SLC1: Structured Ledger Compiler

Status: prospectively frozen before any GPU capability output on 2026-08-10.

Purpose

The closed DSET-to-BSOT cascade showed that a language model can often repair a known local error but that free-form natural trajectories do not provide stable edit addresses. A holdout-blind CVG1 audit then rejected raw text pointer editing: an unrestricted ordered diff copied only 44.9179% of development target characters and required 56.97 copy runs on average.

SLC1 tests the necessary replacement interface: can one model compile a source problem into a compact, model-owned, addressable operation ledger without an external solver at inference? SLC1 is an interface qualification, not a claim of general reasoning and not the final edit architecture.

Data

The immutable RG-v4 source has SHA-256 a87e7c8279048d9fbdf19b245aeff194305415ea06d28258902b397bbe96875f. An independent exact-arithmetic compiler admitted only five executable families and rejected every false intermediate step or terminal mismatch.

• train: 75,935 source-disjoint rows, 198,335 exact operations;
• development: 3,917 rows, 10,120 exact operations;
• operations: ADD, SUB, MUL, DIV;
• depths: one through five records;
• holdout: unused and unavailable to SLC1;
• train JSONL SHA-256: 6a5876f2b8eed1387c31459062102b9bd007bff99d556ee6f63c75613310f671;
• development JSONL SHA-256: 760044d9b3851197988b361eba021ffbfe013600fab464731ab72de5e617ec87.

Every target uses one canonical grammar:

<LEDGER_V1>
R0|SUB|8|3|5
R1|ADD|@R0|5|10
COMMIT|@R1|10
</LEDGER_V1>

Record references create stable semantic addresses. Canonical materialization round-trips every record exactly. With the pinned tokenizer, complete source-plus-target sequences have train/development maxima of 380/355 tokens; a 384-token context therefore has zero truncation.

Frozen Fit

• host: pinned Qwen3.5-0.8B revision 2fc06364715b967f1860aea9cf38778875588b17;
• base weights frozen;
• final four layers, rank-8 LoRA over the established full projection scope;
• alpha 16, learning rate 2e-5, cosine decay;
• 1,024 updates, microbatch 4, accumulation 2;
• maximum 75,935 rows, deterministic data seed 2026081022;
• model seed 2026081021;
• 384-token context;
• greedy generation, maximum 320 new tokens, seed 2026081023.

One two-update mechanics run may verify finite loss/gradient, memory, and generation plumbing. It cannot change any scientific setting or count as a capability result.

Development Evaluation

All 3,917 source-disjoint development rows are evaluated exactly once in the aligned arm. A matched causal control replaces each source with another source from the same family and ledger depth whose gold ledger differs, while keeping the recipient gold fixed. The control changes no model, decode, prompt shape, or target distribution.

SLC1 passes only if all conditions hold:

1. syntax-valid ledgers >=99%;
2. exact record count >=99%;
3. exact operation sequence >=95%;
4. exact complete record sequence >=90%;
5. exact terminal value >=95%;
6. exact terminal value >=90% in every family;
7. exact terminal value >=85% at depth five;
8. aligned terminal accuracy exceeds source-shuffled accuracy by at least 65 percentage points and source-shuffled recipient accuracy is <=25%;
9. zero generation exhaustion and complete hash/identity coverage.

There is no prompt, rank, layer, learning-rate, update-count, seed, or threshold rescue. A pass qualifies the ledger compiler and opens one separately frozen record-edit/replay transducer with clean/fault/natural pairs, deterministic generic edit execution, forced-script intervention, and matched hidden, shuffled, label-permuted, and full-regeneration controls. A failure closes this standard-LoRA compiler interface; any successor must change the compiler architecture rather than repeat this fit.

Claim Boundary

SLC1 may establish only reliable source-to-ledger compilation and causal source use. It does not establish autonomous general reasoning, natural-language coverage, or a benefit from editing. Those claims require the downstream model-owned record editor, tied recurrent replay, final renderer, and matched answer-level controls.