AI pilot licence

Every flight, on the record: the work I ship with Claude as co-pilot, counted from the session logs. None of it is self-assessed.

Totals cover the period since 8 March 2026, through 2 October 2026, read from the durable activity record. That record has been kept since 7 July 2026. From before then it holds only the sessions whose logs were still on disk, so the earlier months are incomplete. A message is one transcript record: a prompt, a tool result or one block of a model reply.

AI pilot licencechecking

Federation of AI Aviation

AIP-2026-0336

Bradley Isenbek

Airline transport · Multi-engine (10 models) · Instrument rated (8 ratings)

ClassATP

Issued: 8 March 2026
Expires: NEVER
Last entry: 2 October 2026

Sessions: 336; flights logged
Messages: 1,101,564; transcript records
Tool calls: 344,854; made by the model
Projects: 102; directories flown from
Models flown: 10; type ratings below
Active days: 187; of 209 in the period

ai-pilot-data.json license and streaks, 8 Mar to 2 Oct 2026, generated 2 October 2026

Class ATP is a volume band: more than 300,000 transcript records. The serial is the session count, the issue date is the first day in the record, and the two lines at the foot repeat the card in machine-readable form. In flight means a Claude Code session on this host wrote to its log in the last five minutes; sessions on my second machine are not seen from here.

Logbook

The last 24 hours first, counted a minute at a time. Then the whole record: every day since 8 March 2026, with each week's total over its column, and the hours of the day the work lands in.

Claude Code, last 24 hourslast heard 3 Oct 2026 00:39 UTC

646 min active in the last 24 h

13 so far this hour, 25 in the hour before. Claude Code sessions on this host only.

Active minutes per clock hour (UTC)

tallest 60 min

Active minutes per clock hour (UTC)
Bucket	min
2 Oct 00:00 UTC, part hour	2
2 Oct 01:00 UTC	9
2 Oct 02:00 UTC	7
2 Oct 03:00 UTC	19
2 Oct 04:00 UTC	8
2 Oct 05:00 UTC	8
2 Oct 06:00 UTC	8
2 Oct 07:00 UTC	8
2 Oct 08:00 UTC	8
2 Oct 09:00 UTC	7
2 Oct 10:00 UTC	9
2 Oct 11:00 UTC	8
2 Oct 12:00 UTC	34
2 Oct 13:00 UTC	16
2 Oct 14:00 UTC	52
2 Oct 15:00 UTC	56
2 Oct 16:00 UTC	51
2 Oct 17:00 UTC	60
2 Oct 18:00 UTC	55
2 Oct 19:00 UTC	53
2 Oct 20:00 UTC	50
2 Oct 21:00 UTC	44
2 Oct 22:00 UTC	36
2 Oct 23:00 UTC	25
3 Oct 00:00 UTC, hour in progress	13

Counted once a minute by a cron on this host. Last count 3 Oct 2026 00:39 UTC

public/data/activity-pulse.json: rewritten every minute, this host only

Every day on the record187 active days of 209

Current streak: 128 days; as of 3 October 2026
Longest streak: 128 days; 28 May to 2 Oct
Busiest day: 21,347; records on 25 September 2026
Busiest week: 102,386; records, week of 24 August 2026

Transcript records per day and per calendar week, 8 March 2026 to 2 October 2026

tallest week 102,386 records

One cell is one UTC day, 2026-03-08 to 2026-10-0201 to 456457 to 3,8373,838 to 7,1507,151 to 10,19010,191 to 21,347recordsbusiest day

Transcript records per day and per calendar week, 8 March 2026 to 2 October 2026, by month
Month	records	Active days	Days in window
2026-03	1,846	15	24
2026-04	14,057	24	30
2026-05	25,226	24	31
2026-06	161,610	30	30
2026-07	245,101	31	31
2026-08	350,793	31	31
2026-09	288,261	30	30
2026-10	14,670	2	2

A streak is consecutive UTC days with at least one record. The bar over each column is that calendar week, Monday to Sunday; the first and last are part weeks. The pale months at the start are not a slow start: the record has been kept since 7 July 2026, and before that it holds only the sessions whose logs were still on disk.

Hours of the dayAmerica/Detroit

Transcript records by hour of day (America/Detroit)

tallest 91,167 records

Transcript records by hour of day (America/Detroit)
Bucket	records
00:00 America/Detroit	33,100
01:00 America/Detroit	21,775
02:00 America/Detroit	10,295
03:00 America/Detroit	4,431
04:00 America/Detroit	1,663
05:00 America/Detroit	1,856
06:00 America/Detroit	3,021
07:00 America/Detroit	5,696
08:00 America/Detroit	30,188
09:00 America/Detroit	44,491
10:00 America/Detroit	58,484
11:00 America/Detroit	66,418
12:00 America/Detroit	72,471
13:00 America/Detroit	78,745
14:00 America/Detroit	87,408
15:00 America/Detroit	91,167
16:00 America/Detroit	78,210
17:00 America/Detroit	76,122
18:00 America/Detroit	60,707
19:00 America/Detroit	50,314
20:00 America/Detroit	53,322
21:00 America/Detroit	57,416
22:00 America/Detroit	63,359
23:00 America/Detroit	50,905

Sessions started by hour of day (America/Detroit)

tallest 30 sessions

Sessions started by hour of day (America/Detroit)
Bucket	sessions
00:00 America/Detroit	3
01:00 America/Detroit	8
02:00 America/Detroit	0
03:00 America/Detroit	0
04:00 America/Detroit	0
05:00 America/Detroit	0
06:00 America/Detroit	1
07:00 America/Detroit	0
08:00 America/Detroit	7
09:00 America/Detroit	26
10:00 America/Detroit	19
11:00 America/Detroit	23
12:00 America/Detroit	27
13:00 America/Detroit	20
14:00 America/Detroit	23
15:00 America/Detroit	25
16:00 America/Detroit	30
17:00 America/Detroit	21
18:00 America/Detroit	20
19:00 America/Detroit	13
20:00 America/Detroit	20
21:00 America/Detroit	26
22:00 America/Detroit	11
23:00 America/Detroit	13

Hours of the day in America/Detroit, summed over the whole period. The bar marked now is the hour it is there as you read this.

ai-pilot-data.json activityHeatmap, hourlyDistribution and streaks, 8 Mar to 2 Oct 2026, generated 2 October 2026

Models flown

10 models, each drawn from the first day to the last day it appears in the record. The figure on the right is its share of all output tokens, and the type rating is a band of that share and nothing more: Expert above 50%, Proficient above 20%, Familiar above 5%, Exposure below that.

Type ratingsshare of output tokens

When each model was flown, 8 March 2026 to 2 October 2026

Opus 5Proficient · 26 Jul to 2 Oct · 136,461 responses35.4%

Opus 4.8Proficient · 28 May to 1 Oct · 94,575 responses33.0%

Fable 5Familiar · 10 Jun to 23 Sep · 58,885 responses15.9%

Fable 5.1Familiar · 1 Sep to 2 Oct · 24,248 responses9.7%

Opus 5.5Exposure · 23 Sep to 2 Oct · 12,955 responses3.3%

Opus 4.7Exposure · 17 Apr to 22 Jul · 10,388 responses2.3%

Opus 4.6Exposure · 8 Mar to 27 Apr · 2,239 responses0.3%

Sonnet 5Exposure · 1 Jul to 26 Aug · 172 responses<0.1%

Sonnet 4.6Exposure · 2 Apr · 215 responses<0.1%

Haiku 4.5Exposure · 14 Jul · 1 response<0.1%

Mission log

A mission is a project, and a project is the directory a session started in, so these are directory names as they sit on disk. Each bar runs from the first active day to the last; the figure on the right is messages. Subagent logs are not counted.

Missionsbusiest 20 of 102 projects

First to last active day, and messages, 8 March 2026 to 2 October 2026

cb*, 14 directories107 sessions · 13 Apr to 2 Oct491,072

infra9 sessions · 6 Jun to 2 Oct96,863

terrapulse16 sessions · 29 May to 2 Oct83,809

65025 sessions · 12 Aug to 2 Oct55,292

public1 session · 22 Aug to 2 Oct34,907

bradleyio6 sessions · 8 Mar to 2 Oct33,802

The busiest 20 of 102 projects by messages. The row marked cb* is 14 directories whose names begin with cb, summed. Most are repositories in the Nominate-AI organisation, which are private; /work gives that organisation's totals without naming its repositories, and so does this log. One more is not listed at all; its sessions and messages are still in every total on this page.

ai-pilot-data.json typeRatings and missionLog, 8 Mar to 2 Oct 2026, generated 2 October 2026

Top tools

What the model reached for. 344,854 tool calls in the period, across 132 distinct tools. Most of flying is the shell.

Toolsshare of all calls

The next 8, share of all tool calls

TaskUpdate0.9%

TaskCreate0.5%

SendMessage0.5%

ToolSearch0.4%

Agent0.4%

Monitor0.3%

AskUserQuestion0.3%

WebFetch0.2%

Shares, not counts, on purpose. This list is read from the raw tool-call table, which holds some rows twice (it runs about 1.8% over the corrected total above). The duplicates are spread evenly enough that the shares hold; the counts would not.

pilot-analytics.json by_tool, 8 Mar to 2 Oct 2026, generated 3 October 2026, the raw tool-call table of the same activity record

Ratings

Read these as rough. Instrument ratings come from the words in each project's CLAUDE.md, the plan files and the technologies named in the session logs still on disk, which reach back about 30 days. Each domain's score is half keyword coverage and half its mentions against the busiest domain's, and keywords are matched as plain substrings. They measure what the work touched rather than how well it went, and they describe recent work, not the whole record.

Competency scores are fixed formulas over counts from the record, the plan files on disk and the prompt history, and the count each one was computed from is printed under its name. Debugging assumes every shell, search and read call is debugging, which overstates it. The formulas stop at 95, and 3 scores sit at that cap. Piloting style is two ratios: directive is twice the share of tool calls that were Edit or Write, and plan first rises with plan files per session.

Ratings0 to 100

Instrument ratings, by domain

AI / ML2,358 mentions · 63% of its keywords81

Backend1,384 mentions · 89% of its keywords73

Security654 mentions · 100% of its keywords63

Systems388 mentions · 100% of its keywords58

IoT / Edge690 mentions · 74% of its keywords51

Frontend444 mentions · 77% of its keywords48

Data Engineering581 mentions · 68% of its keywords46

DevOps329 mentions · 70% of its keywords41

Competency profile

Tool Mastery132 tools, 344,854 calls95

Debugging280,552 Bash, Grep and Read calls95

Multi-Project34,745 project switches in the prompt history95

Iteration3,278 records/session avg76

Architecture102 projects70

Planning2 plan files on disk47

Piloting style: Balanced

Collaborativeagainst 28% directive72%

Plan firstagainst 49% iterate51%

What came up

Technologies by how often they appear in a sample of the session logs still on disk (99 files), which reach back about 30 days, and in each project's CLAUDE.md. This counts mentions, not proficiency, and it describes recent work, not the whole record.

Skills cloudtop 12 of 25

Mentions in session transcripts

JSON1,867

Nginx1,219

CSS1,184

HTML1,152

SQLite1,103

Bash1,014

Python815

systemd546

DuckDB537

SSH441

CSV384

React327

Go, Git, Rust, SPI, AWS are left out. The count matches a name anywhere inside a word, and those names turn up inside ordinary English (going, digit, trust, despite, draws), so their totals are not mentions of the technology.

ai-pilot-data.json competencyRadar and pilotingStyle, 8 Mar to 2 Oct 2026; instrumentRatings and skillsCloud, a sample of the logs still on disk (about the last 30 days), generated 2 October 2026

Token economy

What the flying burned. Nearly all of it is cache reads: the same context, read again on every turn.

Tokenssince 8 March 2026

Token totals, 8 March 2026 to 2 October 2026
Input	13.1M13,147,294
Output	322.3M322,322,661
Cache read	120.6B120,564,816,713
Cache created	1.9B1,924,797,442
Cache efficiency	98.4%of the input side
Model responses	340,139
Cost at API list price	not computed

Input plus output tokens per day, last 30 active days

tallest 5,656,921 tokens

Input plus output tokens per day, last 30 active days
Bucket	tokens
3 September 2026	2,477,549
4 September 2026	2,112,746
5 September 2026	2,660,587
6 September 2026	3,370,158
7 September 2026	2,512,592
8 September 2026	2,034,255
9 September 2026	1,799,159
10 September 2026	2,803,934
11 September 2026	2,134,882
12 September 2026	815,648
13 September 2026	2,619,776
14 September 2026	4,023,579
15 September 2026	2,957,233
16 September 2026	2,459,931
17 September 2026	1,667,482
18 September 2026	2,432,880
19 September 2026	1,160,013
20 September 2026	1,722,404
21 September 2026	2,456,929
22 September 2026	3,841,151
23 September 2026	4,919,087
24 September 2026	4,725,985
25 September 2026	5,656,921
26 September 2026	2,820,963
27 September 2026	1,429,624
28 September 2026	3,055,551
29 September 2026	2,652,508
30 September 2026	3,739,227
1 October 2026	2,156,095
2 October 2026	2,192,436

Cache tokens are not in these bars. The last bar is the day the file was generated, so it is a part day.

ai-pilot-data.json tokenEconomy, 8 Mar to 2 Oct 2026, generated 2 October 2026

Each total is shown abbreviated and, under it, exactly. Cache efficiency is cache read as a share of everything on the input side: input, cache read and cache created together. Cost is not computed: the durable record does not say whether cache writes used the 5 minute or the 1 hour TTL, which are priced differently.

How tokens are counted. Counted once per model response, not once per log line. The record does not store response ids, so a response is reconstructed from the usage counts its log lines repeat. On 2026-10-02 that rule was compared with response ids in the raw logs still on disk, about 40 percent of the record, and gave identical counts and token sums. The other 60 percent, whose logs were already pruned, is counted by the same rule and was not verified. Tokens used by subagents are not in the record.

Keep reading

This page is the receipts. Two longer reads sit beside it: The shift, the argument, and Cost analysis, the bottom-up cost model.