Frozen donor representations
Hidden states, RepE behavior vectors, and task-vector deltas captured offline from larger models. Donor checkpoints are pinned to HuggingFace revisions and never ship with the fabric.
VIVERE Research Platform
Move knowledge between models — and publish what survives.
VIVERE is a research platform for knowledge extraction and transfer: compressed latent cards, RepE and task vectors, and a six-level Knowledge Fabric. Every number on this page points at a committed measurement.
v1.12.0 · 3,107 tests passing · as of 2026-09-22
01 · What VIVERE is
Measured, not asserted.
Knowledge transfer between models is usually asserted, never measured. VIVERE extracts compressed representations from donor models, stores them in a versioned fabric, and ships the controls that show what each piece does — and what it does not.
VIVERE cards, rank 8
Singular-value truncation turns donor states into compact cards: a packet pointer stays under one kilobyte, and wire records run far below the uncompressed representation size.
Controls before claims
Byte-exact conformance between the Python pipeline and the Rust/WASM wire-format crate, memory-intervention falsification controls, and published null results. An untraceable number does not get printed.
02 · The stack
One wire format. Two languages. Byte-exact.
The Python pipeline and the Rust vivere-cards crate agree byte-for-byte, checked by conformance harnesses against the same golden fixtures — including, in the browser, through WebAssembly.
VIVERE cards
Rank-8 SVD compressions of donor hidden states. Small enough to address under a kilobyte, reconstructable at the far end.
RepE + task vectors
Activation directions for behaviors and tone; weight deltas for procedural knowledge. Each with provenance attached.
Six-level Knowledge Fabric
From an L0 answer cache down to the L5 VIVERE latent store, with metadata indexing, a reasoning graph, and provenance on every record.
Rust, WASM, MCP
The vivere-cards crate compiles to wasm32 and decodes packets in the browser; an MCP server exposes the fabric to local agents.
L0 · CACHEHot answersexact repeat lookups
L1 · COARSEMetadata filterdomain and type cuts
L2 · ANNEmbedding recallapproximate neighbors
L3 · FINERanked rerankscored shortlist
L4 · GRAPHReasoning graphmulti-hop traversal
L5 · VIVERELatent cardscompressed originals
Deep-dive: wire format, compression, and a live packet demo →
03 · Open research
We publish our null results.
The repo's most uncomfortable file ships right next to its most polished ones. That is the point.
Student transfer null result
The student-transfer study measured no significant improvement: every transfer method landed within about a point of baseline at full evaluation, and the blocking mechanism — tokenizer disjointness between the donor phrase bank and the student vocabulary — is documented next to the data it came from.
Source: docs/experiments/STUDENT_KNOWLEDGE_TRANSFER_REPORT.md (committed snapshot report)A sparse recurrent CPU recipient using selected frozen representations extracted from a larger offline donor, with mechanistic routing and causal memory-binding controls.
docs/CLAIMS_BOUNDARY.md — the only sanctioned one-sentence description of the packaged demo
- That a teacher model was “compressed” into a smaller one
- Teacher-equivalent performance from a student model
- Production readiness — for any of this
- Any efficacy number: none exists yet, and we will not invent one
3,107tests passing
as of 2026-09-22 · py3.12 + py3.10
20/20WASM conformance checks
vs Python golden fixtures
98%minimum coverage gate
enforced in CI — below fails the build
100%of hub model ids pinned
40-hex revisions, ratchet-tested
04 · Status
30 of 34 fabric phases complete.
The remaining phases — and the first full student training run — are gated on GPU hardware, not on design. The v1.12.0 release shipped with the full gate suite green.
30/34roadmap phases complete
fabric roadmap, phases 3–33
1,610records in the consolidated fabric
from 2 extraction teachers
<1 KBpacket pointer, tested
asserted byte_size < 1024
05 · The substrate
One substrate. Three planned surfaces. Nothing shipping until it's measured.
VIVERE is the lab. The products below are planned to commercialize on top of it — and every one of them is earlier than it looks. Nothing here is a product you can buy today.
Amortyx
The capability-fabric product: request a capability, resolve it to local, remote, expert, or adapter. The resolver exists in the codebase today as AmortyxCapabilityResolver — as design-level infrastructure, not a product.
TaskWizer
A broad agentic development platform, planned to draw on the same fabric for retrieval-grounded tool use.
VibeGenie
A narrow consumer entry point. Its predecessor failed by faking determinism — the failure that became this lab's verification gate.
06 · Support the lab
Fund the unglamorous parts.
Conformance suites, coverage gates, pinned donors, and the next null result are the parts nobody funds — and the parts that make everything else trustworthy.
Support channels are being configured. They will appear here — each one linked, none faked.
07 · Contact
Ask us for the unflattering numbers.
Services, research collaboration, or a hard question about any claim on this page — the inbox opens the same way the repo does: with the boundary document.
Contact channels are being configured. They will appear here — a real address, a working form, no placeholders posing as either.
Claims boundary
Research platform, not a product pitch.
Nothing on this page is a capability claim about deployed models. What VIVERE ships today is the substrate: extraction pipelines, the fabric, the wire format, the controls, and the receipts. Where a measurement is missing, we say so in the same breath as the feature.
The full boundary — what the package supports, what it does not, and how its benchmarks may be interpreted — is the repository's CLAIMS_BOUNDARY.md, and it applies to this site too.