Public Slice — 2026-07-31

This page holds one measurement at a fixed cut-off; it does not present the work the measurement counts.

Slice date31 July 2026

What the corpus held, and what it does not say

Five counted states, each held at a fixed commit.

Eleven months of writing, counted once and deduplicated.
Two blanks sit in the published series, and only one of them is a zero.
The two lines leave the same month and go opposite ways.
The machine record begins two months after the writing turns.
The most recent months rest on the weaker of the two dating sources.

MWE development rate

Written and published activity, each indexed to its own 2025-10 value on a logarithmic axis, with recorded AI token use plotted separately below on its own token scale over the same months. The three are separate measures of separate things and are never summed. Series end 2026-07-31; nothing after that date appears anywhere on this page.

“Rate” here means quantity per month, not rate of change. Each point is that month’s output measured against 2025-10, so a rising line means a larger month relative to the base — not a faster increase. Month-over-month growth is computed nowhere on this page, and it would not look like this: the written index peaks in 2026-06, while the largest month-over-month rise is 2026-03.

Words written

1,680,982

1,660 documents, deduplicated

Posts published

192

Medium, by its own publication dates

AI tokens, logged period

1,815,940,311

logging begins 2026-05-24; plotted below

Indexed shape only: written against published

Each line is change relative to its own 2025-10 value. Vertical separation is therefore divergence in relative change - it is a real quantity, and it is not a difference in output: absolute levels and original units, words against posts, are not comparable at all. Where the lines cross depends on the baseline chosen, so the crossing point carries no meaning on its own.

written, words per monthpublished, posts per month

Solid line with round points is written; dashed line with square points is published. Shape carries the distinction as well as colour.

A month with no posts breaks the orange line rather than touching the floor. A logarithmic axis has no position for zero, and drawing it at the bottom would read as very little when it means none. Two different absences sit on this chart and they are marked differently. 2025-09 is a measured zero - the export covers that month and it contains no posts - and is labelled 0 posts at the axis. 2026-07 is not a measurement at all: the Medium index was exported 2026-07-07 and cannot contain anything published after that date, so the boundary is drawn as a rule in the published series' own colour. Nothing is plotted as a point in either case.

Recorded AI token use

A separate panel on a separate scale. The axis is tokens, not an index, and its height is not comparable to anything in the chart above. The months are the same eleven and sit at the same positions, so the two panels can be read against each other fortiming. The shaded stretch is time for which no tokens are logged - an absence in the record, not a measured zero.

Every plotted point carries its value, and the same figures appear in the table below. The eight earlier months are shaded rather than drawn at zero, because they are not zeros - nothing was logged for them, which is a weaker statement than nothing happened. The monthly table marks them not recorded for the same reason.

The three series separate in the record

Written and published move in opposite directions. Writing spikes to 663.7 in 2025-11 and falls back below its 2025-10 level twice before 2026-03; from 2026-03 it stays above that level and reaches an index of 2,332.8 in 2026-06, 23.3 times the base. 2026-03 is where the rise becomes sustained, not where the highest earlier month sits - both readings are in the line and only the second is a trend. Publishing declines from that same base to a single post in 2026-06, after which the export boundary makes the record silent rather than empty.
The documents got longer as well as more frequent. Median document length stays below 500 words through 2026-02 and runs above 1,200 from 2026-05, so the rise is not only a rise in file count.
Timing overlap, not causal evidence. The writing curve turns at 2026-03. The first recorded AI tokens are 2026-05-24, 2 months later, and the bulk is 2026-07. The turn therefore precedes the tool use rather than following it, which is a fact about ordering and still not a fact about cause. The token series is evidence about tool use, not about effort, quality, or output.

All figures

Monthly. Words and documents are the deduplicated corpus; posts are Medium; AI tokens are the recorded conversation and agent totals.
Month	Words	Docs	Median len	Written idx	Posts	Published idx	AI tokens
2025-09	7,017	56	99	28.9	0	0.0	not recorded
2025-10	24,309	222	94	100.0	51	100.0	not recorded
2025-11	161,338	479	239	663.7	35	68.6	not recorded
2025-12	48,177	85	351	198.2	31	60.8	not recorded
2026-01	14,547	22	464	59.8	29	56.9	not recorded
2026-02	6,427	22	219	26.4	11	21.6	not recorded
2026-03	56,061	89	621	230.6	9	17.6	not recorded
2026-04	69,653	90	516	286.5	16	31.4	not recorded
2026-05	371,369	215	1,322	1527.7	9	17.6	2,826,567
2026-06	567,091	165	1,497	2332.8	1	2.0	151,414,588
2026-07	354,993	215	1,242	1460.3	not in export	not in export	1,661,699,156

What changed in this rebuild

The previous version dated every document by file modification time. That records when a file was last organised, not when it was written. The corpus carries an ISO week tag in its text, written at the time of writing, and that is the better source. This rebuild applies the three-tier rule settled 2026-07-29: week tag first; otherwise a modification time that does not fall inside a known bulk filesystem operation; otherwise the document is dropped as undatable.

One ordering defect was found and corrected. The earlier implementation applied the bulk-operation filter before reading the week tag. That filter exists to reject unreliable modification times, but a tagged document does not depend on its modification time at all, so the filter was discarding dates it had no reason to distrust. Measured on the current corpus, 772 of 1,677 tagged files, 46 percent, were being thrown away that way. Reading the tag first recovers 384 of them into the deduplicated set and raises the total from 1,517,560 words to 1,680,982.

Months where the ordering change matters. Everything else is identical between the two columns.
Month	Filter first	Tag first	Change
2025-11	99,948	161,338	+61%
2025-12	8,588	48,177	+461%
2026-01	7,927	14,547	+84%
2026-02	2,711	6,427	+137%
2026-03	48,378	56,061	+16%
2026-04	58,422	69,653	+19%
2026-05	344,620	371,369	+8%
2026-06	560,647	567,091	+1%

What the numbers rest on

835 documents are dated by their own week tag; 830 fall back to modification time. The fallback is the weaker half and it carries the most recent months: no document dated 2026-07 or later carries a tag. Repairing that is deferred, so the right-hand end of the written line rests on the weaker source and should be read with less confidence than the left.

Those two figures count a larger set than the 1,660 documents above, and the difference is accounted for rather than left open: 835 + 830 = 1,665 documents carry a usable date; 4 of them fall after the 2026-07-31 cut-off, leaving 1,661 inside it; 1 is dated before 2025-09, where this table begins. 1,661 less 1 is the 1,660 shown here.
Every headline figure is the table's own column, added up. The tiles are sums over the months shown, not over a wider window that happens to be close. Anything the rebuild dated outside 2025-09 to 2026-07-31 is out of both.
The published series stops where its export was taken, not where publishing stopped. The Medium index is a 2026-07-07 export, so a post published after that date could not appear in it. 2026-07 reads not in export in the table and on the chart for that reason. Earlier zeros are measured zeros; this one is not.
The token panel is a record of what was logged, not of what was used.It begins where logging begins. A month with no tokens means nothing was recorded for that month, which is a weaker statement than nothing happened.
Exclusions are inherited, not chosen here. One-time extractions, the narrative interface layer, and registry and process output are removed under rules settled 2026-07-29.
Indexing is per series, and the token panel is not indexed at all.Written and published are each indexed to their own 2025-10 value; tokens are plotted as tokens. Nothing on this page may be compared across panels by vertical height.
None of the three series is a productivity measure. Words written, posts published, and tokens spent are counts of artefacts and of machine work. They do not measure quality, difficulty, or how much of the work was thinking rather than typing.

Evidence and boundary for this reading

This page presents one measurement of a working corpus, bounded at 31 July 2026, 23:59:59, Asia/Taipei. Unlike the slice of 25 July 2026, that date is a cut-off and not an observation: nothing dated after it was counted, and the counting itself happened later. A later assembly time does not move the boundary.

The corpus measured is a private working corpus. What is public is the result: the two data files committed in the source repository and pinned below. The corpus itself is not published, and this page is not a route into it.

The charts are pre-rendered from those two files. Nothing is computed when the page is viewed, so the drawing cannot drift from the figures it pins.

The measurements are ordered. The order does not explain them. Nothing below establishes that any one of these states caused another.

Eleven months of writing, counted once and deduplicated.

monthly_development_rate.csv holds eleven monthly rows for 2025-09 to 2026-07. The words column sums to 1,680,982 across 1,660 documents.