DIVERGE-CGL1: Causal Grounding Lattice

Status: data mechanics frozen after the QTE1 development ceiling and before any CGL1 neural training. QTE1 confirmation and composition remain independent.

Hypothesis

GTI1 demonstrates that direct role labels are cheaply fit through renderer lookup. CGL1 removes those labels from the candidate-visible corpus. It represents both complete READ transactions, executes each against a sealed two-candidate state, and supervises only the observed terminal answer:

[ L_{outcome} = -\log \sum_{j: execute(T_j)=y^*} p(T_j\mid source,state). ]

Each semantic source appears under three state interventions: two distinct value assignments that swap the target outcome, and one equal-outcome state where both transactions produce the same answer. The two clause orders for each meaning are also present. A semantic transaction must therefore stay coherent across six records even when the outcome alone is underdetermined.

This is not claimed as novel latent-variable learning. Its purpose is to test whether downstream consequences plus intervention consistency can train a small model-owned semantic owner after direct role fitting failed. QTE1 is the fixed 0.8B capability ceiling, not a runtime teacher in this gate.

Frozen data contract

The source is the immutable 100,000-row RRG1 QUERY corpus at SHA-256 2d325c860e707307886f782350e7ec35ae8c23ae275260b0a937bbb738078c1c. The deterministic builder emits 300,000 public records and 300,000 separate outcome-supervisor records: 200,000 distinct-outcome and 100,000 equal-outcome cases across 50,000 complete semantic pairs.

Public records contain source text, source-owned symbols, anonymous candidate values, state-orbit identity, and commitments. They contain no target, distractor, symbol-role, role-order, or gold-transaction field. Supervisors contain only the committed public identity and terminal answer. Exhaustive generation verifies every six-record orbit, value swap, equal-outcome case, clause-order answer invariant, identity, and forbidden-field audit.

The first local/Stokes cross-host build reproduced the public and supervisor files exactly, but exposed that report v1 serialized absolute host paths. That receipt format is rejected before neural use. Report v2 stores only canonical roles and relative artifact names; public and supervisor bytes are unchanged. The canonical hashes are:

• public: bc438f793a3ced67a3b5493d70c14cbc39db4c20f3fe0fb50579af6b5f1daea9;
• supervisor: affa2cc36412f07f2816a00bbe2abfb06ee93be3b602c79b79b248f4ccf2552d;
• report: e9267aadaa1413778d7ac54db6a72a95500da92acee733fccaa655830b9cb1a6.

Stokes must reproduce all three canonical hashes before neural admission.

Neural admission boundary

QTE1 is closed and independent Stokes job 767029 reproduces all three data hashes exactly. The frozen neural interpreter uses each parent model's final eight blocks with LoRA rank 16 and alpha 32. For each complete candidate it scores YES versus NO likelihood for the fixed claim that the candidate is the requested source and the other is the distractor. The transaction distribution is trained only through the terminal-outcome marginal. No direct role or transaction label enters the candidate-visible or supervisor files.

The complete 300,000-row objective is evaluated through exact sufficient statistics: each of 50,000 semantic pairs has two clause orders, two informative distinct-outcome copies, and one zero-gradient equal-outcome copy. The compressed mean multiplies source cross-entropy by 2/3, exactly matching the six-row mean. A 0.25 symmetric clause-order consistency penalty aligns the two distributions by physical mention identity without revealing which identity is TARGET. Training is one deterministic epoch over 50,000 pairs, pair batch 32, AdamW 1e-4 cosine decay, seed 2026080702.

Three independent single-H100 arms run concurrently: protected Shohin, SmolLM2-135M, and a matched SmolLM2 control whose distinct terminal outcomes are flipped. All receive identical data, order, updates, objective geometry, and evaluation. Development uses the source-disjoint CCR1 board at SHA-256 299237068f436ba33a68487b5300fcd724f8c98bd8bfe6b1916a4ebc7541ebf7.

A treatment pass requires at least 765/768, every mode 254/256, every renderer 127/128, mapped mention-swap equivariance 765/768, a context- scrub drop of at least 250, bit-exact entity-renaming behavior, and an unchanged frozen parent. The flipped control must remain at most 430/768. If both treatments pass, selection is deterministic by exact count, signed margin, then SmolLM2. A development pass admits exactly one fresh balanced confirmation board. A miss closes the exact outcome mechanism without seed, width, duration, renderer, prompt, or threshold variants.

The confirmation generator is frozen before development results. It uses seed 2026080703, 256 new exact TFS1 programs, a disjoint 32-entity bank, six new query families, and two clause orders per family. Every renderer contains exactly 64 transaction-0 and 64 transaction-1 queries. Generation runs on Stokes with exhaustive source/query/entity overlap audits against CGL1 training, the development board, and the already-open PQI/QTE board. The board remains unopened unless assessor 744568 passes.

The first independent build job 767032 rejected the proposed entity bank before writing a board: 15 of its natural-word symbols already occurred in a prior board. Timeout-only recovery 767035 was canceled as soon as that deterministic data-contract failure was known. Before any CGL1 neural result, the bank was replaced with 32 fixed opaque symbols independently checked to have zero overlap with the 300,000 CGL1 public rows and both prior boards. Seed, typed programs, query templates, renderer/order balance, overlap rules, and evaluator are unchanged. The corrected build must still pass the complete Stokes audit before a confirmation job can be dependency-released.

The confirmation evaluator is frozen at commit c08be09. Its minimal read-only Newton overlay is runtime_overlays/diverge_cgl1_confirm_runtime_c08be09_r3, with SHA256SUMS SHA-256 367a85bf08767842983ebf636f38f5bda580c54201f9093ef2dd361a11cf50d7. The overlay contains only the confirmation data contract, evaluator, job wrapper, source receipt, and checksum manifest; all other imports resolve from the already-qualified CGL1 base runtime. A fail-closed CPU dispatcher may run only after 744568 exits successfully. It verifies the fixed board, verifies that the assessment selected exactly one admitted treatment, submits exactly one H100 confirmation, and records the child job and artifact hashes. It does not alter an arm, threshold, board, or score.

Zero-training backbone ceilings

While the frozen CGL1 arms ran, six independent single-H100 jobs measured the already-open development board without training or changing the CGL1 decision. Each arm scored one fixed intervention with candidate-wise YES versus NO likelihood under an exact pinned parent:

Qwen normal is exact on renderers 0/2/4 and zero on 1/3/5. SmolLM3 normal is exact on renderers 0--4 and zero on renderer 5. Both scrubbed arms fall to the same 384/768 balanced-order baseline with zero mean signed margin. SmolLM3's perfect mapped swap shows that a stronger pretrained parent can expose the relation equivariantly, but the normal renderer-5 inversion prevents treating it as a qualified semantic owner. These are capability ceilings only, not trained CGL1 results or promotion evidence.

Jobs 744606--744611 all completed normally in about 13 minutes. Report SHA-256 values in normal/scrub/swap order are Qwen eccff4c9...a12, e493244b...c75, 6a4e647e...025; and SmolLM3 796c53a2...bc7, e3d345fa...525, e8cb6b8c...a7f. Full reports are preserved read-only under artifacts/reasoning/diverge_cgl1/capacity_3bcbc81_r1.

Development result and attribution boundary

The frozen three-arm development gate is closed as a conjunctive failure. Shohin and SmolLM2 each fit all 100,000/100,000 true training assignments and score 768/768 normally. Their context-scrub scores are both 384/768 with zero mean margin, entity renaming is bit-exact, and frozen parents are unchanged. Shohin mapped mention swap is only 384/768; SmolLM2 is 512/768. Every swap error occurs as a complete renderer block. The flipped SmolLM2 control fits all flipped supervisors, fits 0/100,000 true labels, and scores 0/768 normally. Outcome supervision therefore causally controls the learned decision, but does not identify a permutation-equivariant referent binding.

Development report SHA-256 values are Shohin a0247957...a38c, SmolLM2 5b788fa7...eaf4, and flipped SmolLM2 2b7befd3...7f7d. Assessor 744568 writes fail receipt 8c2e2975...eb67 and exits nonzero. Fail-closed dispatcher 744617 is canceled without running, so confirmation remains unopened.

The corrected independent confirmation board nevertheless completed its pre-result Stokes construction and zero-overlap audit. It has 256 rows, 768 balanced queries, 1,048,576 represented worlds, zero source/query/identity/ entity overlap against all bound inputs, board SHA-256 ede48fc5...e7cea, and report SHA-256 b399b686...46758. Exact bytes are staged read-only on Newton and locally but are not model-scored.

Exactly one read-only attribution is frozen after closure. For each opened development query it computes normal candidate log evidence s(x), computes the alpha/beta-swapped counterfactual s(gx), maps the latter into the original physical-identity frame, and reports the fixed orbit product

[ s_{orbit}(x) = s(x) + g^{-1}s(gx). ]

This is a diagnostic ceiling, not a CGL rescue or promotion. It changes no weight, data, prompt, threshold, or confirmation state. If the product recovers the renderer-block failures, the next architecture may enforce an exact permutation-quotient commit boundary; otherwise that hypothesis is rejected. No additional CGL seed, width, duration, loss, or prompt variant is authorized.

Independent replay and orbit-attribution result

The corrected independent development replays closed without changing any scientific input. Shohin job 744620 reproduced the primary development report byte-for-byte at SHA-256 a02479576a5d404f3d5a7bc55d5218652877656db7c76dd8e130dd46d431a38c. After correcting only a stale immutable-tokenizer directory receipt, SmolLM2 job 744624 reproduced its primary development report byte-for-byte at SHA-256 5b788fa7fe8a79e59e80b814c09f2d4084868450b64b1f458c6b83f99058eaf4. Both jobs used separate writable output roots. The failed/canceled launch attempts remain preserved as infrastructure evidence and never produced a score.

The one frozen attribution completed on jobs 744622/744623:

SmolLM2's product is exact across all six renderers, but Shohin still inverts renderers 1 and 5 completely. Exact two-element log-evidence pooling is therefore a useful stronger-backbone capacity result, not a Shohin mechanism and not a CGL1 rescue. The CGL1 confirmation remains unopened. The successor must change the learned computation boundary so that candidate identity is a typed equivariant object throughout compilation and commitment, rather than trying another CGL1 fit or relying on inference-only score pooling.