Dweve Loom | Deterministic AI on Supported Paths

Choose the audience that matches your question

The page contains three selectable readings of the same subject.

528 domain specialists, 4 to 8 active per query, deterministic top-k routing

does this clause meet the retention rule?

multimodal, domain, meta, verify, transfer

reasoning, maths, science, code, language

Answered by your Health specialist. You can see exactly who helped.

Only the domain specialists who know step forward. The rest stay quiet rather than guess.

Is this medicine safe to take with my other ones?

Coding agent and operator assistance. Demo metrics are illustrative.

Terminal, search, lint, test, git, and more.

Remembers your codebase and team context.

Specialised agents collaborate across concerns.

Review diffs, request changes, final sign-off.

Replay any session bit-for-bit when something needs review.

Autonomous agents that write code and keep receipts

to handle network timeouts, 5xx responses, and idempotent-safe conditions. Pure helper, fully unit-tested.

A weave trace on every output. Built for procurement, audit, and regulated workflows inside Europe.

Through Fabric, or directly under a licence

the same jobs as typed nodes with distinct outputs

deep weave: reasoning, solver, specialists

a later constraint check points back to the exact span and cell

unrelated history stays visibly closed to the classifier

the final evidence stays linked to its source passages

Cheap representation up front, deep judgement only at the tail.

outputs stay solver outcomes, never prose confidence

a metacognitive selector compares the paths

branches that fail their own check are pruned and left visible

the selected specialists feed the existing graph

The exact location comes from extraction, unchanged.

facts enter through hard sockets, not a prompt summary

The cited document arrives verified, with provenance.

The SAT or UNSAT result comes from the solver.

models, embeddings, routers, transformers, specialists

the weave combines them without pretending they are identical

The rest of the catalogue, which stays available at no added cost.

The nodes the weave actually lit for this query.

the meter follows lit nodes, never the full catalogue

earlier patterns remain pinned and unchanged

replay turns green, neighbours stay pinned

every handoff numbered, every node versioned

deliberately excluded: private model thoughts

the failed verification node and the missing input stay visible

The weave means the same thing everywhere.

organises work with agents and authority

sibling descriptors carry different output types

perception fan-out, memory join, solver branch, specialist branch, verification gate, expression

passes through model, memory, retrieval, and routing nodes

a boundary marks conversion to native CNF, graph, interval, or proof

downstream nodes consume exact fields and spans

a consolidation event promotes an episode into semantic state

a decay event removes low-value working state

a gate model decides whether a deeper branch activates

source-addressed evidence leaves the final verifier

query A to causal, counterfactual and probabilistic

query B to deductive, temporal and deontic

outputs remain typed, the mode joins the trace

classification routes the problem into the matching region; the native artefact comes back out

features [ boolean, clauses, no arithmetic ]

three candidate solvers scored by theory, features, and budget

the selected solver returns UNSAT plus core

a second branch times out to unknown, no fabricated answer

AION attaches a certificate where the path supports one

specialist outputs rejoin solver, constraint, and verifier nodes

Four to eight wake per query, the rest stay dormant

a trace cursor prints the decision record at each level

averaging SAT with model confidence is rejected

no post-hoc quantiser stage appears in the path

a later weave swaps one ranker version, the diff isolated

live replay diverges at one external evidence window

a private-thought channel is absent by design

agents with identity, authority, and tasks

the shared faculties serve different responsibilities

the Loom graph lowers into Core operation paths

One Core, lowered to the metal underneath

external sockets enabled or sealed according to policy

The manifest is identical, the policy differs

all four pass through one Fabric conversation

a single generic chat bubble does not fit all six

the answer stays connected to what was seen

an unrelated personal conversation stays closed

many source cards narrow to three, then one evidence set

a valid schedule exits with every condition ticked

an impossible case exits as no valid schedule, with the conflict named

the lanes rejoin at a checked conclusion

higher tariff plus winter usage, a capped plan helps

no private thoughts, only the useful record

It came back in plain words you can read, next to the record of how it was reached.

if a check does not pass, it says so instead of guessing

the scale changes with the task, not the idea

the same weave file, only the scale differs

earlier questions show different archived weaves

Constraint learning first. A sparse domain-specialist layer narrows each task. Small bipolar language components render the result graph; they do not decide it. PAP traces, bit-identical replay under a fixed seed, and integer and bitwise inference keep the path inspectable and CPU-first.

The substrate numbers the systems claim rests on.

Domain rules live with the specialist that enforces them, not in a shared global set.

New specialists join without retraining the full catalogue. The router scores the new specialist against existing signatures.

Each cluster trains on its own corpus. A defect in the finance cluster does not affect the science cluster.

Content-addressed by weights, constraint set, and training corpus snapshot. Loom can expand a constraint set when its own calculations produce a candidate and Spindle researches and validates it; each addition is versioned. Upgrade is a deliberate, pinnable event.

Approximately 3 billion domain-specific constraints per specialist. Conflicts surface as typed refusals, not soft warnings.

Typed graph of Spindle nodes. Answers cite a content-addressed reference.

Domain corpus, domain evaluation suite. Failed specialists are tagged and not activated.

Each specialist owns a domain responsibility.

Fix the seed. Bit-identical results, compared to the last bit, in any environment.

An AION proof certificate attaches to every emit. Verify with your own tooling, not ours.

The suites ship with the repos. Use your own hardware. Candidate pool size is a tunable parameter.

PAP, BitWeave, and the router are open foundations. Public repos with their own benchmarks.

A proof certificate is attached on emit. Check it offline with your own tooling.

None. The path from input to answer is enumerable from the trace.

Every decision enumerates the constraints that fired. No implicit dispatch.

Every answer ships with this. The trace enumerates the constraints that fired, names the active specialists, and carries an AION proof certificate you verify on your own tooling.

The inference path uses integer and bitwise operations end to end. No floating-point non-associativity, no reduction-order surprises, no vendor-specific math kernels. Ordering of iterable state is sorted, random draws are seeded, time is not consulted. Replay is byte-identical on supported targets.

The 528 domain specialists span 10 clusters: foundational (core reasoning, mathematics, science and research, code and systems, language) and specialised (multimodal, domain-specific industries, meta-cognitive, verification, transfer and adaptation). Clusters are independent. Adding specialists to one does not require retraining the others.

A Loom specialist is trained on its domain, carrying its own knowledge graph, its own constraint set, and its own evaluation suite. Each specialist carries approximately 3 billion constraints in its domain catalogue. Two specialists working on the same token produce different outputs because they are different specialists, not because they were given different slices of the same neural network.

Routing is staged. Coarse filtering via hypervector signatures and approximate neighbour search narrows the catalogue. Per-specialist gate subsets decide pass or fail before heavy constraint evaluation runs. Permuted Agreement Popcount adds structural similarity on top of Hamming. The active frontier never scales linearly with specialist count.

Loom is a sparse cognitive graph built on binary constraint learning rather than a single weight tensor. A layered router selects a small set of domain specialists, solvers and other typed nodes for each query. Small language components render the completed graph. Inference is integer and bitwise, XNOR, AND, OR, and popcount. Outputs carry a full constraint trace, AION proof certificates, and are bit-identical under a fixed seed.

Specialists instead of one generalist. A constraint trace on every answer. Bit-identical replay under seed. EU residency, on-prem, or air-gapped. Scope a pilot, read the system card, or open the pricing page.

Loom assembles this trace as the inference happens

The trace is created as the inference runs, not assembled after the fact. Every output carries the source it came from, the specialist responsible, and a replay path. That is the record your reviewer accepts.