01Proponent
The claim, at full strength
One agent states the claim as well as it can be stated — sources attached, confidence declared. Weak drafts never reach review.
Research / adversarial AI reviewOpen source · Apache 2.0
Ipcha
Mista
bra.
No claim ships until it has survived its strongest opponent.
Transform review from a polite formality into a structural guarantee. Every AI-generated claim faces three agents: a proponent argues it, the ipcha agent writes the strongest possible objection, an auditor arbitrates with the paper trail in hand. For organisations where an unchallenged answer is a liability, disagreement becomes architecture — not office politics.
Agents on every claim3
Claims challenged before release100 %
Hallucination floor — derived0.15 %
NLI score separation0.923
Open licence — fully auditableApache 2.0
Decisions without a written objection0
Method / three seats at the table
Disagreement,
by design.
“The opposite is plausible” — a Talmudic rule of argument, turned into an architectural constraint. The review is not a person having a good day; it is a role that cannot be skipped.
02Ipcha
The strongest objection
A dedicated adversary writes the best available counter-argument. Not nitpicks — the opposite case, argued to win.
03Auditor
A ruling with reasons
A third agent weighs both briefs and rules. Every verdict carries its reasoning, so humans review judgements, not vibes.
04Always
Nothing ships uncontradicted
The challenge round is structural. There is no fast path, no override for busy days, no claim too small to oppose.
05On record
A paper trail for every ruling
Claim, objection and verdict are stored together. Compliance reads a complete argument, not a log line.
06Neutral
Model-agnostic by principle
Proponent and adversary can run on different providers — the method audits the argument, never the vendor.
Why enterprises adopt it
Review that scales past the reviewer.
Human review is your scarcest resource. Adversarial structure spends it where it matters: on rulings, not on first reads.
01
Risk surfaces early.
Objections are written before release, not discovered by your customers afterwards.
02
Audits answer themselves.
Every shipped claim carries its own challenge and verdict — evidence your auditors can read verbatim.
03
No rubber stamps.
The adversary is rewarded for finding the crack, not for agreeing. Approval means something again.
04
Open and inspectable.
Apache 2.0. Your security team reads every line before it touches a workflow.
“The opposite is plausible.”
Ipcha mistabra — the working rule
Why an ancient rule
Talmudic scholars formalised a duty most review processes only gesture at: before you accept a claim, someone must argue — seriously — that the opposite is true. It survived two thousand years because it works.
Not a second opinion.
Opinions are optional; the objection here is mandatory. Every claim meets its counter-case, every time.
Not a compliance checkbox.
The output is a reasoned verdict with a paper trail — material your compliance process can actually use.
Not a debate club.
Arguments end in a ruling. The auditor decides, records why, and the pipeline moves on.
Stress-test the claims.
Open source and ready to sit inside your review workflow. Briefings and pilots go through the main page.