Project ShohinOPEN BUILD
Menu
← Complete research archive
Architecture researchPreregistered156 lines

PSTC1: Pushdown-Stack Typed Compiler

Predecessor: FSTC1 closed at 85.4736% complete skeleton

docs/research/SHOHIN_PSTC1_PUSHDOWN_STACK_COMPILER.mdOpen original Markdown ↗

PSTC1: Pushdown-Stack Typed Compiler

Status: closed development failure; holdout sealed

Date: 2026-08-10

Predecessor: FSTC1 closed at 85.4736% complete skeleton

Holdout: sealed

Hypothesis

FSTC1 is nearly exact on flat programs but fails predictably under nested scope, unary groups, and mixed precedence. Increasing flat recurrence would not represent the missing invariant. PSTC1 replaces its five independent output slots with a model-owned pushdown transition over a small action alphabet:

  • PUSH(source_number);
  • NEGATE;
  • APPLY_ADD, APPLY_SUB, APPLY_MUL, APPLY_DIV;
  • STOP.

The recurrent controller attends a separately encoded raw source and the top two stack states. Hard actions update a bounded tensor stack. PUSH copies one model-selected source-number representation; NEGATE changes its typed polarity; APPLY pops two states, writes one ordered operation record, and pushes one result-owner state. A generic stack executor materializes only the predicted actions and pointers. It cannot parse the source, choose precedence, execute arithmetic, repair a program, inspect an answer, or invoke a verifier.

This is structurally different from FSTC1: hierarchy and scope are represented by stack topology rather than inferred inside one flat slot state.

Training-only supervision

An independent exact supervisor extracts the arithmetic expression from each admitted family, parses it only during corpus construction, emits postorder PUSH/NEGATE/APPLY actions, executes those actions with exact rational arithmetic, and requires operation-by-operation parity with the immutable gold ledger. The supervisor and exact values are absent at model inference.

Admission requires every train and development row to have:

  • one unique source expression span;
  • exact source ownership for every pushed number;
  • exact action execution and terminal parity;
  • exact ordered binary-operation/result parity with the existing ledger;
  • bounded action length and stack depth;
  • zero source overlap changes and zero holdout access.

Any row that cannot meet all conditions is excluded as a whole and reported. The gate may proceed only if at least 99% of every development family remains.

Frozen model and budget

  • pinned frozen Qwen3.5-0.8B source encoder;
  • width-512, four-block bidirectional source memory;
  • one tied width-512 controller with source cross-attention and top-two stack reads;
  • one shared stack-write cell and action/pointer heads;
  • at most 22 actions and stack depth 6, fixed by the CPU receipt;
  • 1,024 updates, batch 32, first 128 updates gold-action feedback and all remaining updates hard autonomous feedback;
  • AdamW, LR 2e-4, betas (0.9,0.95), weight decay 0.01, gradient clip 1;
  • exactly the FSTC1 charged example budget (32,768) and one seed;
  • trainable sidecar must remain below 30M parameters.

No FSTC1 weights are warm-started. This makes the changed state topology the principal intervention.

Controls

  1. same-family/binary-depth source shuffle, paired by nearest action length;
  2. stack reset before every action while retaining identical weights;
  3. stack-top permutation at evaluation;
  4. parameter/FLOP-matched recurrent decoder whose action state cannot read the stack;
  5. frozen FSTC1 and SLC1 references.

Development gate

The pass is conjunctive:

  • action length exact >=99%;
  • action sequence exact >=97%;
  • source pointer value exact >=97%;
  • generic-executor valid program >=99%;
  • exact materialized operation skeleton >=92%;
  • every-family exact skeleton >=88%;
  • mixed-precedence exact skeleton >=85%;
  • unary-group exact skeleton >=80%;
  • three-plus-parenthesis exact skeleton >=80%;
  • aligned-minus-source-shuffled exact skeleton >=65 points;
  • source-shuffled exact skeleton <=25%;
  • stack reset loses at least 30 points on mixed/unary/parenthesized rows;
  • stack-top permutation loses at least 20 points on depth-three-or-greater;
  • no invalid pop, overflow, forward reference, or decode exhaustion.

One development pass opens exactly one sealed holdout. Failure closes PSTC1 without action-vocabulary, width, depth, duration, seed, LR, prompt, tokenizer, or threshold variants. Arithmetic-state learning and record editing remain closed until the compiler passes.

Claim boundary

A PSTC1 pass would establish learned hierarchical source-to-program compilation with a causal model-owned stack. It would not by itself establish broad reasoning, arithmetic execution, or a novel universal architecture.

Data admission result

CPU job 749652 admitted all 75,935 train and all 3,917 development rows with zero exclusions. Every postorder action sequence reproduces the existing binary operation and result sequence exactly. Maximum action count is 22 and maximum stack depth is 6. Train/development SHA-256 are b522d58b3403064987a47d7b22657ef862945f18dfb352a78625fbcc27ebf008 and bf90044cd17e5e63cdeb7f91bffaae6e4bc59e992fb95cb8da9d2b29da00c914; report SHA-256 is 31913bcd4fb83a7481221b56dc675ead4179e0f1bb8bec5eaa3d494e45f94b37.

Development result

The mechanics replacement 749654, full fit 749655, and evaluations 749669--749672 completed cleanly. The compiler has 21,026,311 trainable parameters. Training charged 1,024 updates / 32,768 examples in 252.4566 seconds (129.7966 examples/s), peaked at 2,374,637,056 GPU bytes, and reduced loss from 3.4148 to 0.1873. Checkpoint SHA-256 is 412e948b3744a2a719cf646775103909db317fdbf8b65dc6aca1e04595afce06.

On all 3,917 source-disjoint development rows, normal exact skeleton is 3,597/3,917 = 91.8305%. Action-length exact is 95.3025%, action-sequence exact 91.8815%, pointer-value exact 94.6643%, and valid-program rate 100%. Same-family/depth source shuffle scores zero exact skeletons. Stack reset removes 68.74 points on hierarchical rows and stack-top permutation removes 75.04 points at binary depth three or greater. Source and the learned stack are therefore causal rather than decorative.

The conjunctive gate fails. Every-family minimum is 71.7662%, mixed precedence 64.5570%, unary groups 35.4331%, and three-plus-parenthesis rows 18.1818%. Exact skeleton by binary depth is 99.9294%, 98.6217%, 92.0530%, 84.6032%, and 73.0832% for depths one through five. Of the 320 exact failures, 318 have a wrong action sequence, 209 have a wrong source pointer value, 184 stop at the wrong length, and 280 are mixed-precedence rows. This localizes the residual defect to free-running source-position and scope tracking, not stack execution validity.

PSTC1 closes with no nearby variants and no holdout access. Its result and training-report SHA-256 are eeff01124c17fcc8e74c52abc57f80367c0f457002fdee6652df1f2f258cc744 and b6989eddbf55327f5eac66be000787f426baa2d909fbdabe467d7f26a2a49a02. The structurally different successor must make source traversal monotonic: predict one lexical role per source token and let a generic precedence stack materialize the operation program, instead of emitting an unconstrained global action sequence.