Measured, merged, or live. Numbers carry their exact context; limits are stated, not buried.
Quantization · SiliquantVerified results
E8 beats the industry quantizer at equal bits
Rounding a real model's weights with the E8 lattice damages its output less than llama.cpp Q4_0 at the same bit count.
Gemma 4-E4B, one scale per 32 weights from the largest entry, 10,200 positions. Mean KL at 4.5 bits: E8 0.075 vs Q4_0 0.117 (first-choice word kept 91.6% vs 89.4%). At 4.005 bits: 0.215 vs 0.448. 10 of 18 preregistered predictions hold.
Limit: One model, one text. No packed file format exists yet and no decode speed was measured, a scored result, not a shippable quantizer.
Moissanite-Labs/siliquant PR #2 · 2026-11-15
Quantization · SiliquantVerified results
The integer ball beats E8, and the lattice is a small slice of the gain
Cutting weights to a ball, with no lattice at all, beats E8 and the industry baseline. The E8 cell shape is worth only about 0.65 dB; most of the gain is scale plus ball.
hZBG128 (ball + half-strength importance matrix): KL 0.0209 on Qwen2.5-1.5B, 0.0282 on SmolLM2-1.7B, ahead of matrix Q4_K_M on both. Decoder tables: E8 561 MB, integer ball 31 MB, predicted ZBG64 4 MB.
Limit: Branch result, not on main (branch bayan/charming-gates-jlas14, PR #12 draft). Qwen record partial, 27 of 32 variants; SmolLM2 complete, 13 of 13.
Moissanite-Labs/siliquant PR #12 (draft) · 2026-11-17
Quantization · SiliquantVerified results
Exact closures: billions of weights provably do nothing
Large parts of a real model can be deleted exactly, with zero output change on the tested text.
2,794,842,624 weights zeroable for this text; 55,050,240 weights zeroable for any text, both scoring exactly as the unchanged model.
Limit: Exact on the tested text and model (Gemma 4-E4B).
Moissanite-Labs/siliquant PR #2 · 2026-11-15
Quantization · SiliquantVerified results
The scale rule matters more than rounding precision
How you choose the scale matters far more than how precisely you round.
Clipping 0.17% of weights (−27.5 dB weight error) gives mean KL 0.318, vs 5-bit rounding of all weights (−21.7 dB) at KL 0.118. Best damage predictor tried: largest singular value of the weight change, rank correlation 0.889 over 26 variants vs 0.603 for squared error.
Limit: Same Gemma experiment as the E8 result above.
Moissanite-Labs/siliquant PR #2 · 2026-11-15
Quantization · SiliquantVerified results
The zero-codeword pitfall is model-dependent
E8's codebook must include a zero point or it breaks, and whether that matters depends on the model.
Without a zero codeword, E8 Voronoi breaks per row: 3.658 vs the scalar rule's 0.632. On Qwen the zero codeword helps at every level; on SmolLM2 it is 7% worse with groups of 32. Why it flips is unexplained.
Limit: Verified measurement, open question.
Moissanite-Labs/siliquant issue #16 · 2026-11-17
Steering · HeedleVerified results
Heedle v1, steer a small model by hand, inside checked bounds
A person can steer a small model's change by hand, propose, steer, hold, check, keep or discard, with every step checked and replayable.
Two hosts: cloth/1 (467 weights, point to colour) and text/1 (4,700 weights, 3 chars to next char, dispute exhaustive over all 21,952 contexts). 30 Python + 21 JS tests pass on the merged head; both sample drafts verify; CI green. The verifier hole was independently re-forged on both hosts and caught.
Limit: Code and tests only in the review; no browser pass, no re-run of the headless Chromium export.
Moissanite-Labs/heedle PR #1 · 2026-11-17
Design · Weft-eidoVerified results
Weft-eido #21, tide palette, interference map, loom research
A new take on the whole design language: tide palette, interference map, and research on the loom.
118 tests pass, 0 fail, on a clean-checkout run; all CI green.
Limit: Code and tests only in the review; the author and both earlier approvers did the browser pass.
Moissanite-Labs/weft-eido PR #21 · 2026-11-17
Verified software · Proven-CommonsVerified results
Lean-verified libraries: parsers and codecs with proved round trips
Twelve format libraries rebuilt in Lean 4 with machine-checked proofs, parsing and printing that round-trip by proof, not by test suite alone.
MessagePack decoder Bencode encoder/decoder (200,000/200,000 requests, 0 failures in review) UUID (RFC 9562, 10,562 inputs, 0 failures) ZIP reader JSON parser/printer RFC 3339 timestamps UTF-8/UTF-16 codec IP parser/printer TOML parser/printer (77/77, 300/300, 60/60, 1869/1869, 0 failures) codec pack (hex, base64, CRC-32, Base32, percent-encoding, PEM, Luhn/IBAN) protobuf wire codec
anthology compilation layer.
Limit: Claims are per-slice: parser first, printer separate. The HTTP request parser is still open (PRs #3/#4). The IP claim scope, parser-only vs expanded to printer, is still undecided.
Proven-Commons org (40 repos) · 2026-10-26
Search · SokobanVerified results
An 84-key route breaks the Sokoban holdout
A verified 84-key route beat the published 87-key record and the benchmark creator's unpublished 85-key holdout, cutting the trilogy shortest-known sum from 3,064 to 3,063.
Room III-20, Heroes of Sokoban III. Route sha256 eecbea96b091, credited to a1kindus on the live Frontier, replayed key by key in the original PuzzleScript build.
Method: the board is vertically symmetric in walls, goblins and goals but chirally asymmetric in hero placement, switch Wizard to Priest immediately and walk the Priest down instead of up; found by role-agnostic vertical-mirror Tube Beam search plus iterative tube refinement.
Limit: 84 is the known record, not proven optimal: the proven floor was 74 at the break and is 77 now.
ahrazzle/soko, docs/frontier-dossier.md · 2026-11-13
Search · SokobanVerified results
A verified lower bound: no III-14 route shorter than 38
A machine-checked proof that no route shorter than 38 exists on the true 15×7 board for room III-14.
Finished f=37 layer: 117,766,522 unique canonical states, 181,959,400 edges verified, 0 Bellman violations, 82.41 s; reproduced by an independent prover.
Limit: Still 6 short of the live floor of 44, the next step is pushing f=37 to f=44.
ahrazzle/soko PR #43 · 2026-11-13
Mathematics · Möbius transformVerified results
Full E1 network check for the GF(2) Möbius transform
The missing full-network run for Theorem 2 of the o(N log N) GF(2) Möbius-transform claim, built locally and contributed upstream.
h = 8–11, zero failures, both controls refuted, against Yates' 1937 bound.
Limit: Contribution to Swapnil Jain's mobius-x1; the gate-count figures reported elsewhere were not re-checked here.
Swapnil-jain/mobius-x1 PR #1 · 2026-10-11
Agents · OptChat memoryVerified results
optchat-memory v1.0.0, endless memory for Hermes Agent
An OptChat-style endless-memory provider: append-only log plus binary summary tree as recall, implementing Victor Taelin's OptMem spec.
Release v1.0.0, 2026-10-11; catalog PR NousResearch/hermes-agent#132907 merged; site live.
aska-digital/optchat-memory · 2026-10-11
Review · ExoVerified results
A claimed ssh gate hole that did not reproduce
A claimed security hole, keep the host read when an ssh remote turns credentials, does not reproduce: the hole is absent on base, and the new test passes unchanged on base, so it pins nothing. The real effect was five different shapes the PR never mentioned.
Review REQUEST_CHANGES on exo PR #664, head fd0e6d3d, 3 inline comments; a second changes-requested review the same day. The PR is still open and blocked, the catch held.
Limit: Full suite not rerun in that review session. The contribution here is the catch, not a shipped fix.
Moissanite-Labs/exo PR #664 · 2026-10-27
Rigidity · Foundation-0Verified results
880 real realizations at 10 vertices
880 realizations confirmed by three independent implementations.
Deposited on Zenodo, 2026-11-11, DOI 10.5281/zenodo.24145786.
Limit: Internally verified; externally unverified.
Foundation-0/rigid-realizations PRs #16–#19 · 2026-11-11
Lattices · Foundation-0Verified results
Incentre lattices v3.0.0
The incentre construction through honeycombs, borides, and rigidity in n dimensions; lattice work verified in dimensions 2–9.
Release v3.0.0, 2026-11-03; elastic proof with exact motion to FCC.
Limit: A mathematician's line-by-line read of the elastic proof is still pending.
Foundation-0/incentre-lattices · 2026-11-03
Geometry · Foundation-0Verified results
Vitruvian3d motion and conductivity
Local rotating-motif motion inside a fixed exterior, proved; conductivity rank contrast from certified passages.
Merged PRs #9/#10, 2026-11-12–13.
Limit: Internal AI review only; no human mathematics review yet. The motion does not transfer to quantization, tested and ruled out.
Foundation-0/vitruvian3d · 2026-11-12
Codes · Foundation-0Verified results
GF(q) code bounds improved
Exact Delsarte certificates improve code bounds over GF(7), GF(8) and GF(9).
Release v1.0.0, 2026-11-11; 12 GF(9) bounds lowered by one, two GF(7) codes excluded.
Limit: The wider ≥2,418-entry scan is still draft.
Foundation-0/gfq-codes · 2026-11-11
Physical trade · ASKA PhysicalVerified results
The Dock fused story, R&D floor plus verified trade stack
Five R&D modules paired to four fail-closed gates (brief, qualify, quote, berth) in one narrative: hardware floor plus verification layer.
Dock PR #6 merged 2026-11-17; OptiFab campus model (10 MW IT on 2,500 m², PUE 1.074) added as #/optifab via PR #7, merged the same day. OptiFab calculator outputs harness-verified against its model.
Limit: The five module figures (Taper-Lock, Kerf-Box, GeoBinder, Mill-Eye, Pooled-Cut) are presented by the deck and site, not independently checked. OptiFab v0.2.0 is claimed and unreviewed.
aska-physical/dock PRs #6, #7 · 2026-11-17
Governance · EthosVerified results
Ethos MEP v0.1, a minimum viable ethos protocol
The most portable already-shipped item from the November evaluation: a minimum viable protocol for ethos records, private by default, publication opt-in and granular, with revocation.
Merged 2026-11-14.
Limit: Exact PR number not yet pinned; link to be added once verified.
Moissanite-Labs/ethos (per internal-stoa#28) · 2026-11-14