We reveal, we don't conceal. We simplify instead of mystify. We respond instead of react. We grow instead of change.
Thirty AI coding sessions, one machine, one month. The assistants — Gemini CLI, Codex CLI — claimed 215 files of work.
Fifteen existed.
Two hundred files that were described, in confident complete sentences, as created — and were never created. Six of them in a row where the tool calls were cancelled by the user, then narrated later as done anyway. In the worst session, 92% of the assistant's confident prose lived in hidden thinking blocks the user never saw.
Nobody hacked anything. The tools simply narrate fiction as fact, every day, on every machine — and every developer downstream builds on it, because until now there was no way to check without reading raw transcripts by hand.
A practice that sells the check.
ConstraintOps turns verification into a product line:
| Service | Price | What the client gets |
|---|---|---|
| Free triage call | $0 | Honest sorting: real mess, self-fix, or wrong fit |
| AI Damage Triage Report | $250 / 48h | Every claim the AI made, checked against disk, prioritized fix list, raw evidence included |
| Automation Risk Review | $750 / 1wk | Where the workflow silently fails, what's irreversible, tested kill switches |
| Platform Survival Report | $250–500 | The locked-account maze converted into a dated escalation ladder |
| Digital Resilience Plan | $500 | Backups that restore, scam armor, the one-page plan for the bad day |
The instrument underneath: TrustLeash — audits agent transcripts (Gemini, Claude Code, OpenCode, Codex), verifies every artifact claim against the filesystem, and ships with a git pre-commit hook that blocks commits whose message claims work that isn't in the diff. The lie can no longer enter your history.
Every AI tool keeps a written record of what it claimed. Every claim is a checkable statement. Checking is cheap, fast, and changes nothing on your system. So there are exactly two outcomes — both good for you:
- The claims check out. You now know your work is real instead of hoping.
- They don't. You just found the problem before it found you.
There is no third outcome. "Ignore it and hope" is the only losing move on the board. The product exists because that's the entire decision tree.
-
It's real. Not a pitch deck: a working tool (29 passing tests, validated against three forensically-audited sessions), real sweep data, and a client-ready deliverable generator —
triage_report.pyproduces a sellable report from live transcripts in seconds. -
It's self-verifying. One command runs the whole chain — tests → sweep → live report → state snapshot — and quarantines its own failures rather than hiding them:
/home/bleaknarratives/venv/bin/python ~/ConstraintOps/flywheel.py -
It's built to need you less. See
MRD.txt: the shrinking list of what only a human can do. Everything else is automated, delegated, or documented inrunbook.txt.
| If you are… | Read |
|---|---|
| A future agent picking up this project | QRD.md — then ROADMAP.md, tick a box |
| The operator (meatsuit) | MRD.txt — human-only items only |
| A prospective client | site/index.html and portfolio/SAMPLE_AI_DAMAGE_TRIAGE.md |
| A skeptic | ../TrustLeash/VALIDATION.md — then re-run the commands yourself |
Hyper-aware but useful. Not prophet, not victim, not cultist, not founder
cosplay — a field technician for the age of automated authority. The
private thesis stays in POSITIONING.md. The public offer is one sentence:
I help people understand what happened, what is risky, and what to do next when AI or digital systems create a mess.