Status date: 2026-09-13
The machine-readable contract
ABI_CAPABILITY_COMPILER_CAMPAIGN_CONTRACT_V1.json controls sequencing and
gates. ABI_CAPABILITY_COMPILER_CAMPAIGN_STATE_V1089.json records live status.
Evidence gates, not experiment counts or dates, close phases.
| Phase | Objective | Status |
|---|---|---|
| 0 | Definitions, baselines, and preregistration | COMPLETE |
| 1 | Capability inventory and normalized acquisition IR | COMPLETE |
| 2 | Strong matched LoRA and distillation baselines | MACHINE + FAIL-CLOSED HANDOFF READY; 0/21,000 EXTERNAL HUMAN PREFERENCES |
| 3 | Causal teacher-to-target acquisition | MACHINE EVIDENCE COMPLETE; BLOCKED BY PHASE 2 PREREQUISITE |
| 4 | Sufficient-information Pareto frontier | CERTIFIED FOR BOUNDED MACHINE-DEVELOPMENT SCOPE |
| 5 | Selective reconstruction and bounded exclusion | CERTIFIED FOR BOUNDED THREE-DOMAIN BEHAVIORAL EXCLUSION |
| 6 | Composition, portability, and multi-source provenance | CERTIFIED FOR BOUNDED THREE-PACKAGE SCOPE |
| 7 | Integrated teacher-relative and systems certification | CERTIFIED FOR BOUNDED SAME-MACHINE CPU/CUDA PRODUCT SCOPE |
| 8 | Independent hostile verification and release | LOCAL CLEAN-EXPORT REHEARSAL CERTIFIED; BLOCKED ON INDEPENDENT OPERATOR AND DIFFERENT CPU/CUDA HARDWARE |
The Phase 8 local packet is sealed by V1070-V1074: 52 exact files,
281,108,851 bytes, 14/14 manifest-verification gates, and 16/16 hostile tests.
It deliberately records phase8_certified: false. Only the independent run
described in docs/PHASE8_EXTERNAL_REPRODUCTION_V1.md can advance this phase.
V1088 exhausts the valid same-machine test: detached exact-commit worktrees, three declared payload restores, 42 host tests, first-run CPU/CUDA passes, 19/19 recomputed gates per device, 244 fresh cross-device identities, 488 fresh-to-sealed identities, and 15 hostile mutations. It does not change the external Phase 8 or Phase 2 human requirements.
R97 is additive evidence after that product campaign. It establishes a strict bounded prospective reasoning transfer into LayerCake (1,399/1,400 candidate versus 1,088 teacher, with raw recomputation, hostile verification, and exact live replay). It does not modify the sealed Phase 4–8 product bytes or satisfy the independent-human and different-hardware gates. The controlling external blockers therefore remain unchanged.
The sealed Phase 1 artifact contains 7,000 eligible English records, exactly 500 for each of 14 frozen capabilities. It preserves raw and normalized material, exact source provenance, authoritative generated token IDs, transformations, labels, hashes, and imported-information accounting.
Search, development, final, domain-isolation, and hostile suites are frozen and disjoint under the preregistered exact and near-duplicate checks. Domain reference inventories are evaluation-only. The original abstention shortfall, the first GPU runtime-conformance failure, and the failed mathematics reference outputs remain published negative evidence.
Phase 1 did not train or evaluate a student. It closes corpus normalization and adequacy only.
All registered baseline training, three-seed reproduction, paired statistics, checkpoint persistence, cold/warm runtime, and machine verification are done. V554 also freezes the exact scoring rule and verifies the blind-lock, attestation, unblinding, clustered-bootstrap, and score-replay workflow under 15 hostile tests. That is workflow readiness, not human evidence. V555 also supplies three separately initialized, answer-key-free, resumable rater sessions so humans need not edit JSONL; its 10 focused tests pass. The next—and only—Phase 2 action is external:
- Follow
docs/PHASE2_HUMAN_RATER_SESSION_V1.mdanddocs/PHASE2_HUMAN_RATING_HANDOFF_V1.md, then assign one complete blinded form to each of three independent human raters. - Keep the answer key and other raters' work hidden until all forms are locked.
- Preserve completed forms separately from the immutable templates.
- Compute the preregistered preference lower bound and rerun the hostile evidence verifier.
- Issue a Phase 2 certificate only if every exit condition passes.
Phase 2 exits only when each mandatory baseline is credible, converged within its locked budget, reproducible, and verifier-backed. No ABI-superiority claim is possible in this phase.
All currently registered Phase 3 machine work is complete. V551 binds the three-seed causal matrix, independent reconstruction, exact endpoint quality, three-host identity, acquisition accounting, physical sparse execution, persistent state, and exact-lineage fully CPU runtime. Do not launch another Phase 3 model or runtime experiment. The sole next step is completion and verification of the three independent blinded Phase 2 forms. If and only if Phase 2 passes, rerun the fail-closed prerequisite audit; an unconditional Phase 3 certificate may then open Phase 4 without new Phase 3 tuning.
The unavailable Phase 2 human ratings were deferred by explicit user direction, not passed. Both hash-bound Phase 3 development branches have executed and failed locked gates. V6 confirms that sequence conditioning is a large improvement, but semantic routing loses to the label-free control and does not beat the monolithic control. Phase 3 cannot issue a certificate or open Phase 4. Any successor must be separately preregistered around a shared sequence transformation; repeating either branch, adding data or steps, or sweeping nearby routes is barred.
- The V16 failure-attribution screen completed the ownership gate. The next bounded branch is self-prefix recovery in the ABI-to-LayerCake bridge. Do not modify the sealed LayerCake host or reopen broad representation sweeps unless that measured repair fails and an oracle-fit control proves a host ceiling. V17 and V18 have now failed, so the oracle-fit capacity control is the next non-promotional ownership gate; nearby recovery-loss variants are closed.
- V20/V21 and the expanded V22 oracle have now failed their fit/capacity gates. No further ABI acquisition, labeling, recovery-loss, or nearby bridge work is authorized. A separate LayerCake investigation must qualify the integration surface before a new ABI protocol can be registered. No runtime claim is inherited by either oracle bridge.
- V45/V46 qualify the sparse routing component, but V47/V49 fail integrated generation and fit. V50/V51 further show that training the entire existing generator topology reaches only 785/1,400 autonomously, 96.53% action fit, and 79.14% exact training sequences. The current topology/data combination is closed. Do not run nearby seed, step, data, token, or learning-rate sweeps. A future experiment requires a materially distinct preregistered realization mechanism after the external Phase 2 human gate is resolved.
- V52 completed that conditional screen and failed at 810/1,400 with 49 collapses. Hard isolation is closed. Preserved outputs identify held-out prompt-entity copying as the next representation target, so one separately preregistered Unicode-safe UTF-8-BPE pointer-supervision screen is permitted.
- The separate LayerCake host investigation passed a construct-only external direct-core interface. ABI V23 then failed the initial absolute screen at 504/1,400 with 77 collapses. Its fixed targets did not supervise available prompt-pointer actions. One new pointer-supervised representation screen is permitted under a fresh preregistration; V23 tuning, controls, and additional seeds remain barred.
- V24 completed at 601/1,400 with 139 collapses and 31 Unicode errors. Its paired +97 passes over V23 do not override the absolute gates. The pointer branch is closed; audit the exposed LayerCake UTF-8 validity gap separately and require a fresh measured ABI bottleneck before any new acquisition run.
- V26 now preregisters a no-training replay of V23/V24 to measure training fit, autonomous drift, held-out action generalization, and unrepresentable targets before any new architecture is selected.
- V26/V27 completed. Both checkpoints fail training fit; V23 represents 657/1,400 development targets and V24 941/1,400. The next gate is a no-training Unicode-atomic open-vocabulary bake-off requiring 100% target representability before another fit run.
- V28 found no qualifier: the training-derived alphabet misses
%and/in development and every candidate exceeds 320 actions at least once. Preserve this result and preregister only a development-independent universal-syntax successor; do not reopen training yet. - V29 reaches exact 8,400/8,400 target coverage with 4,634 fixed actions, but nine development targets exceed 320 actions. Attribute those records before proposing compact sublexemes or any host limit change.
- The separate LayerCake v2 host has closed that bounded UTF-8 contract gap for future artifacts. ABI's next step is still diagnostic, not training: measure sealed-checkpoint train fit, held-out fixed/pointer action accuracy, and autonomous prefix drift before selecting a materially different architecture.
- Phase 3 must show a causal ABI acquisition candidate beats parent, label-free, shuffled, bridge-only, and monolithic controls while passing autonomous quality and repetition gates.
- Phase 4 measures only a tested sufficient-information frontier with an adjacent lower failure.
- Phase 5 proves bounded exclusion and capability segregation.
- Phase 6 proves composition, exact artifact portability, provenance, and selected-only execution.
- Phase 7 certifies one teacher-free integrated candidate against its teacher and strongest matched baselines.
- Phase 8 recomputes everything in a clean environment and attacks evidence, accounting, labels, identities, teacher absence, and claim language.
ABI ends at independently validated acquisition artifacts and integration requirements. Runtime implementation and product certification belong to the separate LayerCake repository. Detailed LayerCake phase history is intentionally not duplicated here.
- Preserve negative evidence; never overwrite a failed phase with a successor.
- Do not repair or select against final-test failures.
- Do not launch a nearby sweep without a measured bottleneck and a bounded, preregistered repair.
- Do not call foreign-teacher acquisition lossless, ontology labeling exhaustive, or a tested information budget globally minimal.
- Do not start a locked phase early.