-position/index.html; git show 20ca47ceb1:public/strata/serial-position/app.js | [prose-13-1] These are hidden initial placeholders in index.html. bankrollText later prints "eProcess.current.toFixed(2)" and "eProcess.max.toFixed(2)" separately. They are not asserted to be the values for every nonempty sample. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | last 3 minus middle 5 | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-14-1] participantContrasts uses "positions.slice(12, 15)" minus "positions.slice(5, 10)": positions 13 to 15 minus 6 to 10. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Bankroll against no contrast: 1.00x now, 1.00x at its peak | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs; supplemented: git show 20ca47ceb1:public/strata/serial-position/index.html; git show 20ca47ceb1:public/strata/serial-position/app.js | [prose-15-1] These are hidden initial placeholders in index.html. bankrollText later prints "eProcess.current.toFixed(2)" and "eProcess.max.toFixed(2)" separately. They are not asserted to be the values for every nonempty sample. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Why these intervals: a confidence sequence is valid at every moment simultaneously, including after every visitor refreshes this page. | https://arxiv.org/abs/1810.08240 | [prose-17-1] Howard et al. define a confidence sequence as "uniformly valid over an unbounded time horizon". That definition and refresh invariance are correct. It does not establish that this page implements one; the separate interval-caption and coverage claims fail the saved numerical counterexample. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. This confirms the definition of a confidence sequence, not this implementation; the page context was corrected under X1. - CONFIRMED | A conventional 1.96-standard-error interval assumes a fixed stopping point. | https://arxiv.org/pdf/1810.08240 | [prose-17-2] Howard et al., pp. 1 to 2, discuss fixed-sample intervals and say monitoring "inflates Type I error substantially". Their Figure 1 distinguishes pointwise coverage from uniform coverage. This concerns repeated looks at accumulating data, not a new error expenditure when an unchanged page is refreshed. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Reusing it under a live counter would quietly spend more false-positive risk at every look. | https://arxiv.org/pdf/1810.08240 | [prose-17-3] Howard et al., pp. 1 to 2, discuss fixed-sample intervals and say monitoring "inflates Type I error substantially". Their Figure 1 distinguishes pointwise coverage from uniform coverage. This concerns repeated looks at accumulating data, not a new error expenditure when an unchanged page is refreshed. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The 15 position sequences divide a 5% familywise error budget by 15. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-18-1] The code sets "const positionAlpha = 0.05 / POSITION_COUNT" with POSITION_COUNT = 15. The nominal allocation is exactly as stated. Actual coverage at that allocation is disproved separately; an alpha argument alone is not a coverage proof. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. The page now states this as an alpha setting, not an achieved familywise guarantee. - CONFIRMED | The two preregistered participant contrasts divide another 5% familywise budget by two, so each contrast's e-process pays alpha 2.5% and its threshold is forty-to-one. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-18-2] Each contrast calls the e-process with "alpha: 0.025"; its threshold is "1 / opts.alpha", hence 40. The two contrast allocations sum to 0.05. The dated changelog records sealing before volunteer data. The e-process calculation is separate from the invalid finite-grid interval inversion. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The bankroll shows two numbers on purpose: the standing wealth, which can fall back when evidence reverses, and the peak, which is what the threshold rule reads because a Ville crossing is permanent evidence. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-18-3] checkedEProcess explicitly distinguishes "current" from "max" and computes the first threshold crossing from the wealth path. The existing reversal fixture and the reference e-process agree. A crossing remains a valid historical rejection event, not a claim that current wealth cannot decrease. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | A peak is never passed off here as the current e-value. | git show 20ca47ceb1:public/strata/serial-position/app.js | [prose-18-4] bankrollText prints "x now" from current and "x at its peak" from max. The page never labels the peak as current wealth. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Volunteer arrival is not random sampling, so none of this makes the crowd representative. | git show 20ca47ceb1:research/serial-position/README.md | [prose-18-5] The source says "Anonymous self-selected visitors are not a random sample" and "the intervals do not repair selection, changing traffic, shared environments, or repeated people who evade limits." The page makes the same limitation explicit. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Murdock reported six between-group conditions, 103 introductory-psychology students in total, and 80 lists per participant across four sessions. | https://pdfcoffee.com/murdock-pdf-pdf-free.html | [prose-20-1] Murdock p. 482 reports "six combinations", "80 different lists", "20 lists per session and four sessions", and "103" students from the introductory psychology course. These are the six-condition main experiment, not the separate supplementary experiment later in the paper. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The suffix in each condition is seconds per word. | https://pdfcoffee.com/murdock-pdf-pdf-free.html | [prose-20-2] The paper says the second number is "presentation time (in sec.) per item" and gives 10-2 as ten words at two seconds per item. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The page recomputes the participant total and checks every transcribed aggregate cell against a separately stored printed check. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-20-3] analyzeAnchor compares all six labels, Ns, means and SDs to printed_checks and EXPECTED_TABLE; it requires totalN = 103 and calls deriveTotalTime. The baseline verifier passed all 60 checks, including its cell-mutation controls. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Murdock 1962, Table 1, page 483. | https://pdfcoffee.com/murdock-pdf-pdf-free.html | [prose-21-1] The source text places "483" immediately before "RESULTS" and "TABLE 1". All six rows match the displayed data. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Published aggregates only. | git show 20ca47ceb1:public/strata/serial-position/page-data.json; supplemented: git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-21-2] page-data.json declares "source_kind": "published_aggregates" and "raw_data_shipped": false. Its only historical table is six group rows; analyzeAnchor hard-codes "fullSampleRawReproduced: false". The page data and renderer provide no participant-level historical curve. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The README documents 88 as an extra-list intrusion but does not explain observed out-of-range codes 0, 16, 31, 41, and 50. | https://memory.psych.upenn.edu/files/pubs/Murd62.data.tgz | [prose-25-3] The complete README says "The code 88 means that the subject made an extra-list intrusion". A numeric-token search finds 88 as its positive control but none of 0, 16, 31, 41 or 50; the file recount finds precisely those out-of-range codes. The tool could not fetch Penn, but curl with certificate verification disabled retrieved the small archive; it was deleted after reading. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | This page therefore ships neither those rows nor a curve derived from them. | git show 20ca47ceb1:public/strata/serial-position/page-data.json; supplemented: git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-25-5] page-data.json declares "source_kind": "published_aggregates" and "raw_data_shipped": false. Its only historical table is six group rows; analyzeAnchor hard-codes "fullSampleRawReproduced: false". The page data and renderer provide no participant-level historical curve. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Every studied token remains available in the model's active context. | git show 20ca47ceb1:research/living-experiments/model-arm/requests/serial-position.json | [prose-27-2] The request places all 15 words between "BEGIN LIST" and "END LIST", followed by "Report now." Nothing removes those prompt tokens before completion. This is visible-context extraction, unlike the timed disappearance in the browser task. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | That is in-context extraction, not human encoding followed by recall. | git show 20ca47ceb1:research/living-experiments/model-arm/requests/serial-position.json | [prose-27-3] The request places all 15 words between "BEGIN LIST" and "END LIST", followed by "Report now." Nothing removes those prompt tokens before completion. This is visible-context extraction, unlike the timed disappearance in the browser task. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | A similar edge shape would not identify a shared mechanism. | https://memory.psych.upenn.edu/files/pubs/HowaKaha02.pdf | [prose-27-4] Howard and Kahana describe decay, displacement, interference and context explanations for recency, then demonstrate a temporal-context alternative. The page draws the justified identification limit: an output shape alone does not distinguish mechanisms across humans and models. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Per model, every draw accounted for | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-32-1] summarizeModel increments exactly one category per raw draw and checks "accountedFor" against total. Independent Python parsing confirms 360 = 203 values + 0 refusals + 0 unparseable + 157 errors. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Exact-match reporting rate by presentation position, pooled as a heterogeneous mixture across the models above; denominator is every completed draw | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-33-1] The source explicitly sets "const completions = total - counts.error" and rates use that denominator. All 203 completed draws remain; the 157 rail failures are listed separately. Per-model totals sum to the pooled 360. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The parser applies Unicode NFKC and lowercase, splits only on comma, newline, or semicolon, trims surrounding ASCII punctuation, then exact-matches whole tokens against the order named by each run's protocol id, so a word scores at the position it was shown, not at a canonical position. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-34-1] classify normalizes NFKC and lowercase, splits only on comma, newline or semicolon, removes surrounding ASCII punctuation, and uses the named protocol order for exact whole-token matches. Independent parsing of all raw answers reproduced every displayed position count. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | It gives no fuzzy, stem, semantic, or manual credit. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-34-2] classify normalizes NFKC and lowercase, splits only on comma, newline or semicolon, removes surrounding ASCII punctuation, and uses the named protocol order for exact whole-token matches. Independent parsing of all raw answers reproduced every displayed position count. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Each raw draw becomes value, refusal, unparseable, or rail error; completions all stay in the position denominator, and rail errors are counted in the open rather than hidden inside it. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-34-3] The source explicitly sets "const completions = total - counts.error" and rates use that denominator. All 203 completed draws remain; the 157 rail failures are listed separately. Per-model totals sum to the pooled 360. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The literal below is part of this HTML, not recalculated to flatter the current file. | git show 20ca47ceb1:public/strata/serial-position/index.html | [prose-36-1] index.html contains the literal "3631e768af9fd774b1a61c0c48a0a3dbc8f6d3f53f01d11e35027075859d1796" in both the pre element and window.SERIAL_POSITION_SEALED. SHA-256 of the pinned and fetched analysis bytes equals it. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. The old digest is now retained as the original seal; the explicitly dated wording revision has digest 0e1ac603693f0c4eec5014253e5035ba37306fa960a61d4fb354950f39e00f13. - CONFIRMED | Table 1 cells, page numbers, and provenance ship in page-data.json. | git show 20ca47ceb1:public/strata/serial-position/page-data.json | [prose-38-1] The fetched page-data.json is byte-identical to the pinned file. It contains table_1.rows, "pages": "483", and the stated transcription provenance. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The code checks six condition labels, Ns, means, SDs, the stated total of 103, and recomputes one published relationship from the cells: the total-presentation-time regularity. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-38-2] analyzeAnchor compares all six labels, Ns, means and SDs to printed_checks and EXPECTED_TABLE; it requires totalN = 103 and calls deriveTotalTime. The baseline verifier passed all 60 checks, including its cell-mutation controls. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Aggregate cells are never presented as raw participants, and no participant-level reproduction exists here. | git show 20ca47ceb1:public/strata/serial-position/page-data.json; supplemented: git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-38-3] page-data.json declares "source_kind": "published_aggregates" and "raw_data_shipped": false. Its only historical table is six group rows; analyzeAnchor hard-codes "fullSampleRawReproduced: false". The page data and renderer provide no participant-level historical curve. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The unexplained Penn subset cannot reproduce the full published sample. | git show 20ca47ceb1:public/strata/serial-position/page-data.json | [prose-39-2] Six archive conditions at an inferred 15 people each provide 90 participant blocks rather than the six printed groups totaling 103. The record describes the subset as unexplained, and does not infer identities for missing participants. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The missing comparison is shown as missing. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-39-3] analyzeAnchor returns "comparisonGateOpen: false" even when its six-row aggregate check passes; app.js renders "CLOSED" with the missing-data explanation. The page displays the absent comparison as absent. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The 1962 study used auditory words, 90-second handwritten recall, 80 lists, and group testing. | https://pdfcoffee.com/murdock-pdf-pdf-free.html | [prose-40-1] Murdock p. 482 says "Group testing was used", lists were read aloud with a metronome, recall lasted "1.5 min.", and participants "wrote down" recalled words. Four sessions of 20 lists give 80. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | This page uses one fast visual list, untimed typed report, a fixed vocabulary, and self-selected web volunteers. | git show 20ca47ceb1:public/strata/serial-position/page-data.json; supplemented: git show 20ca47ceb1:public/strata/serial-position/app.js; https://pdfcoffee.com/murdock-pdf-pdf-free.html | [prose-40-2] page-data.json says "presentation": "visual", 600 milliseconds per word and a fixed 15-word vocabulary. app.js removes each word and opens an untimed text field. These differ from the primary paper in modality, duration, repeated lists and recruitment. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | It is a new analogue, not an exact replication. | git show 20ca47ceb1:public/strata/serial-position/page-data.json; supplemented: git show 20ca47ceb1:public/strata/serial-position/app.js; https://pdfcoffee.com/murdock-pdf-pdf-free.html | [prose-40-3] page-data.json says "presentation": "visual", 600 milliseconds per word and a fixed 15-word vocabulary. app.js removes each word and opens an untimed text field. These differ from the primary paper in modality, duration, repeated lists and recruitment. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The server accepts only recall_mask and list_id. | git show 20ca47ceb1:worker.js | [prose-41-1] The registry declares only "recall_mask" and "list_id". collectiveValidate requires both integer fields and rejects any unknown key. No delay, compliance, intrusion-count or nonce field is available in this fixed schema. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | It cannot support the scout's delay randomization, compliance flags, intrusion counts, or one-use nonce. | git show 20ca47ceb1:worker.js | [prose-41-2] The registry declares only "recall_mask" and "list_id". collectiveValidate requires both integer fields and rejects any unknown key. No delay, compliance, intrusion-count or nonce field is available in this fixed schema. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Exact normalized whole-word matches count. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-42-1] scoreTypedRecall applies "normalize('NFKC').toLowerCase()", then uses "normalizedWords.indexOf(token)" and counts each matched word once. The local three-word fixture scored "3 of 15 exact matches." Misspellings have no rescue route. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Fuzzy spelling rescue does not. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-42-2] scoreTypedRecall applies "normalize('NFKC').toLowerCase()", then uses "normalizedWords.indexOf(token)" and counts each matched word once. The local three-word fixture scored "3 of 15 exact matches." Misspellings have no rescue route. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Duplicate and unmatched counts are shown locally, then discarded. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs; supplemented: git show 20ca47ceb1:public/strata/serial-position/app.js | [prose-42-3] The local scorer returns duplicates and intrusions; app.js prints them in localScore but places only mask and list id in preparedValues. "discarded" concerns the submitted dataset; the displayed counts can remain visible locally. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | This can undercount misspellings. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-42-4] scoreTypedRecall applies "normalize('NFKC').toLowerCase()", then uses "normalizedWords.indexOf(token)" and counts each matched word once. The local three-word fixture scored "3 of 15 exact matches." Misspellings have no rescue route. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Confidence sequences permit arbitrary peeking at the running arrivals. | https://arxiv.org/pdf/1810.08240 | [prose-43-1] Howard et al. explicitly allow "Arbitrary stopping rules". This is a correct property of genuine confidence sequences under their assumptions. It is not evidence that the finite-grid implementation on this page has that property. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. This confirms the definition of a confidence sequence, not this implementation; the page context was corrected under X1. - CONFIRMED | They do not turn voluntary, anonymous traffic into a random sample, block determined bots, or repair changes in who visits. | git show 20ca47ceb1:research/serial-position/README.md | [prose-43-2] The source says "Anonymous self-selected visitors are not a random sample" and "the intervals do not repair selection, changing traffic, shared environments, or repeated people who evade limits." The page makes the same limitation explicit. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The store caps at 25,000 rows; at the cap the page says the arm is closed instead of pretending it still grows. | git show 20ca47ceb1:worker.js; supplemented: git show 20ca47ceb1:public/strata/serial-position/app.js; git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-43-3] The registry has "maxRows: 25000"; the GET response derives full from that cap, armState reports a final closed record, and the UI disables sending. This is also the contradiction to the separate "never a finished sample" headline. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Only the first completed list in a visit can be sent; later completions are labeled practice and blocked on this device. | git show 20ca47ceb1:public/strata/serial-position/app.js | [prose-44-1] app.js increments "completedAttempts += 1" and disables sends when "completedAttempts > 1". resetTrial does not reset that counter. The statement concerns the ordinary page interface, with the reload limitation separately disclosed. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | A reload evades that lock, the frozen two-integer payload cannot record attempt counts, and the server cannot tell a rehearsed volunteer from a fresh one. | git show 20ca47ceb1:public/strata/serial-position/app.js | [prose-44-2] completedAttempts starts at zero in page memory and is not transmitted. Reloading resets it, while localStorage separately blocks browsers that already contributed. The server has only the two bounded integers, so it cannot infer pre-send rehearsals. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Models see all words in active context. | git show 20ca47ceb1:research/living-experiments/model-arm/requests/serial-position.json | [prose-45-1] The request places all 15 words between "BEGIN LIST" and "END LIST", followed by "Report now." Nothing removes those prompt tokens before completion. This is visible-context extraction, unlike the timed disappearance in the browser task. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Humans no longer do. | git show 20ca47ceb1:public/strata/serial-position/app.js | [prose-45-2] After the timed loop app.js replaces the word display with "report what remains" before revealing the recall field. The claim concerns the normal UI, not determined access to source code or note-taking. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The same curve shape cannot prove the same memory mechanism. | https://memory.psych.upenn.edu/files/pubs/HowaKaha02.pdf | [prose-45-3] Howard and Kahana describe decay, displacement, interference and context explanations for recency, then demonstrate a temporal-context alternative. The page draws the justified identification limit: an output shape alone does not distinguish mechanisms across humans and models. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | A perfect 15 of 15 machine report is allowed and would be a flat ceiling, not a failed experiment. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-45-4] All 15 exact matches produce mask 32767, accepted by both parser and wire schema. Five models have mean 15.00 in the shipped completed draws, so ceiling performance is actually preserved rather than rejected. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The request ships the same 60 counterbalanced 15-word orders the living arm deals, one draw per order per census model, at temperature 1. | git show 20ca47ceb1:research/living-experiments/model-arm/requests/serial-position.json | [prose-46-1] The complete request has 60 protocols, each n = 1 and temperature = 1; each prompt matches wordsForListId(id). The actual results contain all 60 attempted protocol ids for each of six models. The truncated source excerpt alone does not show the temperature; the full file does. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The pooled curve is a mixture across whatever models the census reached, shown per model above it. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-46-2] The source explicitly sets "const completions = total - counts.error" and rates use that denominator. All 203 completed draws remain; the 157 rail failures are listed separately. Per-model totals sum to the pooled 360. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Primacy and recency do not uniquely prove separate short-term and long-term stores. | https://memory.psych.upenn.edu/files/pubs/HowaKaha02.pdf | [prose-47-1] Howard and Kahana conclude that recency and contiguity "probably do not depend on short-term store". Their alternative account is a direct counterexample to inferring two stores uniquely from curve shape. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Inhibition, two-store, temporal-context, and distinctiveness accounts make different commitments. | https://memory.psych.upenn.edu/files/pubs/HowaKaha02.pdf | [prose-47-2] The supplied excerpt covers distinctiveness only. The full paper also contrasts temporal context, short-term-store accounts, and transient inhibition in response suppression. Thus the modest statement that these accounts have different commitments is supported, with the excerpt supplemented. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | This page estimates a shape, not anatomy. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [prose-47-3] The actual outputs are position rates, contrasts and word counts. There is no anatomical model or neural measurement in summarizeLiving. The separate coverage defect does not turn the page into an anatomical inference. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The primary paper for Table 1, the six conditions, 103 participants, 80-list procedure, and its fitted descriptions. | https://pdfcoffee.com/murdock-pdf-pdf-free.html | [prose-50-1] The source is Murdock's 1962 primary article text: its procedure, Table 1 and later fitted descriptions are present. Crossref independently identifies the title "The serial position effect of free recall." Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Long-context question answering and key-value retrieval often lost information in the middle. | https://aclanthology.org/2024.tacl-1.9/ | [prose-52-1] The ACL abstract reports that performance "significantly degrades" when relevant information is in the middle. The full paper tests multi-document question answering and key-value retrieval. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | It did not test free recall. | https://aclanthology.org/2024.tacl-1.9/; supplemented: https://aclanthology.org/2024.tacl-1.9.pdf | [prose-52-2] Read the full paper's two experimental task sections. A search for "free recall" finds only the Murdock reference, while positive controls find "multi-document question answering" and "key-value retrieval" 12 times each. No free-recall experiment is described. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Position effects across language-model classification labels and reordered summarization. | https://aclanthology.org/2025.findings-acl.52.pdf | [prose-53-1] The full Guo and Vosoughi paper has "Label Shuffling Experiments" and "Summarization Experiments", with observed serial-position effects in its results. Its publication predates this layer. The PDF was downloaded and read after the checker marked it unreadable. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Machine serial-position effects are not a novelty claim here. | https://aclanthology.org/2025.findings-acl.52.pdf | [prose-53-2] The full Guo and Vosoughi paper has "Label Shuffling Experiments" and "Summarization Experiments", with observed serial-position effects in its results. Its publication predates this layer. The PDF was downloaded and read after the checker marked it unreadable. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Another operable question about memory, prediction, and the information left in a sequence. | git show 20ca47ceb1:src/content/strata/you-already-know-the-rest.md; supplemented: git show 20ca47ceb1:src/content/strata/you-already-know-the-rest.md | [prose-54-1] The related layer says "guess a hidden public-domain passage one letter at a time" and explains that its resulting sample estimates are illustrative. This layer removes presented words and asks for recall. The connection describes these two actual instruments. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | A second sealed instrument whose claim can turn red after it ships. | git show 20ca47ceb1:src/content/strata/the-guess-it-showed-you-first.md; supplemented: git show 20ca47ceb1:public/strata/the-guess-it-showed-you-first/index.html | [prose-55-1] The related page says "Every choice waits for a SHA-256 digest" and verifies its revealed preimage. This page prints and checks the analysis digest. The supplied prose-63-1 excerpt ends before the SHA-256 passage; reading the complete pinned page supplies it. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | A human behavior result that changes when the measurement rule is audited. | git show 20ca47ceb1:src/content/strata/nothing-funny-happened.md; supplemented: git show 20ca47ceb1:public/strata/nothing-funny-happened/index.html | [prose-56-1] The related layer asks "Does a laugh ever overlap the laugher's own words?" and reports "Two. Not two percent. Two laughs." It then discusses zero-duration annotations. Its source-record excerpt for related-3-2 stops before that discussion; the full page supports the connection. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Both turn memory into an operable sequence rather than a metaphor. | git show 20ca47ceb1:src/content/strata/you-already-know-the-rest.md; supplemented: git show 20ca47ceb1:src/content/strata/you-already-know-the-rest.md | [prose-61-1] The related layer says "guess a hidden public-domain passage one letter at a time" and explains that its resulting sample estimates are illustrative. This layer removes presented words and asks for recall. The connection describes these two actual instruments. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Both instruments print a SHA-256 commitment that can turn red after publication. | git show 20ca47ceb1:public/strata/the-guess-it-showed-you-first/index.html; supplemented: git show 20ca47ceb1:public/strata/the-guess-it-showed-you-first/index.html | [prose-63-1] The related page says "Every choice waits for a SHA-256 digest" and verifies its revealed preimage. This page prints and checks the analysis digest. The supplied prose-63-1 excerpt ends before the SHA-256 passage; reading the complete pinned page supplies it. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Both are human-behavior measurements whose strongest conclusion depends on auditing what the instrument could record. | git show 20ca47ceb1:src/content/strata/nothing-funny-happened.md; supplemented: git show 20ca47ceb1:public/strata/nothing-funny-happened/index.html | [prose-65-1] The related layer asks "Does a laugh ever overlap the laugher's own words?" and reports "Two. Not two percent. Two laughs." It then discusses zero-duration annotations. Its source-record excerpt for related-3-2 stops before that discussion; the full page supports the connection. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | A sealed living experiment presents 15 words once, scores recall privately, and sends only a bitmask plus the list order after explicit consent. | git show 20ca47ceb1:public/strata/serial-position/app.js | [plain-protocol] The caller presents 15 words, computes scoreTypedRecall locally, previews exactly the two integer fields and invokes contribute only from the send button handler. The analysis hash matched the pinned and live bytes. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Six published aggregate rows sum to 103 participants, but the full-sample curve cannot be reproduced here. | git show 20ca47ceb1:public/strata/serial-position/page-data.json; supplemented: https://pdfcoffee.com/murdock-pdf-pdf-free.html; git show 20ca47ceb1:research/serial-position/README.md | [social-closed-comparison] The paper prints Ns "18 16 19 19 15 16", which sum to 103. The retained build record says "The archive is therefore an unexplained subset". The downloaded archive independently recounted to 1,200 usable rows per condition, and the page ships aggregate cells only. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | A sealed, living 15-word serial-position experiment with its missing historical comparison left visibly missing. | git show 20ca47ceb1:public/strata/serial-position/app.js | [social-sealed] verifySeal compares fetched bytes to the literal printed digest; analyzeAnchor always returns "comparisonGateOpen: false". The page's 15-word trial and visible closed historical comparison are implemented. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Fifteen words. One pass. | git show 20ca47ceb1:public/strata/serial-position/app.js | [one-pass] runTrial loops once through the 15-element activeWords array, then removes the display. Later completed trials are separately labeled practice. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The table passes. The reproduction gate stays closed. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [table-pass-heading] analyzeAnchor returns "comparisonGateOpen: false" even when its six-row aggregate check passes; app.js renders "CLOSED" with the missing-data explanation. The page displays the absent comparison as absent. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | weighted mean is derived here, not printed by the paper | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs; supplemented: https://pdfcoffee.com/murdock-pdf-pdf-free.html | [weighted-not-printed] Independent arithmetic gives 809.06 / 103 = 7.854951456310679. The primary Table 1 prints six group means rather than that pooled number; its text also describes six group comparisons. The source displays "weighted mean is derived here" honestly. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The interactive experiment, aggregate anchor, sealed analysis, model-arm accounting, and complete uncertainty ledger are at | git show 20ca47ceb1:public/strata/serial-position/index.html | [md-location] The pinned HTML contains the experiment, historical table, analysis seal, machine accounting and twelve limitation blocks at the declared href. The live route returned HTTP 200; its visible content matches the pinned page. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Murdock, 1962 | https://api.crossref.org/works/10.1037%2Fh0045106 | [cite-murdock] Crossref gives Bennet B. Murdock, the title "The serial position effect of free recall.", DOI 10.1037/h0045106 and issued year 1962. The page's short citation is accurate. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Penn archive | https://memory.psych.upenn.edu/files/pubs/Murd62.data.tgz | [cite-penn] The URL serves a small gzip tar archive, with six condition files and a README headed "MURDOCK 1962". Its SHA-256 equals the author's recorded hash. The automatic checker's fetch failure was resolved manually. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Liu et al., 2024 | https://aclanthology.org/2024.tacl-1.9/ | [cite-liu] ACL identifies "Nelson F. Liu" and publication year 2024. The tool's "gives Liu the initial F." warning is a regex error that treats the middle initial as the first name; the supplied citation is correct. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Guo and Vosoughi, 2025 | https://aclanthology.org/2025.findings-acl.52/ | [cite-guo] ACL identifies "Xiaobo Guo and Soroush Vosoughi" and publication year 2025. Crossref agrees with the DOI. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Murdock 1962, Table 1, page 483. Published aggregates only. | https://pdfcoffee.com/murdock-pdf-pdf-free.html | [table-row-10-2] The primary Table 1 prints this condition, N, mean and SD in the displayed column order. Exact rendered cells: 10-2; 18; 6.39; 6.39; 0.76; match. The OCR renders 0.76 as "0:76"; the remaining figures and table structure make the decimal transcription clear. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Murdock 1962, Table 1, page 483. Published aggregates only. | https://pdfcoffee.com/murdock-pdf-pdf-free.html | [table-row-20-1] The primary Table 1 prints this condition, N, mean and SD in the displayed column order. Exact rendered cells: 20-1; 16; 6.87; 6.87; 1.16; match. The OCR renders 0.76 as "0:76"; the remaining figures and table structure make the decimal transcription clear. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Murdock 1962, Table 1, page 483. Published aggregates only. | https://pdfcoffee.com/murdock-pdf-pdf-free.html | [table-row-15-2] The primary Table 1 prints this condition, N, mean and SD in the displayed column order. Exact rendered cells: 15-2; 19; 8.25; 8.25; 1.40; match. The OCR renders 0.76 as "0:76"; the remaining figures and table structure make the decimal transcription clear. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Murdock 1962, Table 1, page 483. Published aggregates only. | https://pdfcoffee.com/murdock-pdf-pdf-free.html | [table-row-30-1] The primary Table 1 prints this condition, N, mean and SD in the displayed column order. Exact rendered cells: 30-1; 19; 8.82; 8.82; 1.98; match. The OCR renders 0.76 as "0:76"; the remaining figures and table structure make the decimal transcription clear. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Murdock 1962, Table 1, page 483. Published aggregates only. | https://pdfcoffee.com/murdock-pdf-pdf-free.html | [table-row-20-2] The primary Table 1 prints this condition, N, mean and SD in the displayed column order. Exact rendered cells: 20-2; 15; 8.53; 8.53; 2.08; match. The OCR renders 0.76 as "0:76"; the remaining figures and table structure make the decimal transcription clear. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Murdock 1962, Table 1, page 483. Published aggregates only. | https://pdfcoffee.com/murdock-pdf-pdf-free.html | [table-row-40-1] The primary Table 1 prints this condition, N, mean and SD in the displayed column order. Exact rendered cells: 40-1; 16; 8.24; 8.24; 1.08; match. The OCR renders 0.76 as "0:76"; the remaining figures and table structure make the decimal transcription clear. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | CLOSED: no authorized complete raw dataset was shipped, so no living-to-Murdock curve comparison is drawn. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [built-anchorGate] analyzeAnchor returns "comparisonGateOpen: false" even when its six-row aggregate check passes; app.js renders "CLOSED" with the missing-data explanation. The page displays the absent comparison as absent. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Historical aggregate check, analysis seal, and living-arm state loaded. | git show 20ca47ceb1:public/strata/serial-position/app.js | [built-pageStatus] init loads page-data, runs renderAnchor, verifies the seal, fetches the living arm and loads models before setting this status. The offline renderer reproduced the exact line. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | This arm opened 2026-08-21. No contributions yet. Counts appear with the first contribution; position rates and confidence sequences appear at 10. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs; supplemented: https://artwaste.land/api/collective/serial-position | [built-empty-arm] armState([]) reproduces the exact empty message with SMALL_N = 10. The live GET also returned "n":0 and "rows":[] on 2026-10-19. This is a conditional UI claim, not evidence that the live arm can never receive a contribution. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. This confirms the old conditional string and n=10 threshold; the repaired string says grid intervals and does not claim confidence coverage. - CONFIRMED | Shannon's reader predicts hidden letters and this reader reports vanished words, while each page keeps its measurement limits visible. | git show 20ca47ceb1:public/strata/you-already-know-the-rest/index.html; supplemented: git show 20ca47ceb1:src/content/strata/you-already-know-the-rest.md | [related-1-2] The related layer says "guess a hidden public-domain passage one letter at a time" and explains that its resulting sample estimates are illustrative. This layer removes presented words and asks for recall. The connection describes these two actual instruments. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | One fixes a predictor before a move, while this one fixes the analysis before volunteer data arrive. | git show 20ca47ceb1:public/strata/the-guess-it-showed-you-first/index.html; supplemented: git show 20ca47ceb1:public/strata/the-guess-it-showed-you-first/index.html | [related-2-2] The related page says "Every choice waits for a SHA-256 digest" and verifies its revealed preimage. This page prints and checks the analysis digest. The supplied prose-63-1 excerpt ends before the SHA-256 passage; reading the complete pinned page supplies it. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. The current related note now names the original seal and dated wording revision, so it does not imply that the revised bytes predate the crowd. - CONFIRMED | Missing archive participants matter here as much as missing overlapping laughs matter there. | git show 20ca47ceb1:public/strata/nothing-funny-happened/index.html; supplemented: git show 20ca47ceb1:public/strata/nothing-funny-happened/index.html | [related-3-2] The related layer asks "Does a laugh ever overlap the laugher's own words?" and reports "Two. Not two percent. Two laughs." It then discusses zero-duration annotations. Its source-record excerpt for related-3-2 stops before that discussion; the full page supports the connection. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Counts are shown; the position display waits for | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [conditional-1] armState returns the small branch only for 1 to 9 rows and app.js then hides the position table. A two-row fixture prints "Position estimates wait for 10 contributions." Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The store cap is reached: this arm is closed to new contributions and the record below is final. | git show 20ca47ceb1:public/strata/serial-position/analysis.mjs | [conditional-2] armState adds this exact sentence when full is true, including the empty, small and ready branches. The worker defines full from the 25,000-row cap. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The living display is closed because the historical aggregate check or analysis seal is red. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-3] refreshLiving tests "!anchor.ok || !sealOkay" before fetching the living arm and reports this exact closed status. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | No substitute data are shown. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-4] The fetch/render catch prints this sentence and supplies no fallback dataset. A previously rendered table can remain on a later failed refresh, but it is real prior data rather than substitute data. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Presenting 15 words. Keep them in mind. Nothing is being sent. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-5] runTrial displays 15 words; no contribute call occurs in that loop. The status refers to the current trial and experiment payload. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | repeated studied tokens and | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-6] scoreTypedRecall separately counts repeated studied tokens and unmatched tokens; only mask and list id enter preparedValues. The synthetic local fixture printed zero of each, without making a network POST. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | unmatched tokens stay on this device and are not in the payload. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-7] scoreTypedRecall separately counts repeated studied tokens and unmatched tokens; only mask and list id enter preparedValues. The synthetic local fixture printed zero of each, without making a network POST. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Payload prepared locally. This browser already contributed, so another send is disabled. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-8] preparePayload reads hasContributed and disables send when it is true. The exact corresponding status is present in that branch. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Payload prepared locally, but the arm has reached its row cap and is closed. Nothing can be sent. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-10] When armFull is true, preparePayload disables sending and selects this message. The handler also guards armFull before submitting. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The raw typed words never left this browser. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-11] The successful-send branch makes this raw-text claim after posting preparedValues, whose fields are recall_mask and list_id. Raw typed strings never enter that request body. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | carry a list_id outside the 60-order design (hand-crafted sends); they are pooled by position but sit outside the word-balance guarantee. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-12] decodeRow accepts wire list ids 60 through 63 but marks them offDesign. summarizeLiving pools their position-indexed masks, counts them, and app.js prints this warning. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | carry an unknown protocol id, so words cannot be mapped to positions. The position table is withheld. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-13] A parsed run with an unrecognized protocol id increments unplacedValues; positionsValid becomes false and app.js hides modelPositionsWrap. The original verifier includes a positive parsed-value control and an unknown-id failure fixture. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The model file exists but contains no runs. The arm is absent, not zero. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-14] An existing empty model file produces total = 0, null position rates and this explicit absence message. It is not interpreted as observed zero recall. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | No machine distribution is inferred. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-15] loadModel catches a failed fetch and prints a pending status ending with this sentence. There is no inferred replacement model distribution. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | The transcribed cells no longer show the published pattern. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-16] This failure message is selected when deriveTotalTime fails. The pinned verifier's altered-cell fixture confirms that a structural change can turn the check red. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | Do not interpret the living display. | git show 20ca47ceb1:public/strata/serial-position/app.js | [conditional-17] verifySeal puts this instruction in the branch where the fetched analysis hash disagrees with the printed literal. The hash comparison itself is independently reproducible. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | 3631e768af9fd774b1a61c0c48a0a3dbc8f6d3f53f01d11e35027075859d1796 | git show 20ca47ceb1:public/strata/serial-position/index.html | [printed-hash] index.html contains the literal "3631e768af9fd774b1a61c0c48a0a3dbc8f6d3f53f01d11e35027075859d1796" in both the pre element and window.SERIAL_POSITION_SEALED. SHA-256 of the pinned and fetched analysis bytes equals it. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. The old digest is now retained as the original seal; the explicitly dated wording revision has digest 0e1ac603693f0c4eec5014253e5035ba37306fa960a61d4fb354950f39e00f13. - CONFIRMED | (added) Total words recalled in arrival order: 0, 3. Position estimates wait for 10 contributions. | git show 20ca47ceb1:public/strata/serial-position/app.js (line 121) | [added-small-counts] Conditional rendering fixture, not a live contribution count: masks 0 and 7 produce "0, 3" in arrival order; n = 2 is below SMALL_N = 10. The exact combined status was omitted from the source record. Recreated offline by node audit-render.mjs; no rows were submitted. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. - CONFIRMED | (added) 3 of 15 exact matches. | git show 20ca47ceb1:public/strata/serial-position/app.js (line 184) | [added-local-score] Conditional rendering fixture, not a reader observation: the first three words of order 0 were entered in the local mock DOM, producing the exact displayed score and payload {recall_mask: 7, list_id: 0}. The source record retained the duplicate/unmatched suffix but omitted this numerical prefix. Recreated offline by node audit-render.mjs. Evidence and verdict carried from N2-codex.md; the source revision remains pinned there. ## Findings - [fixed] X1 (finite-grid-coverage): withdrew the unsupported coverage claim from the HTML dek, meta description and JSON-LD, and the markdown dek. The numerical bounds remain finite-grid approximations with coverage explicitly not guaranteed. All six copies share the single Corrected 2026-10-18 line in the HTML and markdown; the layer verifier now distinguishes arithmetic agreement from coverage and retains the exclusion counterexample. Review-and-repair made the independent Bernoulli sampling assumption explicit in both correction copies and the README changelog, matching the iid counterexample recorded above; the event probability is not a claim about volunteer arrivals. - [fixed] X2 (finite-grid-coverage): withdrew the unsupported coverage claim from the position table caption. The numerical bounds remain finite-grid approximations with coverage explicitly not guaranteed. All six copies share the single Corrected 2026-10-18 line in the HTML and markdown; the layer verifier now distinguishes arithmetic agreement from coverage and retains the exclusion counterexample. - [fixed] X3 (completed-model-order-balance), MINOR: the page and research README now distinguish balanced attempted orders from the unbalanced completed sample. No numerical data changed and no dated page line was added. - [fixed] X4 (practice-reshapes-curve), MINOR: replaced the unmeasured reshaping prediction with the same-vocabulary design fact and an explicit statement that this page has not measured its effect on recall. No dated page line was added. - [open note:d24c18] X5 (placard-html-extraction), WRONG: the verifier imports the separate analysis.mjs and reads HTML for assertions, while the shared MECH.lift placard template says it extracts source from HTML. Left unchanged under the shared census/classification rule. No downgrade and no per-layer correction of generated text. - [fixed] X6 (finite-grid-coverage): withdrew the unsupported coverage claim from all duplicate dek and metadata promises. The numerical bounds remain finite-grid approximations with coverage explicitly not guaranteed. All six copies share the single Corrected 2026-10-18 line in the HTML and markdown; the layer verifier now distinguishes arithmetic agreement from coverage and retains the exclusion counterexample. - [fixed] X7 (never-final-sample), MINOR: shortened the headline to the arrivals count. The existing cap message still states that the arm closes and the record is final. No dated page line was added. - [fixed] X8 (finite-grid-coverage): withdrew the unsupported coverage claim from the position interval column heading. The numerical bounds remain finite-grid approximations with coverage explicitly not guaranteed. All six copies share the single Corrected 2026-10-18 line in the HTML and markdown; the layer verifier now distinguishes arithmetic agreement from coverage and retains the exclusion counterexample. - [fixed] X9 (practice-reshapes-curve), MINOR: the practice status now says later attempts repeat the same vocabulary and retains the first-attempt rule. No dated page line was added. - [fixed] X10 (finite-grid-coverage): withdrew the unsupported coverage claim from the social description. The numerical bounds remain finite-grid approximations with coverage explicitly not guaranteed. All six copies share the single Corrected 2026-10-18 line in the HTML and markdown; the layer verifier now distinguishes arithmetic agreement from coverage and retains the exclusion counterexample. - [fixed] X11 (finite-grid-coverage): withdrew the unsupported coverage claim from the ready-state status and the empty-state interval wording. The numerical bounds remain finite-grid approximations with coverage explicitly not guaranteed. All six copies share the single Corrected 2026-10-18 line in the HTML and markdown; the layer verifier now distinguishes arithmetic agreement from coverage and retains the exclusion counterexample.