← Complete research archive
Architecture researchAudit195 lines

FSTC1: Fixed-Slot Typed Compiler

Predecessor: SLC1 closed (1.2765% terminal, 0/613 depth-five)

docs/research/SHOHIN_FSTC1_FIXED_SLOT_TYPED_COMPILER.mdOpen original Markdown ↗

FSTC1: Fixed-Slot Typed Compiler

Status: closed development failure with strong causal signal; holdout sealed

Date: 2026-08-10

Predecessor: SLC1 closed (1.2765% terminal, 0/613 depth-five)

Holdout: sealed

Capability hypothesis

SLC1 failed because one autoregressive text stream had to learn formatting, program length, operation order, references, and exact arithmetic at once. Its dominant behavior was a one-record collapse. FSTC1 replaces free-form ledger generation with a bounded typed state transition:

[ z_{t+1}=R(z_t,\operatorname{Attend}(z_t,H_x),e(\hat s_t)),\quad t=0,\ldots,4 ]

where each hard slot prediction is

[ \hat s_t=(\text{ACTIVE/STOP},\text{op},r_L,p_L,r_R,p_R,v_t). ]

r is either one of at most seven source-number spans or a causally prior slot; p is identity or negation; and v_t is an exact signed-rational digit state. The same recurrent cell and typed heads are reused at every slot. A generic renderer may serialize predicted fields but cannot infer an operation, execute arithmetic, repair a value, or inspect an answer.

The hypothesis is narrow: explicit typed recurrent state will learn program unfolding and preserve causal state substantially better than SLC1's free-form text decoder. Passing FSTC1 is a compiler qualification, not a claim of broad reasoning.

Admitted geometry

The immutable RG-v4 ledger has 75,935 train and 3,917 source-disjoint development rows, depths one through five, and at most seven source numeric spans. Questions are at most 174 characters. Exact state requires at most 23 numerator digits and 11 denominator digits in training (16/10 on development).

The first pointer audit exposed 3,231 train and 159 development operands that the old ledger mislabeled as fresh literals. Every one is exactly the negation of a causally prior result. Adding a polarity bit to prior-slot references makes all 79,852/79,852 rows pointer-supervisable with zero unmatched operands. This result is hash-bound by 2fe9c48638f9ebd0ffbba41fa03f4a07817e2c64fe8ac2f1806f2213a7be18a9.

Generic numeric-span discovery is lexical only. It may expose byte offsets and decimal characters but may not parse expression structure, precedence, operations, dependencies, or answers. Equivalent repeated source occurrences are scored by value equivalence; exact occurrence identity is reported separately.

Architecture

The source stream is encoded once by a frozen language backbone plus a trainable bidirectional source memory. The state stream is a tied recurrent slot cell with cross-attention to that source memory. Source and state custody remain separate; only attention reads source state.

  • frozen host: pinned Qwen3.5-0.8B used by SLC1;
  • source memory: width 512, four bidirectional Transformer blocks, eight heads;
  • recurrent core: one tied GRU/cross-attention transition, width 512;
  • heads: ACTIVE/STOP, four operations, left/right reference, left/right polarity, and typed result state;
  • source references: seven lexical numeric spans plus masked padding;
  • state references: five slots with a strict < current slot mask;
  • result representation: sign, 23 numerator digits, 11 denominator digits, canonical length/EOS masks; zero and denominator constraints fail closed;
  • hard autonomous feedback: straight-through categorical states during joint training and ordinary argmax during evaluation;
  • expected trainable sidecar: below 30M parameters; exact receipt is required before a fit.

No host expression parser, calculator, symbolic executor, answer label, verifier, or semantic repair is available at inference.

Factorized gates

A. Skeleton compiler

Train source memory, recurrence, STOP, operation, reference, and polarity heads. Gold result-state embeddings may be supplied only as an explicitly labeled teacher-forced input arm. The autonomous arm feeds predicted prior references and predicted slot state. Result digits are not scored in this gate.

Development pass is conjunctive:

  • depth exact >=99%;
  • operation sequence exact >=97%;
  • operand kind/value/polarity exact >=97% each;
  • complete skeleton exact >=90%;
  • every-family complete skeleton >=85%;
  • depth-five complete skeleton >=80%;
  • aligned-minus-source-shuffled complete skeleton >=65 points;
  • source-shuffled complete skeleton <=25%;
  • recurrence-reset intervention loses at least 20 points at depth >=3;
  • zero invalid forward references and zero decode exhaustion.

B. Model-owned arithmetic transition

With oracle operation and exact operand digit states, a tied digit-circuit transition predicts one exact canonical rational result. Data are generated from the admitted value/operator distribution with a source-disjoint random seed and an unopened wider/deeper confirmation partition. No arithmetic routine runs inside inference.

Development pass is conjunctive:

  • exact result >=99.5% overall and >=99% for every operation;
  • sign, numerator length, denominator length each >=99.9%;
  • exact result >=98% on the longest admitted digit bucket;
  • shuffled operation or shuffled operand control <=25% exact;
  • carry/borrow-state reset loses at least 30 points on affected examples;
  • zero malformed rational states.

C. Joint autonomous compiler

This opens only if A and B pass unchanged. It composes their qualified owners, then permits one preregistered joint release. No gold slot or value enters the autonomous path.

Development pass is conjunctive:

  • valid typed program >=99%;
  • exact complete records >=90%;
  • terminal exact >=90% overall and every family;
  • depth-five terminal exact >=80%;
  • aligned-minus-source-shuffled terminal >=65 points;
  • source-shuffled terminal <=25%;
  • autonomous-minus-teacher-forced terminal gap <=10 points;
  • zero invalid references, malformed values, or exhaustion.

One development pass opens exactly one sealed holdout. A component failure closes that exact mechanism without rank, width, layer, duration, seed, threshold, renderer, or prompt variants.

Controls and accounting

The primary causal controls are source shuffle, recurrence reset, operation/ operand shuffle, and hard-state permutation. A parameter/FLOP-matched independent-slot decoder is required before a joint claim; it shares source memory but replaces tied recurrence with five independently queried slots. SLC1 remains the frozen autoregressive reference.

Every report records source/data/checkpoint/runtime hashes, trainable and total parameters, charged examples/tokens, forward/backward FLOPs, activation peak, wall time, generated slots/digits, per-depth/family outcomes, and all control interventions. Holdout remains unopened until the complete development gate passes.

Resource estimate

CPU tensorization and exact-reference audit: less than one CPU-hour. Mechanics: one H100 for at most 15 minutes. Skeleton development: one H100, estimated two to four hours. Arithmetic transition can run independently on a second H100, estimated two to four hours. Joint release is not allocated until both pass. Expected pre-joint charge is below eight H100-hours.

Frozen result

Mechanics job 749647 passed both gold and hard recurrent feedback with finite loss and gradients. Full fit 749648 completed 1,024 updates over 32,768 examples in 174.21 seconds (188.09 examples/s), using 2.38 GB peak H100 allocation. The sidecar has exactly 21,816,330 trainable parameters; checkpoint SHA-256 is 13777429b2c01047e7514f201d325566c44d2df113f68dd54286da6fac5f759a.

Parallel evaluations 749649--749651 closed the gate. Normal complete skeleton accuracy is 3348/3917 = 85.4736%; source-shuffled is 0/3917, and resetting recurrence removes 70.98 points on depth-three-or-greater rows. Thus source and recurrent state are both causal, but the absolute gate fails. Depth, operation sequence, reference kind, operand value, and polarity are 95.7365%, 90.6051%, 93.6176%, 89.4817%, and 92.8772%. Depth-five complete accuracy is 66.0685%; weakest-family complete accuracy is 56.4677%. No holdout opened and arithmetic transition/joint gates remain closed.

The read-only attribution localizes the miss to syntax requiring hierarchical scope. Products are 100%, chain sums 97.88%, and decimal chains 91.94%. Non-mixed expressions score 95.20%, but mixed precedence scores 46.96%. Zero-parenthesis examples score 92.52%, one-to-two parentheses 56.68%, and three-plus parentheses 14.14%; unary-group expressions score 25.59%. This closes fixed-slot recurrence without nearby variants. The successor must introduce an explicit model-owned parse stack/tree and scope transitions. Comparison, training, and attribution SHA-256 are 74035d8175145533b5a6d985c67e46564e73ced565fdca152cb7bcea17149cee, bd9751bb1b32fb2164178d706653152314431f9a4c85e20bd1e950ae164f8b24, and 5b7f8d2c162950d8ddfb405ed56733630b0781bbb626ae588eea51b87faa62ea.