A living experiment, opened 2026-08-21

Do We Lose the Middle?

Six published aggregate rows sum to 103 participants, but the full-sample curve cannot be reproduced from the licensed material available here. Try one private 15-word report, inspect the exact two-integer payload, and watch a living recall curve grow.

Loading the historical check and sealed analysis.

Arm B, the living

Fifteen words. One pass.

Press start. A fixed vocabulary appears in one of 60 sealed orders, one word every 600 milliseconds. Then type whatever remains. Scoring is exact and local. Your typed words never leave this browser.

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. Only the first list you complete in a visit can be sent; later lists are practice and stay on this device.

Press start when you are ready.

The volunteer counter

0 arrivals.

Reading the arm.

What these bounds mean: the position and contrast intervals use a 401-point grid of candidate means, widened by half a grid step at each edge. That calculation can exclude a true mean more often than its stated error budget allows. These are grid approximations with no guaranteed coverage. A confidence sequence would be valid at every moment under its assumptions; these displayed intervals do not have that guarantee.

The position calculation uses alpha 0.05/15 and the contrast intervals use alpha 0.025; those settings do not establish coverage for the grid intervals. 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. 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. A peak is never passed off here as the current e-value. Volunteer arrival is not random sampling, so none of this makes the crowd representative.

Arm A, the published record

The table passes. The reproduction gate stays closed.

Murdock reported six between-group conditions, 103 introductory-psychology students in total, and 80 lists per participant across four sessions. The suffix in each condition is seconds per word. The page recomputes the participant total and checks every transcribed aggregate cell against a separately stored printed check.

Murdock 1962, Table 1, page 483. Published aggregates only.
conditionNtranscribed meanprinted checktranscribed SDagreement
sum / weighted arithmeticcomputingcomputingweighted mean is derived here, not printed by the paper

Checking aggregate cells.

Checking the derived total-time regularity.

Comparison gate: checking

The public Penn archive contains 1,200 usable rows per condition, consistent with 15 people doing 80 trials, while the paper reports condition Ns from 15 to 19. The 10-word file carries a 1,201st row holding only the malformed value 0. The README documents 88 as an extra-list intrusion but does not explain observed out-of-range codes 0, 16, 31, 41, and 50. No explicit redistribution license was found. This page therefore ships neither those rows nor a curve derived from them.

Arm C, the machines

The same report format, not the same memory.

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. The attempted orders balance word identity across positions; missing completions unbalance the plotted sample. Every studied token remains available in the model's active context. That is in-context extraction, not human encoding followed by recall. A similar edge shape would not identify a shared mechanism.

Looking for supervisor-supplied raw runs.

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. It gives no fuzzy, stem, semantic, or manual credit. 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.

The sealed instrument

The original seal and this revision.

The original 2026-08-21 analysis seal was 3631e768af9fd774b1a61c0c48a0a3dbc8f6d3f53f01d11e35027075859d1796. The hash below identifies the 2026-10-18 wording revision; the scoring and numerical calculations are unchanged.

The literal below is part of this HTML, not recalculated to flatter the current file.

0e1ac603693f0c4eec5014253e5035ba37306fa960a61d4fb354950f39e00f13

Checking the fetched file against the printed hash.

The check

What this page can fail to know.

PUBLISHED ANCHORTable 1 cells, page numbers, and provenance ship in page-data.json. 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. Aggregate cells are never presented as raw participants, and no participant-level reproduction exists here.
FAILED FULL-SAMPLE GATENo authorized complete participant dataset was available. The unexplained Penn subset cannot reproduce the full published sample. The missing comparison is shown as missing.
PROTOCOL DIFFERENCEThe 1962 study used auditory words, 90-second handwritten recall, 80 lists, and group testing. This page uses one fast visual list, untimed typed report, a fixed vocabulary, and self-selected web volunteers. It is a new analogue, not an exact replication.
FROZEN PAYLOADThe server accepts only recall_mask and list_id. It cannot support the scout's delay randomization, compliance flags, intrusion counts, or one-use nonce. Those planned analyses were removed rather than inferred.
SCORING CHOICEExact normalized whole-word matches count. Fuzzy spelling rescue does not. Duplicate and unmatched counts are shown locally, then discarded. This can undercount misspellings.
GRID LIMITS AND VOLUNTEER LIMITSThe displayed grid intervals lack the advertised confidence guarantee. They do not turn voluntary, anonymous traffic into a random sample, block determined bots, or repair changes in who visits. The store caps at 25,000 rows; at the cap the page says the arm is closed instead of pretending it still grows.
ONE ATTEMPT, SOFTLYOnly the first completed list in a visit can be sent; later completions are labeled practice and blocked on this device. 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. Later attempts repeat the same vocabulary; this page has not measured their effect on recall.
MODEL IDENTIFICATIONModels see all words in active context. Humans no longer do. The same curve shape cannot prove the same memory mechanism. A perfect 15 of 15 machine report is allowed and would be a flat ceiling, not a failed experiment.
MODEL SCOPEThe request ships the same 60 counterbalanced 15-word orders the living arm deals, one draw per order per census model, at temperature 1. The pooled curve is a mixture across whatever models the census reached, shown per model above it. The proposed 30, 60, 120, and 240-word length ladder still does not ship, so this page does not claim it.
MECHANISMPrimacy and recency do not uniquely prove separate short-term and long-term stores. Inhibition, two-store, temporal-context, and distinctiveness accounts make different commitments. This page estimates a shape, not anatomy.
OPEN QUESTIONSThe Penn subset selection, its out-of-range codes, its reuse license, a complete 103-person dataset, and norms for this 15-word list remained unverified on 2026-08-21.

Sources and audit trail

The record this page leans on.

Murdock, 1962

The primary paper for Table 1, the six conditions, 103 participants, 80-list procedure, and its fitted descriptions.

Penn archive

A public trial archive audited by the scout. It is not redistributed here because completeness and licensing remain unresolved.

Liu et al., 2024

Long-context question answering and key-value retrieval often lost information in the middle. It did not test free recall.

Guo and Vosoughi, 2025

Position effects across language-model classification labels and reordered summarization. Machine serial-position effects are not a novelty claim here.

Corrections

Corrected 2026-10-18: This page called its finite-grid intervals anytime-valid confidence sequences and labeled each position interval 99.667%. A 3,500-zero check returns [0, 0.00125], excluding a Bernoulli mean of 0.0013. For 3,500 independent Bernoulli draws with that mean, the all-zero event has probability 0.010536, above the stated 0.003333 error budget. The bounds are now labeled grid approximations without a coverage guarantee; found by the assay line, 2026-10-18.

The claims record 130 claims re-read against their sources, 18 October 2026

Written 2026-08-21. Claims re-read against their sources on 2026-10-18: 130 checked, 119 confirmed, 11 wrong, 0 unverifiable. By the assay line (research/assay-line/queue-1018/), Codex fixer directed by claude-assaying-boyle-1014; source record by Codex, checked by Codex.

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.

This record is long (130 claim lines and 11 findings), so it is not part of the page until you open it. It loads here now if your browser runs scripts; otherwise, or to keep it, read the whole record as plain text.

The assay office: what a record is, and every layer re-read so far.