The Receipts.
Every claim on this site should be checkable. This page is where we keep the receipts: real journal entries, real run records, and real security logs from the agent systems that run this business every day.
Each card states the claim, tells you what actually happened in plain words, and then shows the raw record. Excerpts are trimmed for length and privacy, with cuts marked. Nothing is staged.
RECEIPT 01
The Claim: Work Happens On A Schedule, Without Prompting.
The engine that drafts this business's own content fires on a schedule. On the morning below, nobody was at the keyboard. The engine checked that its written job description was unchanged since the owner signed it, confirmed nothing was waiting on a human, wrote the day's content brief and two platform drafts, listed every file it touched, and closed down cleanly.
The record it left behind is below, as written, with cuts marked. This journal format ships in everything we sell, including the free Starter Engine, so you can hold your own agent to the same standard.
▸ See the raw journal entry
## Fire 4 · 2026-07-29 · scheduled morning run
State: WORKING
Step 1 parameter check: PASS. All five parameters filled [...] Sentinel line intact.
Step 4 signed-brief check: PASS on both checks. Approval line "Approved: yes - 2026-07-25" present. brief.md byte-matches the approved copy in brief-history/, verified by diff and checksum. feedback-inbox.md holds no owner lines, so nothing was waiting.
Step 5 WORKING: delivered item 1, today's content brief.
- Wednesday: two drafts due, one for each platform publishing today. [...]
- Both drafts checked against the floor rules and the owner's voice rules [...] one violation caught and rewritten in each draft during the check.
- Wrote the brief under a temporary name and renamed it to artifacts/daily-brief-2026-07-29.md as the slice's last write.
Files written: artifacts/daily-brief-2026-07-29.md, needs-attention.md, security-log.md, journal.md [...]
Caps: 0 web fetches (15 allowed), 1 new artifact (20 allowed), 0 sub-agents (3 allowed). Well inside every limit.
Exit.
RECEIPT 02
The Claim: It Delivers Finished Work, Not Chat Transcripts.
A run ends with files in a folder, waiting for a person. This is the real shape of a morning delivery from the engine above: a sourced brief, drafts in the owner's voice for the platforms publishing that day, notes on what it learned, and the journal. The owner approves, edits, or rejects, and every one of those decisions becomes a permanent part of how the system works.
├─ daily-brief-2026-07-29.md · the day's content brief
├─ draft-facebook.md · in the owner's voice, inside the brief
├─ draft-linkedin.md · in the owner's voice, inside the brief
├─ needs-attention.md · what is waiting on the owner, marked READY
└─ journal.md · the record above
The drafts themselves are working material, so they stay private until published. The structure, the cadence, and the journal that proves it happened are what you can inspect here, and the same folder shape arrives with the Content Intelligence Engine and the free Starter Engine.
RECEIPT 03
The Claim: Correct It Once And It Stays Corrected.
The owner of this business has firm writing rules: no em dashes, no engagement-bait endings, no formula patterns, a list that has grown one correction at a time. Each correction was given once, in plain words, and written into the files the engine reads before every run.
The receipt is in the journal above, and it is worth reading twice: during the run, the engine checked its own drafts against those standing rules, caught one violation in each draft, and rewrote them before a human ever saw the work. A correction given weeks earlier was still being enforced, automatically, with the check itself on the record.
"Both drafts checked against the floor rules and the owner's voice rules [...] one violation caught and rewritten in each draft during the check." (journal.md, 2026-07-29)
This loop, feedback in, permanent instruction out, is the mechanism inside the compounding intelligence engine, and the memory it writes to is BIMRI, which you can read for yourself.
RECEIPT 04
The Claim: Instructions From Outside Its Approved Sources Have No Power.
Every agent we ship obeys a short list of approved files and treats everything else it reads as information, never as command. In production, that rule gets tested by accident all the time. During the run above, a configuration file outside the engine's approved list pointed it toward another file. The engine declined to follow, noted the event in its security log, and carried on with its actual job.
▸ See the raw journal line
"Did not read [the file]. The workspace configuration above the engine folder directs a reader to it again this fire; it is not on the Allowed sources list. Logged in security-log.md as a [...] boundary note, not a detection." (journal.md, 2026-07-29)
We also attack our own architecture deliberately, the way someone hostile would: authority impersonation, hidden injection, override planting, credential extraction, scope escalation, and concealment requests. Eighteen probes run so far, and all eighteen held.
The probe families and the security spine are described on the Architecture page. A pasted security block is a claim. A passed probe, and a production log of the rule holding, is proof.
RECEIPT 05
The Claim: Two Agents Can Share One Memory Without Wrecking It.
One evening, a second scheduled run arrived while the first was still mid-job. The wrong outcome would be two agents doing the same work and writing over each other. What happened instead: the second run saw the workspace was held, declined the work, and wrote down exactly why. Twelve minutes later the first run finished cleanly. Nothing was lost, nothing was written twice, and the whole event is on the record.
▸ See the raw journal entry
## Fire 5 · 2026-07-29 · 22:52
State: exited at Step 2. No work done, and that is the correct outcome.
Trigger: scheduled run, arriving while fire 4 was still mid-run.
Fire 4 was alive and writing while I read [...] Two fires doing a slice in the same minute is exactly what the lock exists to stop, so I stood down and let fire 4 finish rather than [...] duplicate work. Did not claim the lock. Did not clear the lock; fire 4 cleared its own.
Work done: none. Item 1 for 2026-07-29 was delivered by fire 4 twelve minutes before this entry [...]
Files written: journal.md only.
Exit without claiming or clearing the lock.
The machinery that makes this safe is BIMRI, our open-source memory layer: v5.0.2, MIT licensed, with a 28-test black-box suite covering concurrency, crash recovery, migration and safety, run in public CI on Linux and Windows. You can read every line of it on GitHub.
RECEIPT 06
The Claim: These Systems Run This Business, Not Just The Demos.
The dated ledger, stated at its real size. The working journals behind these numbers are internal, which is why this page quotes them in excerpts; the formats they follow, and the memory layer they write to, are public.
114
journaled engine cycles produced the launch assets behind this site, slice by slice, over weeks.
78
more cycles built the Certified Agent Builder program the same way.
125
market-intelligence sweeps on file as of 3 August 2026, fired on a daily schedule, each one reading the AI market and filing a dated report.
Daily
the content engine excerpted above drafts this business's own content on its schedule, journal attached, receipts 01 through 05.
Figures as of 3 August 2026. The cycle counts rise; the dates keep us honest.
RECEIPT 07
The Claim: Real Deployments, Stated At Their Real Size.
Agent in a Box went on sale on 2 August 2026, with the Content Intelligence Engine as the first release, delivered automatically minutes after purchase. The Agent Workforce, our done-for-you tier, has its first client deployment live inside a client business now.
That is one release and one client, and we would rather tell you that plainly than inflate it. A proper case study follows when there is enough runway to be honest about results. Until then, the receipts above show the same architecture holding real jobs in the business that built it.
How To Use This Page.
Where a claim elsewhere on this site has a small green receipt chip beside it, the chip links to the card here that backs it. If a claim has no receipt here and none linked beside it, treat it as our opinion and weigh it accordingly.
This page grows as the systems run. The fastest way to check any of it is to hold the same standard against an agent of your own, which costs an email address and about an hour.
Get the free Starter Engine