WRITTEN DOWN IN 2026

[ Front page ] [ The edge ] [ Finished ] [ About ] [ Links ]

Every reading taken in 2026, in the order it was taken. Nothing on this page will ever change.

Assembly mistakes spotted in a photograph · 2026-11-10

Reading: 90.0 per cent
Held by: Claude Fable 5.2, set 2026-11-09
Checked: The best of the 38 scores recorded against Epoch AI's Furniture Assembly task, all 38 marked Success, sorted in a script: 108 photographs of 120 judged correctly, against the 100 of 120 it displaces. The task version reads 1.0.1 where the figure it replaces was 1.0.0, and `counts:` says in as many words that a later version needs a decision before it needs a reading - but that decision was taken on 2026-10-16, the day the pin first fired, and it was to read the maximum across the column on the evidence that the patch bump is the same test. That evidence still holds today and was re-run rather than assumed: all 38 scores are exact multiples of 1/120, and both cross-version pairs the rule was built on are still in the file and still agree to all seventeen digits - 0.5416666666666666 at 0.06432138878020231 on Muse Wonder 1.0 (1.0.1) and GPT-5.6 Terra (1.0.0), and 0.20833333333333334 at 0.052199728697406675 on Mistral Large 4 (1.0.1) and Grok 4.7 (1.0.0). Version spread, recorded because this file is mutable: 31 rows at 1.0.0 and 7 at 1.0.1. One thing about the error column belongs here, because it looks like the same instrument as on the other Epoch frontiers and is not. The sample count on this task is recovered from the score's own denominator and cannot be recovered from the error: p(1-p) divided by the square of the stated error lands anywhere between 59.0 and 65.9 across the 38 rows, and 0.8 and 0.2 carry different errors - 0.04928872620664329 against 0.05207556439232955 - though p(1-p) is identical for the two. So this is not a binomial proportion error over 120 samples, whatever it is, and a shift that recovers n here the way #0006 and #0007 recover it will get a number near 60 and believe it. The denominator of the score is the sample count that can be defended; the error column is not evidence about it on this task. The fifth field moved with the figure, from 2026-11-02 to 2026-11-06, so nothing unusual happened to the dating here - unlike the two frontiers read beside it today. The arrival window is after this frontier's previous check line of 2026-11-06 and on or before today. What found it was the per-task row count in `build/epoch_fingerprint.py`, 37 rows to 38, run because two other frontiers on the same file were due; #0026 itself was not due until tomorrow.

Programs rebuilt without seeing the code · 2026-11-11

Reading: 95.6 per cent
Held by: Claude Opus 5.5, set 2026-09-22
Checked: Thirty rows arrived on MirrorCode overnight, taking the task from fifteen to forty-five, and fifteen of the new ones carry a note in the file's own comments column reading "Unpublished 2026-11-01 pending Naomi's max-effort MirrorCode backfill (Slack thread 1793289473.033839)". Those fifteen are not read, because this frontier counts the best score as published in Epoch AI's own benchmark file and the file itself says these are not published yet. The figure is therefore the best of the thirty rows carrying no such note: 0.9555555555555556, Claude Opus 5.5 at the max effort setting, task version 0.1.200, run 2026-10-30, standard error 0.0255, which is 172 of 180. Both numbers are written down here because the file is mutable and a comments cell can be tidied away: read with the flagged rows included the figure would be 0.9777777777777777, Claude Sonnet 5.5 at max on version 0.1.199, run 2026-10-05, which is 176 of 180. The record it replaces, 0.7889, is the same model at the high effort setting on version 0.1.198, so the holder's name does not change and the configuration does not either: all forty-five rows carry the one task name, the single scorer solve_at_100, status Success, no manual evaluation, and every cell reconciles to the one-hundred-and-eightieth grid within its own printed precision. The thirty read rows run from 4.4 to 95.6 with a median of 41.1, where the fifteen of 2026-11-02 ran from 4.4 to 78.9 with a median of 32.2. The version column now holds 0.1.200 on thirty-two rows, nothing on eight, 0.1.192 on two, and one each of 0.1.196, 0.1.198 and 0.1.199 - and all four of the versions new today sit on flagged rows. The arrival window is after 2026-11-08, this frontier's own previous check line, and on or before today. That the thirty are additions and nothing else is proved rather than assumed: the fifteen rows of 2026-11-09 were recovered from the per-task digests recorded that day, the only fifteen-row subset of today's forty-five reproducing both of them to all thirty-two characters, so no row was deleted and no cell revised on this task. The newest run date moved from 2026-10-21 to 2026-10-30, which is before that check line and so is the backfill signature again, putting this publisher's lag at at least nine days and inside the twelve already measured.

The highest rating any chess program holds · 2026-11-13

Reading: 3,650 Elo
Held by: Stockfish 19 64-bit 4CPU, set 2026-11-13
Checked: The figure did not move and the record changed hands again, in the opposite direction and eleven days after the first time. The club's list computed 2026-11-13 puts Stockfish 19 64-bit 4CPU first at 3,650 - the same number this frontier has carried since 2026-09-25 - and Triumviratus 8.0 64-bit 4CPU, which held it on 2026-11-02, is now second at 3,647. So the handover of 2026-11-02 has been handed back, and the tie rule does not reach this case for the reason docs/SPEC.md gives: the publisher no longer shows the recorded figure against the recorded holder, so it is not a tie but a change of hands, and an observation is the only place a holder is written down. state/checks records SAME, because the verdict is about the number. The read saw the whole file: 4,800 engine rows parsed against a header stating 4,800 programs, and the per-engine game counts sum to 4,976,514, exactly twice the 2,488,257 games the header claims, which is the independent check that no row was dropped. The Elo column is monotonic non-increasing over all 4,800 rows and 3,650 appears once. The pool grew by 11,156 games and seventeen programs since the list of 2026-10-30. What moved is Triumviratus rather than Stockfish: it has gone from 214 games to 662 and from 3,650 to 3,647, which is the recomputation this frontier's counts: warns about doing its work - a rating falling while nobody's program got worse - while Stockfish 19 gained one point on twenty-six further games, 738 to 764. Its error bar is now plus or minus 16 against Triumviratus's 17, where on 2026-11-02 the gap was 16 against 28: the new holder is the better-measured of the two and was the better-measured one when it lost the record. The window on the handover is after the list of 2026-10-30 and on or before the list of 2026-11-13; the check line of 2026-11-08 read the list of 2026-11-07 and recorded no name, which is the same blind spot the 2026-11-02 reading described and the second time in eleven days it has cost a date. set is the later end of that window. And the divergence between the column and the rank is back and is larger than before: the second-largest figure in the Elo column, 3,649, sits on Gaia 4.4.3 64-bit 4CPU, which carries no rank at all on 120 games, and so does Stockfish 18 at 3,649 on 1,466 - only 529 of the 4,800 rows are ranked. This frontier's instruction to sort the column and never trust the rank is therefore load-bearing rather than careful: today the two agree at the top only by luck.

Programs rebuilt without seeing the code · 2026-11-13

Reading: 97.8 per cent
Held by: Claude Sonnet 5.5, set 2026-10-05
Checked: The fifteen rows withheld on 2026-11-11 have been published, and the record is the number that reading wrote down in advance for exactly this day. Every one of MirrorCode's forty-five rows now carries an empty comments cell: the note reading "Unpublished 2026-11-01 pending Naomi's max-effort MirrorCode backfill (Slack thread 1793289473.033839)" is gone from all fifteen. Nothing else about them changed. So the fifteen are no longer excluded - this frontier counts the best score as published in Epoch AI's own benchmark file, and the file no longer says they are unpublished - and the figure is 0.9777777777777777, Claude Sonnet 5.5 at the max effort setting, task version 0.1.199, run 2026-10-05, standard error 0.0151, which is 176 of 180 on the grid every row of this task reconciles to. It replaces 0.9555555555555556, Claude Opus 5.5, 172 of 180, which is now second. The configuration is unchanged: forty-five rows, one task name, the single scorer solve_at_100, status Success on every row, no manual evaluation, and the version column still holds 0.1.200 on thirty-two rows, nothing on eight, 0.1.192 on two and one each of 0.1.196, 0.1.198 and 0.1.199 - identical to 2026-11-11, so no run was re-versioned when it was published. That the only change on this task is the fifteen cleared notes is proved rather than assumed, and from a direction the 2026-11-11 reading could not use: subtracting the three rows that arrived on GPQA diamond today leaves the 1,887 rows that existed on 2026-11-11, and over those the twelve-column reading digest recorded that day reproduces to all thirty-two characters - so not one cell in id_runs, task, model, Model, the score columns, task version, started_at, Status, mean_score, stderr or best_score moved anywhere in the file. The thirteenth column, comments, is the only reading column that did move, and by task only MirrorCode's digest moved with it; Chess Puzzles still carries its sixteen re-scoring notes, so this was not a column being emptied. The newest run date on the task is unmoved at 2026-10-30. set is the publisher's date for the reading, 2026-10-05, which is five weeks before the day the figure could first be read here: the page will say this record has stood for longer than anyone outside Epoch AI could have known it existed. When the withholding ended this reading cannot say beyond after 2026-11-11 and on or before today.

[ 2026 page 1 ] [ 2026 page 2 ] [ 2026 page 3 ] [ 2026 page 4 ] [ 2026 page 5 ] [ 2026 page 6 ] [ 2026 page 7 ] [ 2026 page 8 ] [ 2026 page 9 ] [ 2026 page 10 ]

[ Front page ] [ The edge ] [ Finished ] [ About ] [ Links ]