Three beliefs this project held, what contradicted each one, and what replaced it. None of it was removed.
CHANGE WITHOUT ERASING
Previous · preservedCurrent · bounded
The contradiction stays between them: it explains why the held belief changed.
01How to read
Two independent marks.
A badge grades a claim we are making now. A cross-and-edge marks a belief we stopped making, and carries no badge because it is no longer a claim.
VALIDATED
Supported by a human Decision (public identity). Lab dogfood does not use this mark.
EXPERIMENTAL
Lab result, with stated limitations.
UNKNOWN
Closed box: contents not established.
VISION
Not closed: a direction, not something started.
UNKNOWNThis page is ordered by revision, not by date. The founding year is not recorded in anything we keep, it is already declared unknown, and we will not invent one to make a line look continuous.
This page was assembled by the people it is about.
Three revisions below. Section 06 is what did not move.
There are no numbers on this page. Section 05 explains why.
02Where this started
Why does this exist at all?
DARXI started from a practical failure: work done with AI assistance kept losing its reasoning. The output survived; why it was chosen did not.
The first attempt was made on real work, with real consequences for getting it wrong.
03Revisions
01
BelievedHeld now
Kept, not corrected.
Believed
We believed the missing piece was a machine that does the work. The bottleneck was capacity: build the thing that executes and the rest follows.
Held now
VALIDATED
The bottleneck was never the model’s ability to do the work. It was that nothing durably held what had been decided, what had been proven, and what was merely asserted — so work got reconstructed instead of resumed.
Reading the project’s own history end to end established that, in its most ambitious form, the work had never once run end to end — while the tests reported success the whole time. Each half was tested against the good side of the boundary between them.
Limitation · This establishes what did not run. It does not establish that the design was wrong.
02
BelievedHeld now
Kept, not corrected.
Believed
We believed the system should settle the disagreement. Several proposals ought to converge into one answer, and the value was in producing that answer.
Held now
VALIDATED
The clearest thing this system has ever done was record a proposal, record a human declining it, and keep both. Not resolve them into a tidy summary: keep the disagreement, intact and findable, after the argument was over.
Put a real, live architectural question through several tools and you get several independent proposals and no merged answer. The useful artifact turned out to be the disagreement, kept whole with its provenance — not a summary of it.
Limitation · One question, one session, and every stance landed in the same conditional shape: the disagreement was narrower than the headline suggests.
UNKNOWNStill open
Whether a preserved conflict is actually useful to a human later has never been measured.
03
BelievedHeld now
Kept, not corrected.
Believed
We believed a number was a result. A figure the system reported about itself was treated as evidence.
Held now
EXPERIMENTAL
An unverified measurement carries no number at all. A number implies a measurement happened. And where two real measurements disagree because they exclude different things, the disagreement is left visible rather than reconciled into one tidy figure.
One rule with teeth — a measurement counts as real only if its input came from running something the program did not author, the arithmetic is reproducible in terms of countable things, and the source is concrete enough for someone else to re-run — reclassified a large part of what had previously been reported as guarantees.
Limitation · It is a lab practice, not doctrine. It has not been tested outside this project.
06What did not change
If you keep changing your mind, what is stable?
Three things held across every revision above.
Cited authority
Nothing is promoted without a cited authority. Not by recency, not by length.
The human decides
A human records Decisions. AI produces Proposals.
Nothing is deleted
Which is why the three blocks above still exist.
EXPERIMENTAL
Working rules, each written from a failure that actually happened
Nothing is born in a constitutional register: everything starts as a mutable entry with permission to be wrong.
A document becomes a Decision only when something demonstrably runs.
The medium of exploration is a runnable spike that gets deleted, not a document.
Many cheap options rather than one expensive bet: the cost of being wrong should be an afternoon.
A fixture-backed harness labels itself a fixture in its own output.
LimitationWorking criteria in an editable file, explicitly not doctrine, untested outside this project.
VALIDATED
When a task uncovers something requiring an irreversible or owner-level action, the system records that the action is required and who must authorize it — and stops.
07Where this leaves us
So what is it now?
An experimental knowledge-organization platform. Humans record Decisions; models produce Proposals. External human evidence N=0.
UNKNOWNWhether the current description lands better than the one it replaced. No perception study exists.
UNKNOWNWhether a preserved disagreement is actually useful to a human later. Never measured.
UNKNOWNDrift between a recorded Decision and the code is detectable on markers the system governs. Detecting it across an arbitrary codebase is untested.
UNKNOWNDecisions are recorded by a named human authority. On a project this size that authority is one person: a limitation, not a design goal. A governance model that only works with a single decision-maker has not been tested as a governance model.
UNKNOWNShared or hosted operation is not available.
Every measurement we have is a measurement we took of ourselves. No outside team has used DARXI and reported back. That is not modesty — it is the actual state, and until it changes, anything we show you is a demonstration and not a result.
VISION
The direction is operation across more than one organization, where patterns can travel and authority cannot.
This page was written against a rule about what may be said publicly. Some of what we learned is not on it, and the omissions are deliberate rather than decorative. A memory system that cannot keep a confidence is not a memory system anyone should adopt.