ABI is an alpha research compiler for acquiring, labeling, segregating, packaging, and verifying bounded capabilities from open-weight teacher models. It produces immutable, provenance-bound artifacts for a separate host such as LayerCake.
Get started | Documentation | Current status | Versions | Changelog
ABI and LayerCake have deliberately separate jobs:
- ABI qualifies teachers, probes bounded capabilities, labels and segregates records, accounts for imported information, constructs artifacts, and verifies the evidence.
- LayerCake installs, composes, routes, and executes compatible capability artifacts. Its runtime and performance claims belong to that repository.
Whole-moonshot verdict: FULL_MOONSHOT_NOT_PROVEN.
ABI has strong bounded results, but it has not yet proved automatic extraction of a minimal fluent-English substrate from an arbitrary pretrained model. It also has not completed human evaluation, independent different-hardware reproduction, global minimality, or a universal superiority comparison against LoRA and distillation.
| Evidence line | What is established | What remains open |
|---|---|---|
| R108 | Pinned inputs and preserved resource-budget stop: 13/96 initial responses, no replay or recipients | Control qualification and trained-task comparison remain incomplete |
| R107 | No-training residual intervention and exact live replay; modest natural-text loss interference, with a repetition trade-off | No fluent core, trained-route retention or native-host promotion |
| R7 public release | Bounded capability runtime/conformance across three named environments; public reconstruction passed | Human ratings and independent hardware |
| V1089 local campaign | Bounded machine evidence through Phase 7 and a same-machine Phase 8 rehearsal | One coherent public final artifact, Phase 5 clean replay, external review |
| R97 | Prospective two-hop reasoning transfer: 1,399/1,400 with causal controls and byte-identical replay | Broad English, arbitrary capabilities, independent reproduction |
| R21-R23 | Bounded label-separated supplied-content generation and prospective semantic replication | General English and autonomous discovery |
| R98 / C1 / R99 | Two source failures preserved; R99 data qualified, but its ABI-plan learner scored 10/48 vs LoRA 27/48 and distillation 20/48 | Failed bounded learning pilot; no fluent-English core promoted |
| R99-D1 / R100 | Fit audit completed; one full-model control scored 20/48 on fresh prompts, equal to retained LoRA | Quality and fit gates failed; acquisition bottleneck not yet isolated |
| R101 | Sampled execution paths agree; 288 exact unassisted replays; primary oracle-prefix continuation contrast is null | Data/generalization, optimizer fit and parent suitability remain unresolved |
| R102 | Matched renaming/exposure factorial completed; neither primary factor supported and no quality gate passed | Recipient/interface suitability remains unresolved |
| R103 | Raw control qualification failed at 41/48; all evidence preserved | Recipient/interface comparison not executed; control/evaluator validity remains unresolved |
| R104 | 155 finite scorer cases passed; raw control failed 29/48, with additional scorer-scope limits disclosed | No recipient comparison; general evaluator validity and control qualification remain open |
| R105 | A frozen NLI evaluator failed external scope labels: 779/1,200; exact live replay | Rejected as a judge; no English-control or acquisition promotion |
| R106 | Separate grammatical-discrimination control failed 178/200; exact live replay | No recipient comparison or training; supplied-token ranking is not fluent generation |
These lineages must not be combined into one certification. The controlling
public technical release remains
abi-final-validation-v2-repaired-r7-2026-08-30.
The current default branch is a later peer-review candidate.
The latest R108 attempt hit its fixed 25-minute control-process budget under reported GPU power/thermal throttling. Only 13/96 initial responses completed; no replay, recipients or training ran. The complete verifier rejects it. This is a resource-limited prerequisite, not an English quality failure, acquisition result or native LayerCake defect. The public reconstruction receipt verifies the preserved interruption and confirms that the complete verifier still rejects it.
The earlier R106 diagnostic failed its narrow grammar control at 178/200 against 190/200. It uses external sentence-pair labels, not the rejected NLI judge; all target tokens are supplied. No recipient or training run followed. Source-label limitations and exact replay are preserved; this does not qualify English generation or promote a product claim.
The R106 research release and additive receipt are public. Same-machine reconstruction matched all 581 evidence assets and passed 52 focused checks; tagged CI is green. Publication does not change the negative scientific decision.
The preceding R105 evaluator falsification rejected an off-the-shelf semantic judge on external gold labels, with exact live replay. This is a measured evaluation limitation, not an ABI or LayerCake breakthrough. No recipient ran, and earlier English-control gates remain closed.
Its research prerelease and reconstruction receipt are public: 40 evidence assets, eight evaluator files and 41 focused checks verified on the same machine. This publishes the rejection, not an English-core breakthrough.
The preceding R104 diagnostic stopped at its control gate: 29/48 under strengthened instruction checks. The legacy screen gives 44/48 but still fails the clarification task floor. Two scorer-scope limitations are preserved, not patched after observing outputs. No recipient ran and no native LayerCake defect follows. See also the weight-space research assessment: direct editing is a future competitor, not an ABI superiority result. The R104 research release is public; same-machine public recomputation matched 566 assets and 729 bindings and passed 43 focused checks. No scientific claim is promoted by publication.
The preceding R103 diagnostic stopped at control qualification: 41/48 raw passes versus the required 44/48. Recipient comparisons and training did not run. A temporal false-positive in the frozen contradiction proxy is documented without changing the result. No native LayerCake defect follows. The sealed research release includes the complete manifest and public-review helper.
The published R102 study completed its matched entity-renaming/exposure controls. Neither primary factor was supported and no checkpoint passed quality. The evidence is preserved; no new fluent core or ABI superiority is claimed. Its immutable research tag, checkpoints and review manifest are public.
The R100 acquisition-capacity control also failed: full-model fine-tuning scored 20/48, equal to retained LoRA and below the unchanged 44/48 gate. More trainable weights alone did not solve this bounded acquisition problem. No new core or ABI superiority is promoted.
The R101 diagnostic found no measured cached/padded execution disagreement. Correct answer prefixes did not improve the primary remaining-fact coverage contrast. Assisted scores are not an English-core breakthrough; the next acquisition question is still open.
Read the project status and claim ledger before citing a result.
ABI requires Python 3.10 or newer. Git LFS is required for a complete research checkout.
git clone https://github.com/Yoder23/abi.git
cd abi
git lfs install
git lfs pull
python -m venv .venv
python -m pip install -e ".[dev]"
python -m abi status --json
python -m abi self-check
python -m examples.segregate_capabilitiesThe example builds and validates a small in-memory English/domain segregation manifest. It does not download a teacher, train a model, or claim scientific quality.
Install optional dependencies only for the workflow you need:
python -m pip install -e ".[host]" # Torch host/runtime helpers
python -m pip install -e ".[extraction]" # Teacher extraction research
python -m pip install -e ".[human]" # Frozen human-rating workflowSee the getting-started guide for platform notes, the supported Python API, and verification commands.
The public alpha API in abi/__init__.py supports:
- immutable source-model manifests;
- bounded capability inventories and user selection plans;
- explicit English-core, specialist-domain, and quarantine records;
- nested teacher-information budgets and cost accounting;
- content-addressed acquisition bundles and verification; and
- deterministic release status and self-check commands.
The canonical flow is:
pinned teacher -> bounded probes -> labeled records -> selection/budget
-> immutable acquisition artifact -> external host conformance
An acquisition artifact is not itself a deployable model and does not silently include a teacher at inference. Host-side installation and execution are separate, explicitly measured steps.
| You want to... | Start here |
|---|---|
| Understand the project | Documentation index |
| Run the supported API | Getting started |
| Understand the code and evidence tree | Repository map |
| Review current claims | Project status -> claims |
| Reproduce R7 | R7 public validation |
| Conduct independent review | Independent review handoff |
| Run human scoring | Human-rating handoff |
| Run external hardware review | Phase 8 reproduction |
| Contribute | Contributing guide |
ABI does not currently establish:
- complete diagnosis of everything a teacher knows;
- fluent, minimal English extraction from arbitrary open-weight models;
- exhaustive or ontology-free domain discovery;
- native tensor transplantation between unrelated architectures;
- zero-loss foreign-teacher transfer;
- human-rated teacher parity;
- independent-hardware reproducibility; or
- general superiority over LoRA, distillation, or fine-tuning.
Within this repository, "lossless" is reserved for exact archive, manifest, tensor, package, or replay identity inside a declared compatibility boundary. It does not describe knowledge extraction from a foreign model.
Run the lightweight supported surface:
python -m pytest -q \
tests/test_public_release.py \
tests/test_capability_pipeline.py \
tests/test_capability_segregation.py
python -m ruff check abi/__init__.py abi/__main__.py \
tests/test_public_release.py examples/segregate_capabilities.py
python scripts/check_docs.pyThe complete default research suite has additional Git LFS and exact sibling LayerCake requirements. Follow the independent review handoff; do not interpret a missing historical payload as a pass.
When documents disagree, use this order:
- an immutable tag plus exact content hashes;
- frozen protocol and campaign contracts;
- raw observations and immutable packages;
- fail-closed verifier output;
- claim and status documents; and
- roadmap or explanatory prose.
Documentation cannot promote a scientific claim. Negative and superseded results remain in the repository for auditability.
ABI is Apache-2.0 licensed alpha research software. Historical experiment modules are not a stable production API and should not be exposed directly to untrusted network input. See SECURITY.md, LICENSE, and CONTRIBUTING.md.