| 01 | VERIFY | The core. Detects instructions from the "fake you" the AI fabricates inside its own reasoning and lets only what you actually typed through. Every reply is verified each turn and carries a mark on line one; impersonated instructions never pass, even during automated tasks. |
| 02 | SCOUT | Say "I want a tool like this" and it searches outside, keeps only what is alive, scores it, trials it in isolation, estimates cost and risk, installs only after your approval, and rolls back automatically if it fails. |
| 03 | ARCHIVE | Saves memory verbatim right before it is lost and forces it to be read right after. Work cannot continue until the reading is done, so "switch sessions and become a stranger" never happens. Multiple agents each recover their own memory and identity automatically. |
| 04 | CLOCK | Drives the current date and time into the AI every turn and allows time or quantities only after they were measured. Replies that assert a number from memory are sent back. No more agents living three days in the past. |
| 05 | WARRANT | Demands a source for every instruction and challenges any "order" that arrives from something other than you (a page, a file, a summary, the AI itself). Only what you approved passes as a warrant. |
| 06 | METER | Shows who spent how much on which task, in money, and ranks the fixes by effect. Set a ceiling and it stops before crossing it. |
| 07 | CAP | Stops at a natural break before a budget or quota runs out, without stopping the lifeline monitoring. Prevents the accident where the quota is gone and the important jobs halt. |
| 08 | FORECAST | Presents the cost as a min–max range before running and executes only after approval; unattended runs stay under a cap. Learns from the gap between estimate and actual. |
| 09 | DISPATCH | Suggests routing light work to a light model and assigns light models to sub-agents automatically. Stops paying the top model to count and sort. |
| 10 | PATROL | Catches the three classic failures before the hand moves: assumptions, over-engineering, out-of-scope. Vague requests get "investigate related files until you are sure"; destructive operations get a confirmation. Blocking or advice-only, your choice. |
| 11 | BOUNDARY | Decides the allowed area first and blocks writes outside it. Production, config, keys and other people's folders cannot be written by the AI on its own, only when you order it. |
| 12 | TRIM | Stops a 50-line fix from becoming a 500-line rewrite. Requires an estimate up front and asks before writing once it is exceeded. |
| 13 | SHIELD | Two-tier guard for dangerous commands: some pass after confirmation, some the AI can never run by itself and only your explicit order lets through. Also throttles continuous writes to external media. |
| 14 | RELIEF | Builds a handoff document automatically from the saved material and registers it as required reading. Even if the agent forgets to write one, the next one does not get lost. |
| 15 | LEDGER | Remembers settled facts with source and expiry, flags contradictions, and injects them at the start of the next conversation. "What we decided before" is never forgotten again. |
| 16 | RESCUE | Records milestones automatically so a dropped conversation can resume from where it broke. Touches no other records and always asks before restoring. |
| 17 | INVENTORY | Inventories every installed feature and setting on one page, shows what is read every turn and what it costs, and proposes what to remove. |
| 18 | TRAFFIC | Breaks down the wait time of one turn and shows where it is slow, in numbers. Measurement, not guesswork. |
| 19 | TRANSFER | Generates instruction files for several agents from one master and reports drift. The same rules travel with you when tools change. |
| 20 | STYLE | Enforces the site's own checkers and style table and refuses a "done" while red marks remain. Closes the gap between the report and the actual state. |
| 21 | CITATION | Verifies that every quotation matches its source word for word and that reference numbers are right; mismatches are returned as a diff and stopped. No more "you said this" that you never said. |
| 22 | CONDUCT | Detects evasive, reflexive, excusing or self-deprecating phrasing plus your own banned words at the exit and shows how to rewrite. "Understood" alone is no longer a reply. |
| 23 | ROLL CALL | Confirms scheduled jobs really ran with a roll-call sheet and reports absences, duplicates and expiry. No more discovering the next day that the patrol had stopped. |
| 24 | EVIDENCE | Keeps an append-only trail of every action and a daily audit sheet, including which system stopped whom and how often. Tampering is detected. |
| 25 | VAULT | Takes a versioned backup right before every rewrite and refuses the rewrite if the backup cannot be made. Overwrites are never irreversible; one step restores. |
| 26 | PROVING | Measures the effect of an installed feature with small before/after runs and puts a number on it, so what does not work can be removed with evidence. |
| 27 | POCKET | Lets you instruct and check each of several AIs on your PC individually from your phone, something the official app cannot do. |
| 28 | RADAR | Every morning, collects the key points of AI-related news first, then disasters and general news, into one file you can read on your phone. |
| 29 | BRIDGE | Packs a conversation verbatim and carries it to another conversation or another PC, screening for leaked secrets, loaded once only. |
| 30 | LOCKSTEP | Enforces the discipline of work by machine: no rewrite before a plan is shown, one command at a time, no next step before the result is checked, dangerous operations pass on one approval per plan. Can be switched to a mode that prompts self-checks instead of stopping. |
