A partner who argues back
Rethink for think
“I argue against your decision — before reality does.”
Free to start · No card required
I attack your conclusion. If it survives, it comes out stronger.
The dashboard takes six seconds to load. Let's put a cache in front of it this sprint.
— TURN 1
A cache makes a slow thing slow less often. It does not make it fast. So before we agree, we should know what is actually slow inside that page.
One system, three planes: your decision, your agents' actions, and the problem that stopped moving.
Each meets the scrutiny it deserves, and every judgment leaves a record.
Adversarial by design
Koestler argues against your conclusion, not for it. What survives the attack is what you keep.
Deterministic at the gate
An explicit rule table judges agent actions before they run, and the judge is a different model family by design.
Honest about hypotheses
When a problem needs a distant frame, what comes back is labeled unverified — and stays that way until you test it.
1.0 · DEBATE
Debate it like you would with a real partner
Write down what you are leaning toward. I lay your reasoning out and ask the questions a good colleague would — one at a time.
You walked in with one claim and out with a better one.
CQ defense rate
80%
Position before
Add a cache in front of the dashboard this sprint
Position after
Remove the repeated query first — cache only if it is still slow
1 still unanswered
How many other pages repeat the same query?
1.1
Pick your adversary
Find the holes, assume it failed, argue the other side, or questions only.
1.2
Questions you must answer
Answer it, concede it, or leave it standing on the record.
1.3
What actually changed
The position you walked in with, next to the one that survived — with the reasons, exportable as Markdown.
2.0 · GATE
Decide together what your agents may run
Agents keep working while you are away. I check the consequential moves before they run, and hand you what I should not decide alone.
| Action | Decision | ||
|---|---|---|---|
| 001 | 10.6 | PASS | |
| 002 | 89.7 | BLOCK | |
| 003 | 54.7 | ESCALATE | |
| 004 | 53.7 | DEBATE |
A rule makes this yours to decide. The score is below the threshold and it does not matter — the rule fires first.
7 criticality axes decided this score — "How hard to undo" weighed heaviest
The total also runs through a domain multiplier and is clamped to 0–100, so these points do not add up to it exactly.
override · reason recorded
2.1
A decision before it runs
Pass, debate, hand it to you, or block — decided by a rule table, not a model. Nothing is silently approved.
2.2
A record you can't erase
Nothing here can be edited or deleted — the database itself enforces that.
2.3
You can overrule it
The gate only asks that you write down why, and keeps that reason.
Start in shadow mode — nothing is blocked, everything is recorded, and the record shows what the gate would have stopped.
Connect over MCP — Claude Code today, more clients as they open up →
3.0 · INSIGHT
Cross the problem with a distant field
For the problem you have circled three times. I cross it with a field far from yours and hand back angles to test — not answers to trust.
The test suite fails at random, and rerunning usually fixes it.
PROBLEM
- goalA test should pass or fail for exactly one reason
- constraintThe suite has to stay under ten minutes
- failureA rerun passes with no code change
DOMAIN
railway signalling
MAPPING
- a test run → a train moving through a block section
- a shared fixture → a section two trains were cleared into
CANDIDATES
- Treat the suite like a relay racetoo close to a common metaphor
- The tests interfere with each otheralready said
- Two runs were cleared into the same sectionkept
The failure isn't in either test. It's in whatever cleared both of them into the same section at once.
If it holds, try: let each test claim the resources it touches, and fail immediately when a claim is already held.
Which fixtures are shared right now with no explicit claim?
Not verified — test it before you trust it.
06 / 06
What survives arrives with something to try and something to ask.
3.1
Mechanics, not resemblance
You see which piece of your problem maps onto which piece of the other field.
3.2
The model proposes, rules dispose
Obvious and repeated angles are cut by measurement, not by taste.
3.3
Everything is marked
Ideas from here stay visually separate from grounded work.
“Rethink for think”