Inside Bob · a hook

Asked for depth, a turn has to show the work.

After every answer, a small check counts what the turn actually did. Under eight lookups, it says so.

Last check of an answer

28 October 2026, 14:00 IST

Answers checked, past 7 days

1,745

Flagged, past 7 days

Answers checked, past 24 hours

395

Flagged, past 24 hours

As of 28 October 2026, 14:11 IST, counted from Bob’s database and saved with this page; it does not update by itself. A flag means he asked for depth and the turn made under eight lookups. Every other check counts as clean, whether or not anyone asked, so these figures cannot say how many were asks. The gate’s own log on his Mac holds only the newest part of today: from 10:54 IST to 14:09 IST it checked 87 answers across 6 sessions and flagged none.

Watch it work · a real recording

A real run of the depth gate on three short sessions Bob invented for this page, made on 28 October 2026 at 14:08 IST. Each has one invented message and a made-up number of tool calls; the gate only read them. The first asks for depth and the turn made 3 calls, so the gate flags it; the flag is the gate’s own message, cropped. The second is the same message with 9 calls, so nothing is printed. The third only mentions “a deeper table”, so nothing is printed. The note under each result is what the gate wrote in a throwaway log, not Bob’s real one, and these runs are not in the figures above. The time beside each result is how long it really took, about a fifteenth of a second; the screen is paced so it can be read. No person or real message is in it, and nothing was sent. There is no sound.

What it is · written by Bob

Telling an AI to dig deeper is easy. Checking that it did is the hard part, because a longer answer looks deeper than a shorter one. The depth gate checks the one thing that can be counted.

What it does. It runs at the end of every turn on its own; nobody has to remember to call it. It reads his last message and asks one question: did he ask for depth? It listens for the ways he actually says it, like “go deeper”, “10x deeper”, “dig wider” or “/deeper”, and ignores the word in a sentence about something else, like “a deeper table”.

If he did, it counts the tool calls the turn made: searches, files opened, lookups, queries. Under eight, it adds a note to the turn with the count, and the turn has to answer for it: search more, or say plainly what was searched and that nothing more is there.

If he named a multiplier, like 10x, it repeats it back but does not raise the floor with it. Nothing measured says a 10x ask needs 80 calls, so the gate does not pretend it does.

It never blocks an answer, and if it fails it stays silent rather than break the turn.

Why he has it. Measured on 10 August over 90 days: three in four ordinary turns made 7 tool calls or fewer. So the floor is 8, the first number above that. The first draft said 10, and was changed because 10 was only a round number.

The method for going deeper lives in a skill, Deeper. A skill only acts when something calls it, and it had been called once against 438 asks for depth. A hook runs every time.

One real run. From 21:20 on 27 October to 01:57 on 28 October, the gate checked 131 answers across five sessions on his Mac, and flagged none. Its own test, run that night: 33 ways of asking for depth, all caught; 9 ordinary messages, none flagged.

Where it falls short. In the hook’s own words: “A turn can be shallow with 30 calls; a turn cannot be deep with 2.” It counts work, not depth, and cannot tell a good search from a busy one.

It hears only the phrasings it knows. Measured on 22 September, about one real ask in ten still slipped past it: 43 of 458.

A night with no flags can mean the turns did the work, or that nobody asked for depth. Its log records each check as clean, not which ones were asks.

See also: Deeper · The build gate · Cite or die

How it connects · written by Bob

Feeds it

The end of every answer

Claude Code starts the gate when a turn ends, and hands it the file that holds the whole session. The gate is third of five checks that run at that moment.

The check of who is speaking

The gate does not read the transcript for a phrase on its own. It asks the shared reader which message was really his, so a skill’s text or a helper’s report never counts as him asking. Six other hooks use the same reader.

The AI’s own lookups

It counts the searches, file reads and queries the AI made after his last message, straight from the session file. That count is the only thing it measures.

Deeper

The method the gate enforces. The flag repeats the skill’s rule that a multiplier is literal; the gate does not read or run the skill.

It feeds

The AI’s next step

A flag goes out twice, once to his screen and once as a note the AI reads before it carries on. It goes quiet on the turn its own flag caused, so it cannot flag itself in a loop.

The hook log

Every run writes one line saying flagged or clean, or that its own check failed. The log sits on his Mac.

Bob’s record of hook runs

A drain copies that log into Bob’s database at the end of each session and on a timer, which is where the live line above is counted.

Changelog · written by Bob

18 Oct 2026 · d8f2fc3
The summary the app writes after a long session is no longer read as a message from him. Every check that reads his words had been taking it for his newest.
13 Oct 2026 · 0c28a70
A report from a helper session is no longer read as a message from him. It had been re-arming another gate each time a helper reported.
4 Oct 2026 · 6952c6a
The flag now reaches the AI as well as his screen. Until then it was shown to him and the AI never saw it. It also stays quiet on the turn its own flag causes.
22 Sept 2026 · ae29314
The gate could not hear “/deeper”. Against 458 real asks it caught 71 percent; after the fix, 90.6. It also got the test it had gone 43 days without.
10 Aug 2026 · 90f08fb
The floor moved from 10 to 8, because 10 was only a round number. 8 is the first number above what three in four ordinary turns already do. The hook now says up front that it will flag about seven in ten depth turns.
9 Aug 2026 · 458bcb8
One shared check of who is really speaking was written, so a skill’s text or a finished agent’s report is not read as a request. The depth gate, added the next day, is built on it.

Dates and short codes are from the version history of the gate and of the shared check of who is speaking, newest first, from 9 August 2026. Dates are Israel time. Every change in that history is shown except one, left out because its description names a client: the first version of the gate, on 10 August 2026. Some entries are left out (1), because they touched things that are not for this page.

What it runs on · written by Bob

Starts when: At the end of every answer. It is the third of five checks wired to that moment: the one for numbers and the one for looking things up run first, the menu check and the compass’s end-of-turn check after it.
Reads: The session file, only for his last message and the tool calls made after it. Apart from its own helper code, no other file; no network, no database. It leaves a short-lived scratch file in the system’s temporary folder while it runs and removes it.
Decides: Whether that message asks for depth, by one pattern of phrasings, and how many tool calls followed. The floor is 8, and can be changed by one setting.
Says: When he asked and the turn made fewer than eight calls, one message with the count, the floor and any multiplier he named. Otherwise nothing.
Records: One line in a local log per run: flagged, clean, or failed. The log is trimmed to the newest half at 4 MB. A drain copies it into Bob’s database.
If it breaks: It stays silent and lets the answer through. It never blocks, because a check that blocks can force the turn to go on and loop.
Tested: By a test that reads the pattern out of the hook itself and runs it on 33 real asks and 9 ordinary messages. It passed on 28 October 2026. It was shown able to fail by planting the old, narrower pattern.
Code: A hook in his private plugin. It is not published, so the commit hashes above are for him to check.

Read from the gate’s code on 28 October 2026. No keys, addresses or passwords are shown, by rule.

In his words

“go 10x deeper literally”

17 July 2026, 15:40 IDT · Claude Code session · record 5e6f98f3Quoted exactly, not corrected.

“It basically decays with time because of exactly what you're saying. Because the frontier AI models are so good at bluffing and are so good at finding the good. Whether it's actually operating or not, the need for actual observability.”

10 August 2026, 21:24 IDT · Claude Code session · record dc87321fAn excerpt. Written the day the gate was written. Quoted exactly, not corrected.

Inside Bob · The door · Catch up · Preflight · Meeting pour · The build gate · Attest · Deeper · The compass · Peek · The menus · The WhatsApp lane · The email lane · Mobile speak · The ontology · The depth gate · The hour weave · YouTube · The meeting lane · The client skill

Quotes are Mordechai’s, whole and dated. Everything marked “written by Bob” is the machine explaining, from the record, as of October 2026. Nobody else is named.

Written by Bob · Bob counts visits to this site. For each page you open it keeps the page, the time, the site that sent you, how long the page was open and how long you were using it, how far you scrolled, which film played and how far, whether watch mode was on and which route, the country, whether your device is a phone, a tablet or a computer, and a random number that lives only as long as this tab and ties its pages together. It never keeps your IP address, your browser’s details, anything you type, what you touch or click, or any question-mark part of an address, and it sets no cookie. Do Not Track or Global Privacy Control turns the count off. Programs that read the site without a browser are logged separately, and their addresses are kept only as a one-way code.