Status date: 2026-11-14
Documentation cannot widen a machine-readable certificate. Each claim below is limited to its exact source, artifact, implementation, benchmark, and evidence identity.
R108 timed out after 13/96 initial control responses under recorded GPU power/thermal throttling. No replay, recipient comparison or training ran. The original complete verifier rejects the run; the additive interruption record verifies only preservation/recomputation of the incomplete prefix and timeout. It establishes neither English-control qualification, acquisition success/failure, ABI superiority nor a LayerCake defect. No partial pass rate or product claim is promoted.
R107 completed nine active/neutral/reference cells and their exact fresh-process replays. The primary parent's residuals raise natural-text loss by +0.03224 passage BPB (95% CI [+0.02064, +0.04419]), but removing them worsens parent repetition. Distillation matches parent behavior on a fixed route not trained in R99; this cannot establish trained-route retention or superiority. No optimizer steps, new artifacts, fluent core, CPU/GPU advantage, native-host defect or product promotion follows. Historical gates are unchanged.
R106 scored 178/200 versus 190/200 on four external BLiMP families, with exact fresh-process raw replay. It performed no recipient comparison or training. A likelihood ranking with supplied target tokens is not autonomous English; no historical generation gate is replaced. Apparent external-label issues are preserved, not used for post-hoc promotion. No ABI acquisition, native LayerCake defect or English-core claim follows.
R105 scores 779/1,200 external scope cases and 62/64 author-specified approval cases. Both registered gates fail. Initial and fresh-process raw replay are byte-identical. The additive completion verifier requires all replay evidence; the weaker frozen stage-checker limitation and its hostile failure are preserved. This is not general evaluator qualification, English fluency, human review, acquisition superiority or a LayerCake defect.
R104 records 155 exact formal-obligation fixtures and 48 raw generations. The control scores 29/48; four task floors fail. Additional approval-scope limitations mean general semantic evaluator validity is NOT proven. Removing all approval flags yields at most 30/48, not a new qualification score. The old screen's 44/48 also fails a task floor. No recipients/training ran; no ABI-transfer, LayerCake-defect or product claim follows. The separate weight-space review reports external work and a possible future control design, not our own reproduction or superiority.
The R103 diagnostic recorded 41/48 raw control passes against 44/48, and three tasks missed their floors. The planned recipient comparison was not executed. A future-conditional false-positive in the frozen contradiction proxy is disclosed without changing any score. Neither a qualified English control, interface rescue, ABI-transfer failure nor native LayerCake defect is established. These mechanical screens do not determine broad English fluency.
The R102 study completed 6,144 finite updates and all fixed evaluations. Neither the renaming-policy contrast nor the repeated-data exposure contrast had positive 97.5% cluster-bootstrap lower bounds in both controls. No checkpoint passed the quality screen. The source qualified at 282/288 normalized passes and the new development references were 95/96 normalized, but source adequacy is not student acquisition success. The renaming policy changes distributions, not just data quantity; its original- example panel is not in-sample for augmented learners. No generic data/compute law, native LayerCake defect, fluent core, ABI novelty or product pass follows. The user authorized publication. Fresh public-only clones matched 782 assets and 833 frozen bindings, recomputed the identical negative decision, and passed 42 focused tests. This establishes public evidence availability and same-machine recomputation, not independent hardware, human quality or fresh live generation.
R101 found zero tolerance violations or argmax disagreements on the registered full/padded/cached logit comparisons, and exactly replayed 288 unassisted sequences. This is measured-path agreement, not universal LayerCake correctness. Supplying half the correct response did not improve full-model remaining-fact coverage (24/34 before and after; paired 95% CI [-0.14706, 0.14706]). Assisted joined-response gains cannot be promoted as autonomous quality. Training/development reference-fit differences warrant further acquisition controls but do not prove a sole cause. No new product or moonshot certificate is issued.
The fit diagnostic and full-scope control distinguish correct-prefix prediction from autonomous generation. The single R100 full-model control completed 384 finite updates but scored 20/48 on fresh development prompts, equal to retained LoRA, and 17/24 on training examples. All registered quality/fit/scope-sensitivity screens failed. This does not prove a LayerCake runtime regression, insufficient data, fundamental host incapacity or ABI superiority. Source qualification and successful optimization are not English-quality certification. No new artifact is promoted.
The R98 protocol and launch record register a bounded plan-supervised English experiment against matched LoRA and sequence-distillation controls. R98 failed its source gate at 206/240. C1 preserved every completion prefix but failed at 226/240, with emails at 22/30. Neither trained a student. The additive R99 record reports 230/240 normalized-source passes and a preserved FP16 gradient overflow. Its completed matched FP32 comparison failed: ABI plan 10/48, distillation 20/48, LoRA 27/48, versus a 44/48 gate. These facts establish neither a fluent English core nor ABI superiority. The inherited 87,710,428-parameter parent counts in full. Mechanical lexical/format screening cannot substitute for human fluency, semantic entailment, same-artifact CPU gates or a successor LayerCake seal.
The autoregressive-plan branch is closed. Its paired mechanical-score deficit against both trained controls is evidence against this specific plan mechanism, not evidence of universal ABI inferiority or a LayerCake architectural regression.
The independent 2026-09-14 audit returned
FULL_MOONSHOT_NOT_PROVEN. ABI accepts that verdict. R7, V1089, and R97 are
separate evidence lineages and must not be combined into one final-artifact
claim. The audit and receipts are preserved byte-for-byte under
reviews/independent_2026-09-14/.
The audit-response commit repairs the public replay surface; it does not add human judgments, independent hardware, broad English, automatic discovery, latent purity, global minimality, or universal LoRA/distillation superiority.
ABI may claim that the exact 530,050-parameter teacher-derived bridge, frozen before the R97 catalog existed, scored 1,399/1,400 on the registered two-hop nonce-reasoning surface versus 1,088/1,400 for the live pinned teacher, 60/1,400 for the unchanged LayerCake parent, and 17/1,400 for the random bridge. The paired candidate-minus-teacher interval is [0.20071, 0.24429]. It retained 1,087/1,088 teacher-correct rows, had zero collapses, used the physical sparse path on 1,400/1,400 rows, rejected 12/12 hostile mutations, and reproduced the raw evaluation byte-for-byte in a second live run.
This promotes only ABI-C4k for one bounded prospective reasoning family on one
compatible frozen LayerCake parent. It does not establish fluent English,
open-ended generation, automatic capability discovery, arbitrary teachers or
hosts, latent segregation, information minimality, systems superiority, or
general superiority over LoRA/distillation. Authority:
experiments/prospective_length_span_r97/CERTIFICATE.md and
results/prospective_length_span_r97/.
Exact historical root certificates and the V6 catalog have been restored so
the supported Phase 6-8 verifiers can replay from their frozen paths. The
retained R97 parent/tokenizer and V1089 Phase 7 handoff payloads are staged for
Git-LFS publication. Phase 5 clean replay remains unavailable because six
exact D0/L1 baseline tensors were deleted in the recorded 2026-08-29 cleanup.
The verifier continues to fail closed; the missing 1,191,079,704 bytes and
their hashes are listed in
reviews/independent_2026-09-14/REPLAY_SURFACE_REPAIR_MANIFEST_V1.json.
ABI may claim that, for four held-out synthetic capabilities learned in the registered ABI-native teacher substrate, four content-addressed 2,053-byte neural-state packages produced exact canonical UTF-8 teacher equivalence through frozen DistilGPT2, Pythia-160M, Qwen2.5-0.5B, and T5-base native output heads. Strict recomputation covers 90,112 raw observations, zero recipient optimizer steps, unchanged model/codec hashes, causal negative controls, exact removal, and exact restoration. Seven hostile cases fail closed. A fresh live execution using the exact revision-001 package files reproduces both raw JSONL files byte-for-byte and independently verifies.
ABI may not translate R11 into extraction from a conventional pretrained
transformer, English or natural-domain transfer, internalization into recipient
transformer blocks, arbitrary-model support, information minimality, LayerCake
product acceptance, or superiority over LoRA/distillation. A second acquisition
was functionally exact but produced different float32 package hashes, so
bit-reproducible acquisition is also not claimed. Authority:
docs/R11_NATIVE_NEURAL_CONSTRUCTION_RESULT.md and
results/native_isa_r11/.
ABI may claim that its exact 281,108,851-byte handoff reconstructed and ran from detached exact-commit ABI and LayerCake worktrees on the development machine after restoring only three manifest-declared untracked payloads. The unchanged CPU and CUDA runs pass 19/19 recomputed gates per device, all 244 fresh cross-device and 488 fresh-to-sealed functional identities match, 42/42 host tests pass, and 15/15 hostile mutations are rejected.
ABI may not call V1088 independent reproduction or Phase 8. The operator and
hardware fingerprint are unchanged, and no external attestation exists.
Authority: ABI_CAPABILITY_COMPILER_PHASE8_LOCAL_CLEAN_REHEARSAL_RESULT_V1088.json.
ABI may claim that the exact Phase 7 product has a locally verified, content-addressed external-reproduction handoff. The manifest binds 52 files totaling 281,108,851 bytes across the exact ABI and LayerCake source lineages; 14/14 read-only verification gates and 16/16 hostile readiness tests pass.
ABI may not call this an independent reproduction, Phase 8 pass, or release.
No independent operator, different hardware fingerprint, fresh external CPU
or CUDA run, signed attestation, or external hostile recomputation exists yet.
Authority: ABI_CAPABILITY_COMPILER_PHASE8_LOCAL_READINESS_RESULT_V1074.json.
ABI may claim that one exact teacher-free product—signed B40 English core
acb787...f25d plus the exact signed chemistry, civics, and Python packages—
passes the inherited Phase 4 machine-quality, Phase 5 behavioral-isolation,
Phase 6 composition, and fresh integrated CPU/CUDA gates on the registered
machine. Independent recomputation confirms 244/244 cross-device output/token
identities, zero V2 repetition collapse, persistent state, selected-only
specialist execution, unchanged artifacts, and zero receiver learning.
On the registered CPU it measures 274.319 UTF-8 bytes/s, 13.2387x its frozen comparator, 101.39% same-core throughput retention, 3.121 s cold TTFT, and 545,308,672 bytes peak RSS delta. On CUDA it measures 881.358 bytes/s, 13.5139x its registered L1 systems comparator, 97.84% same-core throughput retention, and 505,427,456 peak allocated bytes.
This is a bounded same-machine claim. It is not independent-hardware
reproduction, human preference, release certification, arbitrary-hardware
dominance, a global minimum, or universal ABI superiority over LoRA or
distillation. Authority: ABI_CAPABILITY_COMPILER_PHASE7_CERTIFICATE_V1.json.
ABI may claim that the exact registered chemistry, civics, and Python packages compose without interference on three fresh compatible LayerCake host initializations with distinct B40 English-core payloads. Independent recomputation confirms 900/900 selected specialist outputs, 900/900 structured three-domain components, 300/300 conflict quarantines with zero execution, 300/300 preserved English outputs, exact cross-host output identity, complete record-level deletion lineage, selected-only physical execution, and zero V2 repetition collapses. Seventeen hostile tests pass.
This is a bounded same-machine portability and composition claim. The archive contains three pinned source models, while these deployed packages all select Phi-3; deployed multi-source quality is not proven. Arbitrary-machine portability, latent semantic purity, completed human review, Phase 8 release, and universal ABI superiority remain unproved.
ABI may claim that the exact registered three-seed B40 LayerCake product selectively reconstructs the signed chemistry, civics, and Python packages on 300/300 frozen final cases per seed, while missing packages explicitly fail closed on 300/300 cases per seed with zero English-core or cake execution. The same runs preserve 100/100 English outputs, exact core identity, exact package remove/reinstall outputs, and all registered spoof, cross-domain, and quarantine boundaries. The nine ABI/L1/D0 result trees were independently recomputed and 13 hostile tests passed.
This is behavioral exclusion at an immutable outer-label control plane, not a claim that specialist facts are absent from all English-core weights. It is bounded to the exact artifacts, domains, seeds, suite, and host. It does not establish multi-domain composition, multiple-source provenance, Phase 6, completed human review, release, or universal ABI superiority.
ABI may claim that, for the preregistered exact five-route v25 architecture and the tested B20/B40 development budgets, B40 is the smallest stable passing tested budget and B20 is its adjacent-lower all-seed failure. At equal B40 sequence information, ABI passes the locked quality contract on all three seeds while the registered L0, L1, and D0 controls do not. The same ABI artifact is teacher-free and preserves the separately verified CPU speed gate. On the registered GPU runtime suite it has 13.8126x L1's median byte throughput, 0.1881x L1's median TTFT, and 0.0332x L1's active tensor bytes.
The authority is ABI_CAPABILITY_COMPILER_PHASE4_CERTIFICATE_V1.json, backed
by the raw V1014 verifier result. This claim is bounded to the named artifacts,
budgets, suites, seeds, and systems. It does not establish a global minimum,
human preference, final-test result, general product superiority, method
superiority, novelty, or universal superiority over LoRA or distillation.
Phase 7 machine scope is certified. Phase 8 independent reproduction and Phase 2's external human gate remain incomplete.
ABI may claim that the exact route-isolated endpoint has completed and passed all registered Phase 3 machine gates. The bounded claim includes three-seed causal replication against label-free, shuffled-target, no-update, and monolithic controls; 1,393/1,400 selected-seed autonomous passes with zero collapse; three byte-identical host initializations; teacher-free canonical execution with persistent state and physical one-expert rank-16 activation; and exact-lineage fully CPU performance of 12.835x median UTF-8 throughput, 10.565x paired lower-95 throughput, 0.0254x TTFT, 0.7169x peak RSS, and 99.12% parent-speed retention.
ABI may not claim an unconditional Phase 3 certificate. Phase 2 is a required
predecessor and currently has 0/21,000 completed blinded preferences. ABI also
may not claim a smallest sufficient information budget, final-test success,
universal teacher improvement, or superiority over LoRA/distillation. Phase 4
is locked. The controlling evidence is
ABI_CAPABILITY_COMPILER_PHASE3_FINAL_CERTIFICATE_AUDIT_RESULT_V551.json.
ABI may additionally claim that the external Phase 2 rating workflow is
operational and fail-closed under
ABI_CAPABILITY_COMPILER_PHASE2_HUMAN_RATING_READINESS_AUDIT_V554.json:
scoring was frozen at 0/21,000 preferences, raw forms must lock before
unblinding, and 15 hostile readiness tests pass. ABI may not translate this
into a claim that human ratings exist, that rater humanity or independence was
machine-proven, or that Phase 2 passed.
ABI may claim that V555 provides a blind-safe resumable rater workflow: all three production forms initialize as distinct 7,000-row sessions without an answer key, and 10 focused integrity tests pass. This is an operational claim, not evidence that a human completed any judgment.
V45 scientifically qualifies the bounded sparse capability router across three fixed seeds: each seed passes 1,400/1,400 original prompts, bodies, and metadata rejections, with 100/100 for every capability. This is a routing claim only.
V47 does not qualify the integrated product. Its routed generator scored 760/1,400 with 20 collapses and failed teacher noninferiority. V49 shows the same checkpoint fits only 90.56% of acquisition actions and 72.37% of complete acquisition sequences. ABI therefore cannot claim fluent English transfer, LayerCake host certification, Phase 3 completion, or superiority over LoRA or distillation.
V50 also does not qualify the integrated product. Allowing all 4,174,280 generator parameters to train raised the locked autonomous score only to 785/1,400 and produced 36 collapses. V51 measured 96.53% action accuracy and 79.14% exact acquisition sequences. This closes only the current router/data/representation/generator-topology combination; it neither proves a fundamental ABI ceiling nor authorizes Phase 4.
V52 and V54 do not qualify it either. Hard-routed experts reached 810/1,400 with 49 collapses; Unicode-safe BPE pointer supervision reached 785/1,400 with 40 collapses. These results close capability isolation and the tested limited pointer policy. They support studying richer teacher representations, but do not prove that logits or hidden activations will work.
ABI_MOONSHOT_CERTIFICATE_V2.json remains valid for its exact 1,700-observation
catalog, teacher-free candidate, three narrow domain packages, tested nested
budget, runtime, and performance comparison.
The later ABI_POSTCERT_GENERALIZATION_AUDIT_DECISION.json controls the broader
product conclusion: the same candidate scored 0/28 on novel-form English while
frozen Phi-3 scored 19/28. Therefore v47 is not certified as a generally fluent
English core or as source-quality parity.
ABI_TEACHER_TO_LAYERCAKE_GRAMMAR_PILOT_V87_DECISION.json proves that one real
teacher-derived grammar mapping causally improved the intended LayerCake route
relative to both an unchanged parent and a matched shuffled-target control.
The paired 95% confidence-interval lower bounds were positive for the locked
likelihood comparisons, and autonomous exact corrections improved from 0/64
for both controls to 14/64 for the real mapping.
V87's formal pilot failed: positive-margin accuracy was 34/64 against a 48/64 gate, and 23/64 outputs exhibited repetition collapse. This is a causal construct signal, not reliable transfer, fluent English, or a deployable model.
ABI_TEACHER_RECORD_LABELING_PHASE2_CERTIFICATE_V89.json certifies the locked
English, chemistry, civics, mathematics, Python, and quarantine ontology on a
120-row source-record-disjoint holdout:
- exact destination-plus-domain accuracy: 117/120, or 97.5%;
- English precision: 100%;
- specialist records mislabeled as English: zero;
- English recall: 90%;
- known-domain macro-F1: 99.36%;
- known-record capability accuracy: 97%; and
- quarantine recall: 100% across all three adversarial families.
The verifier independently replayed raw labels, metrics, gates, hashes, provenance, and authoritative token counts and rejected five deliberate mutations. This is a point-estimate pass with 20 observations per class; it is not exhaustive domain discovery or a population-wide guarantee.
ABI does not claim or duplicate runtime-host results. The intended consumer and its evidence live in the separate LayerCake repository. A future ABI artifact must be independently accepted and recertified by its target host.
The LayerCake repository has separately passed a construct-only signed direct neural core interface. ABI V23 generated through that interface, but scored only 504/1,400 with 77 collapses. The correct claim is that the host construct can carry an external artifact while this ABI acquisition representation fails quality. It is not a LayerCake regression, a Phase 3 pass, or evidence of ABI superiority over LoRA or distillation.
V24 establishes only a paired representation effect: source-pointer target supervision raises the matched development score from 504 to 601/1,400 (95% CI for the paired delta +4.43 to +9.43 points). It also raises collapses from 77 to 139, introduces 31 Unicode errors, and lowers prompt grounding from 50 to 39. The branch fails and cannot support a superiority claim.
The external LayerCake lc-direct-neural-core/2 host is now construct-ready
and Unicode-safe, but ABI does not own or inherit that evidence. It removes one
future host-contract confound; it does not improve or reinterpret V23/V24.
V26 establishes a narrower failure attribution. V23 and V24 fail the locked training-fit thresholds, autonomously reproduce only 136/280 and 167/280 fixed training responses, and can represent only 657/1,400 and 941/1,400 held-out teacher targets. Both action surfaces reject all coherence and fact-free- reasoning targets. This supports an ABI fit/representation/generalization failure; it does not establish a LayerCake regression, a Phase 3 pass, or an ABI advantage over LoRA or distillation.
V28 establishes that the four preregistered Unicode fallback surfaces do not qualify. Each exactly represents all training targets but misses six development targets and exceeds the 320-action bound. The result supports a universal syntax alphabet selected independently of development outputs; it does not prove learned open-vocabulary generation or a host limitation.
V29 establishes complete representability on the bound 7,000 training and 1,400 development targets using a development-independent syntax alphabet. It does not qualify because nine development targets exceed 320 actions. This is representation feasibility evidence only, not learned quality or transfer.
ABI_CAPABILITY_COMPILER_PHASE0_CERTIFICATE_V1.json certifies the initial
teacher, immutable LayerCake host, mandatory LoRA/distillation/ABI systems,
data boundaries, statistics, numeric gates, accounting, and stop rules. It
contains no new training and establishes no comparative quality result.
ABI_CAPABILITY_COMPILER_PHASE1_CERTIFICATE_V1.json certifies a bounded,
immutable 7,000-record English acquisition IR with exactly 500 eligible records
for each of 14 capabilities. Records bind raw and normalized forms,
authoritative generated token IDs, source revision, provenance, destination,
transformations, and hashes.
The frozen split audit reports zero exact prompt overlap and zero detected near-duplicate clusters. The English inventory contains zero specialist acquisition records. Domain inventories are evaluation-only. The V1 abstention failure, fresh V2 successor, and failed mathematics references remain explicit negative evidence.
This is a data-suitability claim only. No model was trained in Phase 1.
The complete preregistered machine campaign is content-verified under
results/abi_capability_compiler_phase2/machine_evidence_v1.json.
- T0: 1,237/1,400 functional passes and 64 collapses.
- L0 three-seed mean: 94.60% functional passes; source retained; oracle route; 0.84× T0 warm bytes/second.
- L1 three-seed mean: 93.86%; 4,200/4,200 prompt routes correct; source retained; 0.83× T0 warm bytes/second.
- D0/D1/D2 three-seed means: 6.69%, 0.50%, and 2.10%; severe collapse despite 9.83×, 5.94×, and 6.84× T0 warm bytes/second.
These results establish credible matched baseline behavior and a measured quality-speed split. They do not establish ABI transfer, candidate quality, or superiority. Phase 2 remains incomplete because the required three independent blinded human forms have not been completed.
ABI_CAPABILITY_COMPILER_PHASE3_CONDITIONAL_OPEN_V1.json records the user's
explicit decision to proceed while raters are unavailable. It authorizes only
the now-completed A0-A4 development campaign. It is not a Phase 2 pass, a
Phase 3 certificate, or permission to access final data.
results/abi_capability_compiler_phase3/conditional_decision_v3.json records
the corrected V4 result. All A0-A4 systems consumed the same successful
28,000-record sequence. A0 scored 379/1,400 with 150 repetition collapses.
Paired stratified-bootstrap pass-rate differences favored A0 over A1 by 19.00
points (95% CI 16.79-21.21), A2 by 7.71 (5.43-10.00), A3 by 26.79
(24.71-28.86), and A4 by 9.07 (7.00-11.14). This supports a bounded causal
claim for the actual labeled teacher payload and segmented routing.
The same A0 trails T0 by 61.29 points (95% CI -63.64 to -58.86), has two capabilities at 0/100, and fails every locked absolute quality family, teacher-noninferiority, and repetition gate. The branch is not promoted, its remaining seeds are not authorized, Phase 3 is uncertified, and Phase 4 is locked. This result does not establish fluent transfer or superiority over LoRA or distillation.
results/abi_capability_compiler_phase3_sequence/conditional_decision_v1.json
records the complete V6 initial-seed result. B0 scored 1,148/1,400 with 43
collapses. It beat deranged B2 and no-response B3, showing sensitivity to real,
correctly paired teacher responses. It lost to label-free B1 by 5.43 points
(95% CI -7.50 to -3.43), did not beat monolithic B4, and trailed T0 by 6.36
points (95% CI -8.64 to -4.07). Semantic routing is therefore not established
as causal for the gain. The branch failed, remaining seeds are prohibited,
Phase 3 is uncertified, and Phase 4 is locked.
The V9 diagnostic bypassed components in memory without training or persisting a checkpoint. Removing the route embedding changes B1 by +0.07 points (95% CI -0.14 to +0.29). Removing output cakes changes it by -5.43 points (95% CI -7.07 to -3.79) and reduces collapses from 60 to 49. This localizes dependence inside B1 only; it does not promote B1 or predict retraining.
V11 C0 scores 1,207/1,400. It beats label-free C1 by 8.36 points, deranged C2 by 68.93, no-response C3 by 85.50, and monolithic C4 by 7.14; every paired 95% lower bound is positive. Against T0 it is -2.14 points with 95% CI -4.21 to -0.07, passing the locked -5-point noninferiority margin. C0 nevertheless has 51 collapses and seven capabilities below 90/100, so Phase 3 remains failed.
On a preregistered balanced 280-prompt diagnostic, C0 reaches 82.63% teacher-forced token accuracy versus 27.61% for the no-payload control. The exact LayerCake repository, checkpoint, and sealed certificates pass identity checks. One wrong C0 self-prefix token raises 16-token continuation NLL by 1.77x and lowers accuracy by 6.96 points. This supports one bridge self-prefix-recovery experiment. It does not prove Phase 3, a LayerCake capacity ceiling, or ABI superiority over LoRA or distillation.
V17's self-prefix candidate scored 1,185/1,400 versus 1,208/1,400 for its compute-matched control. The paired 95% interval for the -1.64-point difference is -2.79 to -0.43 points. Its 64 collapses also exceed the control's 56 and the parent's 51. This rejects arbitrary first-mismatch recovery; it does not reject the causal ABI payload signal or establish a LayerCake failure.
V18 S0 scored 1,213/1,400 versus S1's 1,208, but the paired 95% interval for the +0.36-point difference is -0.29 to +1.00 points. Its 52 collapses do not improve on V11 C0's 51. This closes the bounded recovery-loss family. It does not prove a host ceiling; that requires the separately governed oracle-fit diagnostic and can never support promotion itself.
The contaminated V20 upper bound scored 1,229/1,400 with 89 collapses. V21 teacher-forced accuracy was 94.11% at 0.2631 NLL, missing both locked fit thresholds. This assigns the next experiment to bridge fit, optimization, or expressivity. It does not invalidate sealed LayerCake, prove a fundamental host ceiling, or permit a quality or superiority claim.
V22 expands the sequence rank from 128 to 256 and the shared-output rank from 64 to 128 while keeping the sealed LayerCake host byte-identical. It scores 1,248/1,400 with 75 collapses on the exact development pairs used to train the bridge. This fails the preregistered 99% aggregate, 95%-per-capability, and zero-collapse gates. It gains only 19 passes over V20 and does not establish a sufficient integration architecture.
This result does not test or disprove ABI extraction or labeling, does not show a LayerCake regression, and does not prove a fundamental host representational ceiling. It requires a separately governed LayerCake integration-interface investigation before ABI acquisition experiments can resume. The expanded candidate inherits no speed, memory, TTFT, sparsity, quality, or superiority claim.
- The V87 real-versus-shuffled causal pilot is implemented and evidence-bound.
- The V89 labeler combines deterministic fail-closed risk screening with a frozen int8 CUDA semantic classifier on every record.
- Labels bind destination, domain, capability, knowledge class, source record, ontology, semantic output, deterministic screen, and implementation hash.
- Unknown-domain, cross-domain, spoofed, malformed, inconsistent, or low-confidence material fails closed to quarantine under the tested method.
- LayerCake was not invoked during V89 and no training artifact was composed.
- The Phase 1 catalog, extractor, IR builder, IR verifier, certificate verifier, and hostile mutation suite are implemented and evidence-bound.
- Phase 2 implements and verifies packed sequence data, uninterrupted top-64 teacher logits, LoRA/routed-adapter and same-size student baselines, three-seed checkpoint persistence, paired statistics, genuine-cold and warm runtime, and a blinded counterbalanced human-rating packet.
- Phase 3 implements a corrected paired-successful-step sampler, exact sample sequence hashes, two five-control architecture branches, autonomous development evaluation, and immutable paired statistical analysis.
ABI has not yet established:
- broad English fluency from the certified normalized acquisition artifact;
- sufficient data for natural English generalization;
- the smallest passing fluent English data or model budget;
- reliable operational teacher-to-LayerCake transfer;
- teacher-relative broad generation quality in a final LayerCake;
- general domain discovery from arbitrary weights;
- general Python, chemistry, civics, or mathematics extraction;
- robust multi-source knowledge merging;
- preservation of LayerCake's sealed CPU/GPU advantages by an ABI-derived core;
- external laboratory reproduction; or
- one coherent public final artifact whose quality, runtime, portability, and external-review evidence all come from the same lineage;
- Assumption-free theorem-level mathematical proof over all possible models.
- Automatic transfer to a new target model without an ABI compatibility certificate.
ABI also has not established superiority over LoRA, routed adapters, or any
form of distillation. ABI_CAPABILITY_COMPILER_CAMPAIGN_V1.md defines the
matched baselines and evidence required before such a bounded product claim is
allowed.
- “Lossless” applies only to exact LayerCake archive, manifest, tensor, and installed-payload identity—not acquisition from a foreign teacher.
- “Minimum” means the smallest passing preregistered tested nested budget with an adjacent lower failure—not a global minimum.
- “Pure English” means a bounded corpus and behavioral isolation result—not proof that neural weights contain literally zero world knowledge.
- “Domain labeling” means classification within a declared ontology—not exhaustive discovery of everything a teacher knows.
- “Independent” means disjoint evidence and adversarial replay unless an external organization is explicitly involved.
- “ABI product superiority” requires quality noninferiority plus strict preregistered advantages on the complete deployment and capability-control contract. It never means universal algorithmic superiority.
- “Stronger than the teacher” is allowed only for a named frozen scope with a positive paired confidence interval and noninferiority on every protected dimension. It is claimed only for R97's bounded prospective reasoning surface, never for English, arbitrary domains, or the full product.
See CURRENT_PROJECT_STATUS.md and RESEARCH_STATUS_AND_GAPS.md for the active
phase and unresolved questions.
FORMAL_UNIVERSAL_TRANSFER.md records the impossibility boundary and the
conditional assumptions required for any bounded transfer certificate. The
capability-compiler campaign does not weaken or supersede that theorem.