ZERO.4
The Q2.6 quantity model passed seeds 1, 2, and 3. It remains the deployed promoted model.
Lab program registry · reconciled 2026-08-28
The active lab,
not an old roadmap.
Braid produces governed data. ZERO runs model experiments. ilXyr records the contracts, evidence gaps, and decisions. This page keeps the promoted ZERO.4 line separate from the active ZERO.5 research line.
Promoted modelZERO.4
Active research lineZERO.5 · 4.85M base
Next experimentC6.1 · built, blocked
Ownership
One lab, three clear boundaries.
Braid · data plane
Compiles governed releases, representations, target streams, rights records, split seals, and verification evidence. A Braid release never authorizes model training.
ZERO · model engine
Owns the C11 model, training loop, frozen evaluator, execution receipt, and private checkpoint. A ZERO result never promotes itself.
ilXyr · evidence plane
Owns the cross-project registry, admission status, forecast history, evidence classification, and promotion decision. New runs require an ilXyr registration.
Model lines
ZERO.4 and ZERO.5 are not one checkpoint chain.
ZERO.5
A separate dependency-free C training line. Its base has 4,852,992 parameters and its C experiments start from the selected C2 checkpoint. C6.1 is built, but no experiment is currently authorized to train.
C5.2 TargetBridge
The authorized run finished, but its frozen contract forbids public results. Its outcome is not used as public evidence for C6.1.
C6.1 Shared-State Bottleneck
A separate public prospective contract couples verified state learning to answer logits. It does not disclose the C5.2 result and cannot run without a later hash-bound authorization.
ZERO.5 decision lineage
Every gate stays visible.
| Stage | Single changed question | Decision | ilXyr evidence |
|---|---|---|---|
| C0 | Governed corpus and lossless tokenizer | Selected byte-BPE512 | Backfill required |
| C1 | Native C training signal across seeds | Pass, no promotion | Backfill required |
| C2 | Atlas corpus scale at fixed parameters | Pilot pass | Backfill required |
| C3 | Evidence-linked claims, cloze, and retrieval | No-go | Backfill required |
| C3.1 | Record-safe interleaving and answer weight | No-go | Backfill required |
| C3.2 | Repaired pairs and task-balanced loss | No-go | Backfill required |
| C3.3 | Pair-atomic optimizer updates | No-go | Backfill required |
| C4.2 | Grouped data repair | No-go | Backfill required |
| C4.3 | Hard retrieval and cloze coverage repair | No-go | Public result; import required |
| C5.1 | 25% Braid structured-state text | No-go | Private result hash; import required |
| C5.2 | Verified factorized state-target loss | Decision withheld | Private terminal record |
| C6.1 | Shared state bottleneck in answer logits | Built; training blocked | Native prospective contract |
C6.1 preregistration
Make the state representation part of the answer path.
A 152-wide bottleneck feeds both the verified 752-tag state decoder and a zero-initialized answer-token adapter. It adds 193,032 parameters—232 fewer than the prior auxiliary ceiling—for 5,046,024 parameters total.
The completed C5.1 seed-0 run is the text-only control. The trained checkpoint must also beat itself with the bridge switched off by at least one point on both retrieval choice and pair-exact accuracy. This same-checkpoint ablation is what makes the mechanism testable.
The implementation and frozen gates exist, but this registry records C6.1 as blocked. No preflight or training is authorized until a separate record binds the exact ZERO contract hash to this ilXyr registration.
C5 causal chain
Content first. Learning bridge second.
C4.3 control: retrieval choice reached 53.76%; claim and retention gates passed, but retrieval and cloze kept the experiment at no-go.
C5.1 StateBridge text: the base remained 4,852,992 parameters. Exactly 25% of the matched stream became Braid structured-state text. Retrieval choice fell to 52.57%, pair-exact to 50.94%, and claim choice to 56.14%. Structured content alone was not enough.
C5.2 TargetBridge auxiliary: the authorized run finished with the C5.1 stream, seed, schedule, compute, and sealed-test boundary fixed. Its outcome, metrics, checkpoint hashes, and result hash remain private under the committed publication rule.
Braid input
BRAID-CORPUS-5.2-STATEBRIDGE is a scientific artifact.
50,000 worlds
Arithmetic, inventory, logic, spatial, program, comparison, pattern, and graph state transitions generated without an external model.
Two representations
A ZERO.5 byte-token text view and an inspectable symbolic uint16 view share canonical state truth.
Verified targets
Proof traces, admissible wrong states, mirrored choices, sealed test membership, and a 752-entry factorized target vocabulary are hash-bound.
Director policy
How the registry stays current.
- An already-started authorized run may finish. It is then imported with its real chronology.
- A new run requires an ilXyr registration before execution.
- External completed runs are never relabelled as forecast-risked ilXyr experiments.
- Plane-prefixed display names distinguish Braid Corpus 5.2 from ZERO5 C5.2.
- C5.2 stays private. C6.1 stands on a separate public prospective contract and makes no public claim about the C5.2 outcome.
- A built or blocked experiment is not an active run. C6.1 still needs a separate hash-bound authorization before compute.