Skip to content

Latest commit

 

History

History
168 lines (143 loc) · 13 KB

ABI_PROOF_LEDGER.md

File metadata and controls

168 lines (143 loc) · 13 KB

ABI proof ledger

Updated: 2026-11-14

This ledger separates software portability from neural transplantation. A result in one row must never be promoted into another row without its own registered evidence.

Latest attempt: R108 ended at its fixed control process budget, with 13/96 initial rows and no replay or recipients. Recorded GPU throttling does not establish any acquisition or host-quality result. Its complete verifier rejects qualification; no claim below is promoted.

Previous completed acquisition diagnostic: R107 supports bounded natural-text residual interference with exact live replay, but no fluent core is qualified. Its untrained-route retention and repetition trade-off do not promote ABI-C6, ABI-C7 or ABI-C8 below. Historical bounded passes remain separate.

ID Claim Current evidence State
ABI-C1 Immutable packages retain exact identity and canonical semantics R7 public release, reconstruction, isolation, and hostile verification Bounded pass
ABI-C2 Compatible hosts can execute a package through a shared capability runtime R7 across the three named conformance environments Bounded pass
ABI-C3 A source learning event can be extracted into a canonical package R13-B: four conventional Qwen sources, four held-out finite packages, 2,048/2,048 oracle rows Bounded enumerable local pass
ABI-C4a Identical source-extracted canonical-IR bytes can be pasted, removed, and restored through a shared runtime across frozen heterogeneous hosts R10: all 61,440 recipient rows passed; exact live replay Bounded synthetic component pass
ABI-C4b One learned ABI-native neural state can be packaged once and executed losslessly through frozen heterogeneous native output heads with zero recipient training R11: 90,112 rows passed; 7/7 hostile controls; exact-package live replay reproduced raw rows byte-for-byte Bounded synthetic construction pass
ABI-C4c A conventional source can expose a complete enumerable capability that ABI canonicalizes and executes exactly across frozen heterogeneous hosts R13-B: 96/96 atoms, 2,048/2,048 oracle, 12,288 AFTER/RESTORED recipient rows, 7/7 byte-exact replays Bounded enumerable local pass
ABI-C4d A conventional source can support reliable non-exhaustive capability recovery from mixed behavior R14: one of three fresh programs recovered; two latent selections failed Failed for tested output-probability frontend
ABI-C4e A generic frontend can recover a fresh deliberately learned synthetic capability from an anonymous foreign before/after weight delta and emit an exact portable package R15A: 8/8 held-out recoveries, 88,000 exact oracle rows, three recipients, strict/live/hostile PASS Bounded local pass
ABI-C4f A generic frontend can recover a bounded pre-existing capability from anonymous internal source representations without source training R15B: 4/4 secret mappings, 40,000 exact package rows, three recipients, strict/live/hostile PASS Bounded local pass
ABI-C4g An unchanged teacher can supply bounded pre-existing facts that ABI segregates into immutable domain packages and executes without the teacher R16: two 16/16 selections (27 distinct facts), 96/96 total package/source rows per run, exact namespace/removal controls, byte-exact live replays, 21/21 hostile cases, clean blind review Bounded replicated blind-reviewed structured-factual local pass
ABI-C4h ABI can extract teacher-derived compositional English realization into a teacher-absent package R19: 96/96 disclosed development plus 48/48 fresh hidden functional rows; zero regressions; 0/48 hidden control; 48/48 removal; 22/22 hostile; byte-exact live replay Bounded replicated local pass
ABI-C4i ABI can extend bounded realization to raw natural instructions and broader supplied-content behaviors with a structural program package R20 source: 10/72 extraction and 19/120 evaluation registered-exact; insufficient six-program support; no package or host run Failed for tested exact delexicalized-program representation
ABI-C4j ABI can transfer six semantic supplied-content behaviors into factorized signed LayerCake packages that generalize prospectively without teacher or receiver training at inference R23: 120/120 fresh hidden on all three seeds versus 96/120 teacher; 1,080 GPU, 54 CPU, and 24 lifecycle/corruption fresh replays Bounded strictly verified local pass
ABI-C4k A compact teacher-derived neural capability can causally create prospective reasoning behavior in a frozen LayerCake host and outperform its live teacher on the same unseen surface R97: 1,399/1,400 candidate versus 1,088 teacher, 60 parent, and 17 random; paired teacher advantage CI [0.20071, 0.24429]; 1,400/1,400 sparse; strict/hostile/live replay pass Bounded strictly verified local pass
ABI-C4 End-to-end foreign-teacher behavior copy/paste into frozen heterogeneous hosts with zero host training R12 strict gate failed; R13-B canonicalized the capability but matched only 372/2,048 source outputs Open
ABI-C5 The recipient's native neural computation internalizes the foreign capability R8 and R9 recipient realization failures Failed for tested mechanisms
ABI-C6 ABI extracts and segregates fluent English and arbitrary domains from an open-weight LLM No sufficient evidence Open
ABI-C7 ABI matches teacher generation quality and supersedes LoRA/distillation No sufficient evidence Open
ABI-C8 The full ABI moonshot is complete Dependencies C4-C7 are not all passed Not passed

Exact distinctions

Runtime-owned copy/paste

The package is immutable and recipient-independent. A fixed capability-blind interpreter executes its canonical semantics, while a host codec realizes the result through the host output interface. Removing the package removes the behavior; restoring identical bytes restores it. This is useful capability portability, analogous to portable code plus an instruction-set runtime.

Native neural transplantation

The same package changes a recipient's neural computation through a fixed capability-blind neural backend, with no capability-specific optimization, and the recipient produces teacher-equivalent behavior. The backend may not be an external answer solver. R8 and R9 did not establish this.

R11 establishes a narrower but positive construction boundary: its package is a learned neural transition state, a zero-parameter neural ISA executes it, and pre-capability frozen codecs inject the state into heterogeneous native output heads. Exact frozen-package replay passes. The learning substrate is ABI-native, not a conventional pretrained LLM, and the package is not internalized into recipient transformer blocks. R11 therefore promotes ABI-C4b but not ABI-C4, ABI-C5, or foreign-teacher extraction.

R13-B establishes ABI-C4c for an exhaustively enumerable finite family. A conventionally trained Qwen source exposed the complete 24-entry transition table, and the frozen R11 stack executed the canonical package exactly. Because native source/package agreement was only 372/2,048, this is not ABI-C4 teacher-function copying. Because the probe set is the entire capability specification, it is not evidence of non-enumerable or general knowledge extraction.

R14 tested ABI-C4d without atomic probes. The query/evaluation ratio was 256 / 4,642,668,576 per capability, and evaluation used unseen depths and order counterfactuals. One capability was recovered exactly on all 11,000 registered oracle rows, but two were not. This is positive mechanism evidence inside a failed reliability certification; it does not promote ABI-C4d.

R15A establishes ABI-C4e for a learned weight-delta interface. The extractor received no behavioral queries or answers and operated in a pivot-root capsule; all eight fresh packages were exact and all recipient rows replayed byte-for- byte. The source capabilities were deliberately installed through complete atomic supervision, so R15A does not establish pre-existing knowledge extraction, semantic labeling, English/domain transfer, or Track A equality.

R15B establishes ABI-C4f for a narrow modular-arithmetic representation interface. It uses an unchanged pretrained source, but the prompts specify the operations, source reasoning states the answer before the captured position, and labels are externally registered. It does not establish autonomous discovery, English/domain transfer, or teacher-behavior equality.

R16 establishes ABI-C4g for a registered two-domain factual workload. The source received zero training; answer-free generations and candidate-sequence scores were compiled in a physical capsule into two structured packages; and package execution matched 48/48 disjoint teacher questions with causal namespace and removal controls. Candidate vocabularies and semantic namespaces were registered in advance. This does not establish autonomous open-world discovery, fluent English transfer, arbitrary-domain completeness, native neural transplantation, or teacher-quality free generation.

R18 establishes only the public prerequisite for ABI-C4h. It compiled free teacher realizations into a feature-factorized 24-template package and generalized to supplied unseen lexical slots without the teacher. This is a registered surface-realization family, not autonomous prompt understanding, conversation, broad English generation, or LayerCake hosting. Promotion required a preregistered hidden lexical replication. That replication failed 46/48 because number pooling copied singular Does into a plural signature; R18 therefore does not promote ABI-C4h.

R19 corrects the measured R18 number-agreement defect without hard-coded surface tokens. It passes 96/96 disclosed development rows and a separately committed fresh lexical replication at 48/48, with zero teacher-correct regressions, a rejected control, exact removal, 22/22 hostile rejection, and byte-exact live replay. This promotes ABI-C4h only for the registered supplied- slot, 24-signature realization family. It does not promote general English or teacher-behavior copying.

R20 tests the next broader structural boundary and fails before compilation. The teacher's summaries, emails, clarification questions, and abstentions were not reducible to six repeated verbatim-preserving output programs. This closes the exact delexicalized-template route for broader English. It does not invalidate R19's narrower pass, and it supplies no evidence about LayerCake because no R20 package or host execution occurred.

R21 replaced the structural program with generative LayerCake packages and passed its public prerequisite, but its frozen lexical hidden gate failed at 13/20 summary rows. R22 then rejected unconstrained teacher-side normalization as a repair. R23 prospectively froze a semantic contract and a new hidden split without changing the R21 package bytes. All factorized seeds scored 120/120, and a separate live process reproduced the complete GPU/CPU and package lifecycle matrix. This promotes ABI-C4j only. The prompts still supply three semantic fields and use a registered six-class ontology, so it is not broad English extraction, autonomous labeling, arbitrary-domain segregation, or teacher-model transplantation.

R97 promotes ABI-C4k after preserved R90/R92/R93/R94/R95 negative evidence and one bounded repair. Its candidate was frozen before R97 existed, source rows were captured and committed before candidate binding, and the unchanged package passed the absolute, per-family, teacher-retention, paired-comparison, causality, sparse, and collapse gates. Strict recomputation trusts no stored scientific boolean; 12 hostile shadow-tree mutations failed closed; live replay was byte-identical at the raw-row level. This is a neural teacher-to-ABI-to-LayerCake transfer result for a bounded nonce two-hop reasoning family. It does not imply general English, open-world domains, heterogeneous-host execution of this neural package, global minimality, or universal method superiority.

Teacher-quality English/domain extraction

An open-weight source is diagnosed, English and domain knowledge are separated, the selected information is packaged, the source is absent at inference, and the resulting LayerCake passes teacher-relative generation-quality gates. No existing ABI release establishes this.

Promotion rule

R10 promotes only ABI-C4a, the bounded runtime component. R11 promotes ABI-C4b, the bounded synthetic neural-ABI construction. R13-B promotes ABI-C3 and ABI-C4c only for local, exhaustively enumerable capability canonicalization. R14 promotes no claim because its registered three-capability gate failed. R15A promotes ABI-C4e only, R15B promotes ABI-C4f only, R16 promotes ABI-C4g only, R19 promotes ABI-C4h only for its bounded replicated realization family, R23 promotes ABI-C4j only for its six supplied-content behaviors, and R97 promotes ABI-C4k only for bounded prospective two-hop reasoning. None promotes ABI-C4 teacher-behavior copying, ABI-C5, ABI-C6, ABI-C7, or ABI-C8. Foreign-teacher behavioral transplantation and English/domain quality remain separate registered lines.