EXP-001 · Q2.3 local replay guard
Can a per-attempt functional replay budget prevent cumulative forgetting while quantity is learned?
200 accepts, 0 rejects; replay +2.685%.
Read the full experiment →Certified micromodel research
Fund uncertainty.
Settle in evidence.
A public guide to frozen questions, exact execution contracts, negative results, and the decisions they earn. Failed gates stay visible. Promotion is never inferred.
Promoted modelZERO.4
Active research lineZERO.5 · 4.85M base
Next experimentC6.1 · built, blocked
Experiment registry
One question, one contract, one decision.
EXP-002 · Q2.4 cumulative replay guard
Every candidate is checked against the same immutable six-source replay baseline before commit.
66 accepts, then 8 guarded rejects; no public checkpoint.
Read the full experiment →EXP-003 · Q2.5 deterministic backtracking
Retry one frozen AdamW proposal at eight geometric scales under the unchanged cumulative replay authority.
66 full + 5 scaled commits, then 8 exhausted attempts; no public checkpoint.
Read the full experiment →EXP-004 · Q2.6 replay-tangent projection
Remove only the replay-increasing component of each candidate direction, then apply the unchanged direct six-slice authority.
700 full-scale commits; public and one-time promotion gates passed.
Read the full experiment →EXP-005 · Q2.6 family replication
Seeds 1 and 3 ran independently under the accepted tangent design, with all three declared seeds required to pass.
Seeds 1, 2, and 3 resolved go; ZERO.4 is current upstream.
Read the completed experiment →EXP-006 · Holo HRR attention at T=512
A public CPU/enwik8 run verified the three-arm implementation and preserved a strict recovery trail after an invalid result envelope.
Softmax led both HRR variants in validation quality and speed.
Read the recovered result →EXP-007 · Zero → Solomon Q22 bridge
Three prospectively frozen Solomon integer class heads ran against the shared Q22 operation-routing task.
Every seed scored 500/500, with identical predictions across seeds.
Read the completed bridge →ZERO5 C5.1 · StateBridge text
Exactly 25% of the matched C4.3 stream was replaced with Braid structured-state text at the unchanged 4,852,992-parameter base.
Structured text alone did not improve the frozen retrieval audit.
Read the active program →ZERO5 C5.2 · TargetBridge auxiliary
The C5.1 stream remains fixed. A verified factorized state-target head is the only treatment, bringing the experimental model to 5,046,256 parameters.
The run finished under a no-publication boundary. Its outcome remains absent from the public program.
Read the publication boundary →ZERO5 C6.1 · Shared-State Bottleneck
A 152-wide state representation is shared by the verified factor decoder and a zero-initialized answer-token adapter, with a mandatory same-checkpoint bridge-off ablation.
The contract and code are frozen. No compute is authorized.
Read the prospective contract →Evidence discipline
What the page promises
Frozen before inference
Seeds, metrics, data hashes, stop rules, and promotion access are bound by machine-readable contracts.
Negative results count
A valid no-go closes a branch. It is not softened, hidden, or converted into a capability claim.
Replay over rhetoric
Clean commits, content hashes, deterministic adapters, and the event ledger define what is grounded.