ABI

ABI is a research codebase for acquiring, labeling, packaging, validating, and eventually minimizing capabilities extracted from open-weight teacher models. It is designed to produce immutable capability artifacts that a separate LayerCake runtime can install, compose, route, and execute.

The repositories have deliberately separate responsibilities:

• ABI: teacher qualification, probing, labeling, segregation, provenance, information accounting, artifact construction, and verification.
• LayerCake: hosting, installation, transfer between compatible hosts, routing, orchestration, and inference performance.

Current release status

An independent audit dated 2026-09-14 returned FULL_MOONSHOT_NOT_PROVEN. That is the controlling whole-moonshot verdict. The untouched audit, its machine-readable summary, and its four receipts are preserved under reviews/independent_2026-09-14/. R97 is the strongest bounded local transfer result; R7 remains the strongest publicly reconstructed release. Neither may be combined with V1089 to claim one coherent final-artifact proof.

The latest technical validation is R7:

https://github.com/Yoder23/abi/releases/tag/abi-final-validation-v2-repaired-r7-2026-08-30

R7 passes a bounded capability-runtime/conformance proof across LayerCake v25, Qwen2.5-0.5B, and Pythia-160M using four immutable packages: English, Python, civics, and chemistry.

This is not a claim that ABI already extracts a fluent minimal English core from any teacher. Human ratings, different-hardware reproduction, registered minimum-information certification, teacher-extraction quality, and comparison with LoRA/distillation remain open.

Later local compiler campaign

The controlling public release is still R7, but it is not the end of the local research record. The separately sealed V1089 compiler campaign has completed all currently registered machine work through Phase 7 and a clean same-machine Phase 8 export rehearsal:

Phase	Local evidence status
Phase 0	Complete
Phase 1	Complete
2	Machine evidence and blinded handoff complete; 0/21,000 independent-human preferences
3	All registered machine gates pass; unconditional certification remains dependent on Phase 2
4	B40 is the smallest stable passing tested budget among B20/B40 for the exact five-route host
5	Three-domain selective reconstruction and behavioral exclusion certified
6	Three-package composition, portability, and provenance certified on three fresh host initializations
7	Same exact integrated product certified on the registered CPU and CUDA device
8	Clean same-machine export rehearsal certified; independent operator and different hardware still required

The B40 English host passed the registered machine-quality contract in all three 1,400-observation screens (4,200 observations total). The exact seed-104729 core later used by the Phase 7 product scored 1,381/1,400 across fourteen capabilities with zero collapse, and its paired advantage over the teacher was 10.29 percentage points (95% bootstrap lower bound 8.64 points). The integrated Phase 7 product retained the same core and three packages while measuring 13.24x the optimized CPU-transformer throughput and 13.51x the registered GPU comparator throughput. These are bounded same-machine results, not human, arbitrary-hardware, global-minimum, deployed-multi-source, or universal ABI-over-LoRA/distillation claims. The authoritative state is evidence/current/ABI_CAPABILITY_COMPILER_CAMPAIGN_STATE_V1089.json.

The latest additive local result is R97. A frozen 530,050-parameter teacher-derived reasoning bridge scored 1,399/1,400 on a separately sealed prospective catalog, versus 1,088 for the live pinned teacher, 60 for the unchanged LayerCake parent, and 17 for a randomized bridge. It retained 1,087/1,088 teacher-correct rows, executed the registered sparse path on every row, showed zero collapse, passed strict raw recomputation and 12/12 hostile mutations, and reproduced its raw evaluation byte-for-byte in a second live run. See experiments/prospective_length_span_r97/CERTIFICATE.md. R97 is a bounded two-hop reasoning transfer result; it is not evidence of a complete fluent English substrate, arbitrary-domain extraction, or universal superiority over LoRA/distillation.

The independent-audit response restores exact root-relative compatibility copies of historical certificates and the V6 catalog so the frozen Phase 6-8 verifiers replay without relying on test order. It also puts the exact local R97 parent and V1089 Phase 7 handoff payloads on the Git-LFS publication surface. Phase 5 clean replay remains blocked: six baseline tensors totaling 1,191,079,704 bytes were deleted during the recorded 2026-08-29 storage cleanup and were not recoverable from local LFS or other working copies. This gap is explicit in the review manifest; no verifier is weakened around it. See docs/INDEPENDENT_REVIEW_HANDOFF_2026-09-14.md.

The additive R8 native-neural-transfer campaign is a Level 0 negative result. Its strongest package-only recipient interface produced AFTER−BASE +0.001953 across 1,024 paired Pythia rows (95% bootstrap CI -0.019531 to +0.024414). Canonical extraction was exact, but recipient behavior did not transfer. The held-out secret was never revealed, and R8 does not change the R7 release. The final public verifier rejected 7/7 hostile mutations and ignored one forged scientific boolean without changing the NO verdict (8/8 expected outcomes).

R9 then tested the backend/compiler hypothesis directly with an intentionally capability-specific Pythia neural backend. V1 failed to fit its own training rows. The preregistered v2 repair exposed both embedding and final recipient states and trained a 1,039,880-parameter backend for 10,000 steps. It still recomputed at only 12.85% training accuracy and 12.5% unseen-depth AFTER accuracy, versus 7.32% BASE; ZERO also reached 12.5%, failing causality. All 8,192 evaluation and 288 training observations replayed live, and 5/5 hostile controls were rejected. R9 therefore closes this recipient-state GRU branch without opening the universal-backend experiment or changing R7.

R10 then connected the synthetic R8 extraction to a clean package slot and three frozen recipients. Its runtime-owned copy/paste component passed: four 2,040-byte AFTER packages produced 1.0 accuracy after paste and restore across 61,440 Pythia, Qwen2, and T5 rows; removal exactly matched BASE, all negative controls were at most 0.15625, recipient parameters were unchanged, and a fresh live run reproduced every row byte-for-byte. The overall R10 contract is still negative because the source model's native decoder reached only 0.59375 to 0.65234375 against the registered 0.99 source gate. See docs/R10_COPY_PASTE_RESULT.md. This proves a bounded canonical-runtime component, not LayerCake integration, lossless source behavior, English/domain extraction, or native neural transplantation.

R11 now supplies the missing bounded neural-ABI construction proof. Four held-out synthetic capabilities were learned as compact neural transition states, packaged into four 2,053-byte content-addressed artifacts, and executed through frozen DistilGPT2, Pythia, Qwen2, and T5 native output heads. AFTER and RESTORED were exactly 1.0 on all 90,112 registered source-plus-recipient rows; all teacher/recipient canonical UTF-8 outputs matched, negative controls were at most 0.19140625, all removal interventions exactly restored BASE, recipient optimizer steps were zero, and model/codec hashes did not change. A fresh live replay of the exact revision-001 package bytes reproduced both raw JSONL files byte-for-byte, and 7/7 hostile mutations failed closed. See docs/R11_NATIVE_NEURAL_CONSTRUCTION_RESULT.md.

R11 is deliberately narrow. Its teacher substrate is ABI-native and synthetic; it does not prove extraction from a conventional pretrained LLM, English or natural-domain transfer, internalization into recipient transformer blocks, or LoRA/distillation superiority. R7 remains the controlling public release.

R12-A has now exercised the missing frontend on one bounded public synthetic capability using conventionally trained Qwen2.5-0.5B. Its best native source checkpoint reached 510/512 = 0.99609375, two rows short of the registered exact source gate. Its 24 atomic probes were exact, and the fixed zero-parameter extractor emitted an unchanged 2,053-byte R11 package whose executor was exact on 512/512 unseen, prompt-disjoint compositions without the teacher at execution time. This is a bounded public frontend construction pass and a strict teacher-gate failure, not held-out R12 certification. See docs/R12_FOREIGN_FRONTEND_PUBLIC_RESULT.md.

R13-B now passes a separately preregistered held-out Track B control. Four conventionally LoRA-trained Qwen sources each exposed 24/24 atomic transitions; the fixed zero-parameter compiler emitted four 2,053-byte packages that scored 2,048/2,048 against the hidden capability oracle and 6,144/6,144 across frozen Pythia, Qwen2, and T5 recipients. Seven fresh-process replays were byte-exact, and the repaired verifier passed 9/9 hostile controls. Native Qwen matched only 372/2,048 long compositions, so this proves bounded enumerable capability canonicalization—not behavioral transplantation. Bulk replay assets remain unpublished. See docs/R13_BOUNDED_CAPABILITY_EXTRACTION_RESULT.md.

R14 then removed the exhaustive-table shortcut. Its fixed frontend received only 256 mixed, non-atomic Qwen probability observations per capability and was tested on 10,000 unseen-depth cases plus 1,000 order counterfactuals drawn from a 4.64-billion-case space. It recovered one of three fresh latent programs exactly; the other two packages reached only 0.2449/0.2520 and 0.2657/0.2340 on unseen/order-counterfactual oracle evaluation. Three fresh source replays were byte-exact and 7/7 hostile controls passed. R14 is therefore a bounded partial result and failed certification—not a non-exhaustive extraction pass. See docs/R14_NON_EXHAUSTIVE_CAPABILITY_RESULT.md.

R15A materially changed the frontend from behavior probing to controlled foreign neural-state access. A generic decoder qualified on 320 public Qwen learning events received only anonymous 7,168-element before/after effective output-weight deltas. On eight fresh committed capabilities it recovered 8/8, emitted eight 2,053-byte R11 packages, and scored 1.0 over 80,000 unseen plus 8,000 counterfactual cases. The packages produced 1.0 AFTER/RESTORED behavior across unchanged Pythia, Qwen2, and T5 recipients; full-state controls were at most 0.265. Strict verification passed, fresh live execution reproduced all 135,168 recipient rows byte-for-byte, and 9/9 hostile mutations were rejected. See docs/R15_FOREIGN_NEURAL_STATE_RESULT.md.

R15A is the first bounded foreign learned-state recovery pass, but it is not the full ABI moonshot. The source capabilities were deliberately taught with complete atomic supervision, and extraction used a before/after delta. The result does not establish extraction or labeling of pretrained English/domain knowledge, teacher-quality natural generation, after-only extraction, LayerCake English-core ingestion, minimality, or LoRA/distillation superiority.

R15B removes the source-training requirement for one deliberately narrow representation prerequisite. An unchanged Qwen2-7B-Instruct generated 24/24 registered atomic answers under four secret slot mappings. A frozen, zero-parameter decoder operating in a Linux pivot-root/no-network capsule recovered all four transitions from pre-answer residuals and frozen digit-head rows. The resulting four 2,053-byte R11 packages scored 40,000/40,000 on the registered deep oracle and reproduced 33,792 recipient rows across Pythia, Qwen2, and T5; strict and live verification passed and 10/10 hostile mutations were rejected. See docs/R15B_PREEXISTING_REPRESENTATION_RESULT.md.

R15B is not English or open-world domain extraction. Its prompts explicitly state the arithmetic operation, source reasoning exposes the answer before the captured terminal digit, semantic labels are externally registered, and the package's deep generalization comes from the registered affine inductive bias. The full ABI moonshot therefore remains open, and R7 remains the controlling published release.

R16 now passes a preregistered bounded local factual-acquisition and semantic- segregation test. From an unchanged Qwen2-7B-Instruct source it recovered 16/16 held-out facts across chemistry and geography, emitted two immutable packages, and matched 48/48 disjoint evaluation questions with the teacher absent. Target-only execution was 48/48, other-domain and removed-package conditions abstained 48/48, a rotated-score control scored 0/48, live evidence replayed byte-exactly, and the expanded verifier rejected 21/21 hostile mutations. A blind review passed with no Critical or High finding. See docs/R16_FACTUAL_SEMANTIC_RESULT.md.

This is a structured closed-candidate memory result with a registered ontology and candidate vocabulary. It is not autonomous open-world discovery, fluent English extraction, teacher-quality free generation, native neural transplantation, production LayerCake ingestion, minimality, or superiority to LoRA/distillation. A second preregistered 16-fact selection, bound to the repaired public protocol before reveal, independently repeated every gate and expanded combined held-out coverage to 27 distinct facts. Blind review of the exact replicated seal passed with zero findings; durable publication and clean public reconstruction remain open. The full ABI moonshot therefore remains open, and R7 remains the controlling published release.

R17 now tests the next, explicitly separate step: teacher-derived compositional English surface realization rather than factual lookup. Its first frozen public source interface failed at 38/72 extraction and 28/48 evaluation outputs. Its sole bounded repair improved this to 67/72 and 47/48 but still missed the locked 120/120 source prerequisite, before ABI compilation or LayerCake was invoked. The exact-template branch is closed; its successor must separate functional language quality from raw string identity. See docs/R17_LINGUISTIC_REALIZATION_RESULT.md.

R18 replaces that exact-string boundary with a frozen feature-factorized grammar package. Reusing the same R17-v2 teacher outputs, its 3,881-byte package scored 48/48 on lexically disjoint functional realizations, versus 47/48 for the teacher and 46/48 for independent per-signature consensus. It had zero regressions on teacher-correct rows, a 0/48 mood-permutation control, and 48/48 removal abstentions. Strict recomputation, 12/12 hostile mutations, and a fresh byte-exact physical replay passed. The preregistered hidden lexical replication then failed at 46/48, with two source-correct plural questions regressed by unsafe cross-number auxiliary pooling. This remains a bounded public prerequisite, not a certified realization mechanism, unrestricted English, or the ABI moonshot. See docs/R18_FACTORIZED_REALIZATION_RESULT.md.

R19 is a materially different successor: it learns same-number structural contrasts between positive and negative forms instead of pooling across number. On the disclosed R17 and failed-R18 evidence sets it scored 96/96 total functional rows with zero source-correct regressions, 0/96 controls, and 96/96 removal abstentions. Its separately committed 20-lexeme hidden replication then scored 48/48, versus 43/48 for the unchanged teacher and 48/48 for independent modal consensus, with zero teacher-correct regressions, 0/48 control behavior, and 48/48 removal abstention. The 4,263-byte package contains no source or host parameters; 22/22 hostile mutations were rejected and a fresh physical run reproduced the package and evaluation byte-for-byte. R19 therefore passes only the registered bounded surface-realization claim. It is not unrestricted English, prompt understanding, conversation, autonomous labeling, arbitrary- domain extraction, LayerCake ingestion, minimality, or LoRA/distillation superiority. See docs/R19_CONTRASTIVE_REALIZATION_RESULT.md.

R20 then tested whether that structural approach could scale to raw natural instructions for prose, summary, email, bullets, clarification, and abstention. The frozen unchanged teacher produced all 192 rows, but reached only 10/72 registered-exact extraction outputs and 19/120 evaluation outputs. Output program support was insufficient for six classes, so compilation, packaging, and LayerCake were not invoked. The evidence shows that exact delexicalized templates are not a viable bridge to broader free-form English; that branch is closed rather than prompt-tuned. See docs/R20_INSTRUCTIONAL_REALIZATION_RESULT.md.

R21 then replaced the failed template compiler with signed generative LayerCake packages and matched sequence-distillation controls. Its factorized system passed the public suite at 116/120 on every seed, versus 96/120 for the teacher, and passed a fresh 1,080-row GPU replay, 54-row CPU replay, and all 24 package lifecycle/corruption controls. Its separately committed hidden split failed: every factor seed reached 113/120 and exceeded the 107/120 teacher, but summary was 13/20 against the frozen 15/20 minimum. Teacher-side semantic labeling also failed; the successful registered-ontology classifier is not autonomous discovery. See docs/R21_LABEL_SEPARATED_GENERATIVE_TRANSFER_RESULT.md.

R22 then tested one explicit semantic-plan normalization call for each of the 600 preserved R21 training responses. It failed before student training: 520/600 targets passed functional scoring and 255/600 preserved every field verbatim once, although all 600 remained grounded and non-collapsed. LayerCake was not invoked. This closes unconstrained teacher rewriting as the immediate repair path and preserves the ABI-versus-LayerCake failure boundary. See docs/R22_SEMANTIC_PLAN_NORMALIZATION_RESULT.md.

R23 then evaluated the unchanged R21 package bytes on a new committed hidden split with a public-corpus-supported semantic contract. All three factorized seeds scored 120/120 versus 96/120 teacher, 99-110 labeled monolith, and 27-32 raw sequence. A fresh replay reproduced 1,080 GPU and 54 CPU outputs, passed all 24 package removal/restoration controls, and rejected 24/24 targeted tensor corruptions. This certifies a bounded six-task supplied-content mechanism, not general English or the full ABI moonshot. See docs/R23_SEMANTIC_REPLICATION_RESULT.md.

R7 at a glance

Evidence	Result
Public archive	844,018,841 bytes
Archive SHA-256	fc50f423986149b5d4670ec9e28698540f64be96034efa26e5704c4469921e88
Archive members	1,193, all exact
Physical certification environments	3/3
Reachable inventory rows	301,543
Locked matrix	5,043/5,043 rows
Live causality	3,072 rows, 24 distinct processes
Live isolation	2,100 rows, zero target successes
Fail-closed controls	19/19 pre-public and 19/19 blind replay
Public reconstruction tests	17/17
Blind tar-prefix controls	12/12
Blind verdict	PASS, bounded scope

The blind reconstruction succeeded with a 368-character extracted path while Windows long paths were disabled.

Install

ABI targets Python 3.10. For the lightweight package and CLI:

python -m pip install -e .
abi status
abi self-check

Teacher extraction and research workflows require the dependencies in requirements.txt. Human-rating helpers use the optional human extra:

python -m pip install -e ".[human]"

Public R7 verification

Download public_release_assets_r7.json from the R7 GitHub Release. It binds the definitive archive and four capability packages by byte length, SHA-256, content address, release tag, and commit.

From a clean checkout of the public tag:

python -m abi_v2.public_reconstruction \
  --manifest /path/to/public_release_assets_r7.json \
  --tag-clone /path/to/clean/tag-clone \
  --workspace /path/to/new/workspace \
  --output /path/to/public-reconstruction-receipt.json

The external different-hardware workflow is under external_reproduction/. Its command sequence is:

abi-reproduce verify
abi-reproduce certify-hosts
abi-reproduce capability-matrix
abi-reproduce causality
abi-reproduce isolation
abi-reproduce performance
abi-reproduce hostile-audit
abi-reproduce report

Executing this sequence on the development laptop is a rehearsal, not an independent reproduction.

Human review

The frozen packet contains 7,000 judgments for each of three independent raters. The gate remains 0/21,000 until real raters complete and attest their forms:

abi human-rate --rater R1
abi human-rate --rater R2
abi human-rate --rater R3

See docs/PHASE2_HUMAN_RATING_HANDOFF_V1.md for the exact handoff.

Published artifacts

Capability	Bytes	SHA-256
English	253,216,208	acb787b3ffa0153c57d88cd37ba81c3f00b370d4ca4937e659cd4c775851f25d
Python	448,404	f1defaef2771ced336a332572a2d2f0e1e542399c877d182c48a6cd2e199231d
Civics	495,919	634ce66958859ec36dc1fbdf5ef34d6d2a9949d10cf2348a68c245d8c325d604
Chemistry	510,981	f9c9b2668fda5ef6b92844c1b7097fbdf8ff0daaae51f5b86f72d4a49000abeb

Claim boundary

R7 supports only the bounded capability-runtime/conformance result described above. It does not establish:

• arbitrary or universal model compatibility;
• tensor transplantation between unrelated architectures;
• complete diagnosis or extraction of teacher knowledge;
• fluent teacher-quality English generation;
• superiority over LoRA, distillation, or fine-tuning;
• human-rated quality;
• independent hardware reproducibility;
• global information minimality; or
• completion of the full ABI moonshot.

See docs/ABI_TECHNICAL_CLAIMS.md, docs/ABI_FINAL_RESULTS.md, and the review_packet/ directory before citing results.

Repository map

• abi/ — acquisition, labeling, artifact, and public CLI code.
• abi_v2/ — canonical runtime/conformance, certification, strict verification, public reconstruction, and external reproduction tooling.
• tests/ — supported automated checks.
• results/abi_final_validation_v2/ — immutable R3-R7 validation lineage.
• results/abi_moonshot/packages/ — published specialist packages.
• experiments/native_isa_r11/ — bounded native-neural ABI construction code.
• results/native_isa_r11/ — R11 preregistered raw evidence and live replay.
• experiments/foreign_teacher_r12/ — conventional-teacher frontend protocol and fail-closed public verifier.
• results/foreign_teacher_r12/ — additive R12-A receipts and exact packages; large source adapters remain excluded.
• experiments/foreign_capability_r13/ — separated Track B held-out protocol, replay, sealing, and hostile-verification code.
• results/foreign_capability_r13/ — local R13-B evidence and content-addressed artifact manifest; bulk replay assets are not yet public.
• experiments/foreign_capability_r14/ — non-exhaustive capability protocol, source replay, diagnosis, and hostile-verification code.
• results/foreign_capability_r14/ — compact R14 negative certificate and local artifact manifest; bulk adapters and raw replay rows are not public.
• experiments/foreign_neural_state_r15/ — R15A preregistration, generic neural-state frontend, physical extraction, strict/live verification, and hostile audit.
• results/foreign_neural_state_r15/ — compact R15A certificate and local immutable run lineage; bulk raw observations are not yet public.
• experiments/preexisting_representation_r15b/ — R15B public qualification, preregistered held-out protocol, isolated representation decoder, and fail-closed verification.
• results/preexisting_representation_r15b/ — compact R15B certificate and local evidence lineage; bulk recipient replay rows are intentionally excluded from Git.
• experiments/factual_semantic_r16/ — R16 public qualification, preregistered held-out protocol, isolated structured-fact compiler, and fail-closed verifiers.
• results/factual_semantic_r16/ — R16 raw held-out/live evidence, immutable chemistry/geography packages, accounting, and local certificates.
• experiments/prospective_length_span_r97/ — frozen prospective reasoning protocol, strict verifier, hostile audit, and bounded certificate.
• results/prospective_length_span_r97/ — source capture, bindings, raw rows, verification receipts, and byte-identical live replay.
• external_reproduction/ — independent-operator workflow and environment lock.
• review_packet/ — ordered technical and external-review handoff.
• docs/ — architecture, claims, results, and review instructions.

Historical experimental evidence remains in Git for auditability. Generated caches, model weights, temporary public reconstructions, and bulk reproducible intermediates are intentionally excluded from the production tree.

License and research status

This repository is a research release. Consult LICENSE and NOTICE before redistribution. Scientific claims are governed by immutable evidence and the claim ceiling above, not by roadmap language.