Where the next layer is built.
Working prototypes of the reasoning architecture — built and tested inside our own operation, promoted into products when they prove themselves. This is what we are working on right now.
intelligance ACTIVE RESEARCH
An intelligence layer above agents. Not an agent and not a framework — an intermediate layer that takes a task in natural language and returns a verifiable artifact: a typed graph of meaning with computed confidence.
REV WORKING PROTOTYPE · SHEET 1/1
A finite alphabet
16 node types, 18 edge types, mechanical composition rules. A parse can be validated, diffed between versions, and a mistake localized to one node — instead of "trusting the model".
Context lives in the graph
At every step the model sees only the current node, its path and the scene. Task size is limited by a question budget — not by the context window.
Confidence is arithmetic
Every edge carries a modality weight; a conclusion's confidence is the product along its path. "Probably mold" is computed — not a hedge the model added.
The completeness pass
Every expansion is asked twice: "is that the only option — what else?". It is this pass that surfaces the non-obvious paths a model stops short of.
Branch intersection
Non-obvious conclusions are born where independently grown branches meet. Intersections are computed on the graph — not guessed.
A verifiable artifact
The output is a graph, conclusions with confidence, a mechanical validation report and a full trace of every question asked — executable by any agent stack.
WHY IT MATTERS: THIS IS THE REASONING LAYER OF THE GITMIR ARCHITECTURE RUNNING AS CODE — KNOWLEDGE STAYS IN THE MODEL, THE DISCIPLINE OF ASKING LIVES OUTSIDE IT.
context-engine ACTIVE BUILD
Builds an object model of a product's business logic straight from its repository — then serves the smallest slice of it that answers a task. A code index knows where a function is defined; this knows that upgrading a subscription touches User → Subscription → Plan → Payment, and from which states that is legal. This is the machine behind GitMir Models.
REV WORKING SYSTEM · SHEET 1/1
Journeys need the whole catalogue
A user journey does not live in any single file — it emerges from holding the entire catalogue at once. That is why a per-file indexer cannot produce one, and why this is a different layer.
Verified by arithmetic
Everything the language model produced is checked mechanically: a reference to an object that does not exist is dropped and counted. Everything downstream trusts the edges.
The price is shown first
A dry run reports what will be read and what it will cost — before a token is spent. An index with a price must say the number itself.
THE MODELS IN THE GITMIR REGISTRY ARE BUILT BY THIS ENGINE — INTERPRET ONCE, REASON MANY TIMES, AS RUNNING CODE.
symbolic mind FOUNDATIONAL RESEARCH
An object neural network built from scratch — its substrate is objects, not tokens. 70 GB of source code translated into layered object data across many levels of abstraction; at every level, the finite types of mechanics and algorithms are extracted into formalized closed lists, together with the admissible links between them.
REV FOUNDATIONAL · SHEET 1/1
Objects, not tokens
The network's substrate is typed objects and their relations — so what it learns can be read, validated and diffed, not just sampled.
Finite alphabets per level
At every abstraction level the discovered mechanics and algorithms form closed, formalized lists — with the admissible link variants between them enumerated.
Knowledge as gravity
Accumulated knowledge acts as an attractive force in graph formation: structure condenses around what the network already understands — the way mass shapes space.
FIRST APPLIED INSTRUMENT: MEANING GRAPHS OF WORKING DISCUSSIONS THAT MARK WHERE PEOPLE FILL THE SAME WORD WITH DIFFERENT MEANING — DIVERGENCES ARE FLAGGED, NOT SMOOTHED OVER.
Agentic Systems Load WHITE PAPER · V2.0
Our research paper: an operational framework for measuring context reconstruction, coordination and verification in AI-assisted software development. The expensive part of agentic work is often not generating code — it is discovering the right business rules, reconciling interpretations and proving that a change is correct. ASL makes that work countable.
A task-conditioned unit
One Agent Load Unit (ALU) is one adjudicated context element required for a specific task — a state rule, an invariant, a permission, a contract. Artifact size stops masquerading as task load.
Count before predicting
Hidden Context Ratio, reuse, deficit and the four observed costs are separated from any predictive model. Ratios predict; only costs add. No universal constants are claimed.
Easier to falsify than to market
Six preregistered hypotheses, a measurement protocol that avoids circularity, and a benchmark template where GitMir is a condition to test — not a conclusion.
More is in progress.
Further Lab work is being prepared for publication — including the visual history of how a product's business logic evolves, commit by commit.
Working on the same layer?
We compare notes with teams building reasoning structures, program understanding and agent verification.