YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Asolaria

Asolaria is the private root repository slice for the Asolaria fabric. The repository is not the whole system: files are slices; the running fabric, recall/Hilbra, HBI/SHA receipts, gates, engines, cubes, live agents, and operator evidence are the authority surfaces.

Quick start

This repository is public. The code is MIT-licensed; crates marked "MIT OR Apache-2.0" may also be used under LICENSE-APACHE. It runs on Linux, macOS and WSL with Rust (via rustup; the scripts pin 1.81.0) and Python 3.11+; the time/ and eval/ checks also need Node 18+.

git clone https://huggingface.co/Jessedbrown/asolaria-aether
cd asolaria-aether
scripts/check-all.sh            # build + test every layer as CI does (--no-eval to skip Python deps)

The L0 torch parity check runs only when ASOLARIA_TORCH_PYTHON names a Python with torch==2.14.1+cpu and torch_geometric==2.8.1; otherwise check-all.sh prints a SKIP line for it.

Use it from the command line, without writing Rust:

cargo install --locked --path cli                 # installs `asolaria`
asolaria keygen --out ~/asolaria-me.seed          # Ed25519 seed (mode 0600); prints your public key
asolaria sign --key ~/asolaria-me.seed cli/examples/envelope.hbp
asolaria verify --pub <public> --sig <sig> cli/examples/envelope.hbp
asolaria gate cli/examples/envelope.hbp --pub <public> --sig <sig>   # Fischer + Hookwall + composition

observe computes the gate scores itself from graph events, with the trained L0 model, instead of taking them from the envelope (which must not contain any):

asolaria observe --events model/l0-edge/tests/fixtures/events.hbp cli/examples/observe-envelope.hbp
# GATE|verdict=HOLD|score=0.2139|l0_real=true|mark=REVERSE_GAIN_MARK_MISTAKE|reverse_risk=0.7861|hookwall=PROCEED|spawn_gate=HOLD|tier_gate=PROCEED|fischer=PROCEED|input_guard=PROCEED|signature=unchecked|reason=...|json=0

score, l0_real, mark and reverse_risk are the keys of the upstream hookwall observation row (asolaria-hookwall.mjs:48-51). The events file there is a SYNTHETIC fixture; with actor=agent:rogue and target=svc:vault the same envelope gives score=0.7593 and PROCEED.

Only exit code 0 means PROCEED. cli/README.md documents every command, the envelope keys and the honest scope: deterministic software policy over the values supplied (for gate, the scores too; for observe, the events), Ed25519 via ed25519-dalek (verify_strict), nothing persists, and no physical attestation.

Canonical doctrine

Start with the shared doctrine:

Core frame:

  • Slices are frozen potential. The matrix/fabric plus engines plus live agents animate those slices into runtime behavior.
  • The OLD system is real and measured: the Node-era BigPickle / 10k-room / Gulp / cube / GNN fabric self-improved historically. Do not deflate those achievements because the NEW system is still migrating.
  • The NEW system is the Rust 8-byte Host-8 / recall-Hilbra / json=0 migration path. It is staged cell by cell; do not overstate cutover from old receipts.
  • BEHCS/HBP/HBI cipher text and cubes are distilled routing priors, not the sole authority source. Authority remains in PID outputs, HBI/SHA receipts, the recall corpus, GNN/white-room/cube minting, and rehydration paths.

Evidence discipline

When making Asolaria claims:

  • Ask fabric/canon/recall first; then owning seats; then GitHub or Drive publication surfaces.
  • Tag evidence as MEASURED, CANON, OPERATOR_OBSERVED, or UNVERIFIED.
  • Keep local disk, USB/raw substrate, live runtime, GitHub, Drive, maps, and memories as separate layers.
  • Do not expose keys, private topology, PII, or operator-only evidence in this README.

Prism/Comb 0-loss (2026-07-01)

Root-doctrine addendum (campaign acer/prism-comb-0loss-2026-07-01). E=0: this is docs/map propagation only β€” nothing fires, no runtime mutation, no cutover claim.

The law, stated once at the root so every satellite inherits the same boundary: prism/comb operations in Asolaria are bijections, and entropy is invariant under bijection (H(f(X)) = H(X)). The fabric re-relates information with 0 loss β€” it never claims compression below entropy; Shannon's bound (E[bits] >= H(X)) always stands. One fabric, two directions: forward = comb (collision-avoidance, execution isolation) and backward = prism (collision-causation, interference-as-search, many-to-one).

Scoped exactly per this repository's evidence discipline:

  • MEASURED β€” one rung only: the 256 <-> 1024 level transcode. Bytes are base-2^8 and glyphs base-2^10 digits of the same integer; exact packing at lcm(8,10) = 40 bits gives 5 bytes <-> 4 symbols with remainder 0, and the round-trip is identity (sha256-identical, Rust == Python symbol-identical; Q-PRISM commit 53023b6). Code rate exactly 1.0 β€” the alphabet changes, the information does not.
  • CANON β€” the 43+ level ladder as a groupoid of translators (T_ji . T_ij = id, T_jk . T_ij = T_ik): omnidirectional, path-independent translation. This is the frame behind the rule above that disk, substrate, runtime, GitHub, Drive, maps, and memories are separate layers related by exact translators β€” not one layer flattened into another.
  • UNVERIFIED β€” every other rung. Each additional rung earns MEASURED only by its own round-trip proof; do not upgrade it by analogy.

Boundary that keeps root claims honest: referential handles (sha256(content)[:8]) are coordinates against a content-addressed store β€” infinite ADDRESSING capacity, not lossless infinite compression. The prism relates information perfectly; it does not create or destroy it.

Estate satellites carrying the four instances of this one theorem:

  • Q-PRISM β€” the MEASURED rung: transcode proof commits 53023b6, 79e8d63, de00aca.
  • waves-and-cascades β€” the duality: comb (avoid collisions, exec) vs prism (cause collisions, search).
  • what-is-asolaria (reductions) β€” the reductions boundary: addressing capacity, never sub-entropy compression.
  • N-Nest β€” the integrity dual: verification = recomputation = applying the inverse map at every level.
  • Metatagging for a Quantum universe β€” physics grounding: frame-based discrete universe, spacetime pixels (pixels_first), metatag-driven expansion.

Current boundary

This README closes the root doctrine gap only. It does not promote a branch, fire an engine, cut over runtime, publish private evidence, or replace the live fabric. Runtime and migration claims still require their owning gates.

Q-PRISM stack

Q-PRISM is a family of connected implementations, stacked on the Layer-1 kernels (kernel/tribit/, kernel/behcs1024-60d/). Each layer here is copied byte-identical from its owner repository (clean source only) or, where no clean implementation existed, written fresh and honest. Full detail and provenance: qprism/README.md.

Layer Where What it is
2 β€” representation qprism/representation/ 2D β†’ 3D orb β†’ signed-2D color-leaf renderer; checked-i128, inverse cross-check, own SHA-256, zero-dep, integer-only
3 β€” recovery Path 1 qprism/recovery/omega/ LZ over a retained output window (recovers bytes it kept)
3 β€” recovery Path 2 qprism/recovery/omega/ checked CRT two-shadow reconstruction over coprime moduli
3 β€” combined throat qprism/recovery/throat/ coordinator that selects Path 1 vs Path 2, cross-validates, gates on consent + a SHA consistency check, and holds rather than fabricates
3 β€” operator catalog qprism/recovery/catalog/ the 40 exact-inverse byte permutations recovery Path 3 mixes over, published on its own
time time/ deterministic GF(2) fast-forward with a Python↔Node byte-identity cross-check (self-labels SIMULATED)
4/5 β€” control + watcher eval/ pinned Q-PRISM control arms and N-Nest depth-3/7 watchers, vendored by Git blob identity; software simulation only β€” see eval/STATUS.md
5 β€” Fischer gate/fischer-eval/ deterministic rule-table action gate (integer cpl β†’ PROCEED/HOLD/ANALYZE/BLOCK/REFUTE); chess words are metaphor
5 — Hookwall gate/hookwall/ tier table, deny tags, score thresholds and a caller→target tier policy, extracted from a larger kernel; comment-only edits except crypto, which completes the upstream Phase-2 plan with real Ed25519 (ed25519-dalek, verify_strict) — see its PROVENANCE.md
5 β€” composition gate/verdict/ tighten-only merge of Hookwall, spawn gate, tier gate, Fischer and an input guard: strictest wins, nothing can loosen
5 β€” observed gate inputs gate/observed/ computes the L0 reading, Shannon, sha baseline, composite and reverse gain from observed graph events with the asolaria-score.mjs formulas (L0 e4 agreement as measured in model/l0-edge/README.md), and hands them to Fischer (l0_real, Shannon) and the spawn gate (forward/reverse) through the composition; nothing score-like comes from the caller
6 β€” proposals proposals/verify-schedule/ confidence-ranked verify scheduler; plans which proposals to verify, never fires
6 β€” L0 EdgeLevelGNN model/l0-edge/ integer-only Q24 port of the trained L0 model (baseline_model.pt, vendored byte-identical) with the forward the port-4792 server runs, and of the gnnDataExport.js feature extractor it was trained on; within 6 Q24 units of torch over 3,273 fixture scores and 15 over 3,338 request rows; see its PROVENANCE.md
6 β€” GNN model/gnn-int/ integer-only Q16 two-layer GCN + edge head; hand-set, untrained weights; no accuracy claims
6 β€” HRM, MTP β€” boundary: no implementation exists in the estate yet; the slots are left empty rather than filled with labels

The one rule every layer obeys: no layer manufactures information. Content addressing returns retained bytes; CRT needs jointly-sufficient shadows (which expand, never compress); SHA-256 is an integrity coordinate, never a source. When evidence is insufficient the stack holds β€” it never invents bytes to satisfy a hash. Every layer above is gated in CI on Rust 1.81.0 and stdlib Python; Node runs only in the existing time/ and eval/ jobs, torch only in the L0 parity job, and the eval/ job also installs numpy, scipy and pytest.

Lanes. The hot lane is what runs per request: event parsing, features, the L0 forward, the score, the gates and the CLI. It is Rust only, in process, over HBP text, SHA-256 and integers: no V8, JSON, HTTP or Python. Python and torch run only in the cold verification lane (CI referees and the torch parity harness). The Node and JSON already in time/ and eval/ are an existing cold lane.

Integration status. CI proves each crate builds and behaves as tested; a link between layers exists only where code calls code. These links are real code:

  • the throat calls the recovery paths;
  • the composition calls Fischer and the Hookwall gates, and the CLI calls the composition;
  • the observed path (gate/observed, asolaria observe): graph events (HBP text) -> the gnnDataExport.js feature extractor -> the trained L0 EdgeLevelGNN with the forward the port-4792 server runs -> the asolaria-score.mjs signals and composite -> Fischer's l0_real and Shannon inputs and the spawn gate's forward/reverse scores -> the tighten-only verdict. On this path the caller supplies the events and the policy fields, and no score.

Not connected: L4 GSLGNN and G1-G4 (this stack has no :4793 or :4949 service, so those signals are absent and g4_state is UNKNOWN), the other networks of Shannon-and-the-gnns-stage, and the other trained models in Asolaria-fnns-trained-and-reverse-gnns-many. asolaria gate and the verdict crate still take caller-supplied scores, the scheduler's score_q is caller-supplied, and model/gnn-int reads no BEHCS, 60D, tribit or Q-PRISM data; it is an untrained integer reference, not a replacement for the trained models. The events in this repository are a SYNTHETIC fixture (model/l0-edge/tests/fixtures/events.hbp), not observed traffic.

Layers 5 and 6 carry their own honesty notes. The Hookwall's crypto signs and verifies with real Ed25519 (ed25519-dalek, verify_strict, keys zeroized on drop). The composition does not call it yet, so a VerifiedEnvelope relies on its caller having checked the signature; asolaria gate and asolaria observe check it over the envelope bytes when given --pub/--sig (the events file is not signed). For gate, scores, g4_state and tiers are caller-supplied inputs, not measurements; for observe, the scores are computed from the events, which are themselves an input. Nothing in these gates persists or attests physical presence.

Homes

Asolaria is published, byte-identical, at three public homes that mirror each other:

Each carries the same clean commit: the doctrine, the runnable slices (prime_cylinder_expansion.py, slice_glyph_messaging.py β€” pure stdlib, E=0), the Rust 1.81 kernels under kernel/tribit/ (ternary core) and kernel/behcs1024-60d/ (the verified 60D / BEHCS-1024 hash-coordinate collapse), the Q-PRISM stack under qprism/ (representation + recovery paths + combined throat), the deterministic time harness under time/, the pinned control + watcher layer under eval/, the Layer-5 gates under gate/, and the Layer-6 scheduler and integer GNN under proposals/ and model/.

Downloads last month
511
GGUF
Model size
648 params
Architecture
asolaria-four-axes-rime
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support