Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1,106 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

slm-training

Novel SLM experiments: harnesses for placeholder OpenUI layout generation (official @openuidev/lang-core), a TwoTower masked-diffusion model, plus a GPU multi-farm MCP.

This is not a natural-language LLM. It is a grammar-constrained symbolic diffusion model that emits templated grammars — scaffolded structure and structural reasoning — with templated content deferred to a real external LLM. Constrained decoding is the product, not a switch: deterministic completion paths bypass inference, a scope-proven singleton is committed with no forward pass, and no lever or experiment may make output less legal. The goal law is AGENTS.md § Non-negotiable architecture invariants and its canonical expansion decode-invariants.md.

What's included

  1. Training-data harness — build/validate versioned train corpora
  2. Testing-data harness — held-out / adversarial / OOD eval suites
  3. Model-building harness — lineage-first TwoTower and causal-LoRA tracks
  4. OpenUI Lang bridge — Node sidecar over official @openuidev/lang-core
  5. GPU multi-farm MCP — list / launch / cost-project across Vast.ai, RunPod, Lambda

Autonomous experiment campaigns use the fail-closed, evidence-grounded autoresearch harness, with isolated, pinned Open Deep Research and OpenResearcher implementations behind one memo/trajectory contract and trusted hypothesizer. Before execution the pipeline requires a persisted matrix of at least five distinct, grounded hypotheses, including categorical candidate-novelty audits adapted from Wang and Buehler (2026). Pre-run audits are not claims of proven discovery or SOTA. Each matrix names its recommended experiment; completed outcomes and diagnoses become typed feedback for the next matrix and for future campaign evidence. The loop improves by evidence, never by rewriting its own code, frozen cases, or gates. RL remains locked until a model passes the frozen production readiness contract. Architecture hypotheses can additionally use Lean formal preflights to reject structural contradictions before training; these prove explicit abstractions and assumptions, never empirical quality or ship readiness.

See docs/design/decode-invariants.md (goal law — constrained decoding, deterministic bypass, symbol-table speculation and scheduling, shared ops vocab, CRDT multi-turn), docs/design/model-lineage.md (canonical two-track cycle), docs/design/openui-twotower.md, docs/design/grammar-topology-diffusion.md (dynamic production-tree diffusion), docs/design/verified-scope-solver.md (VSS0 verified scope-solver contract — prefix legality vs verified support), docs/design/research-lineage.md (papers → code), docs/design/semantic-planning-valid-state-disposition.md (SPV4-02 final disposition), docs/design/recurrent-semantic-computation-looped-latent-disposition.md (RSC4 blocked; not ship), docs/design/research-correction-critics.md (V4 remask / trust-gate / honest inventory; V6 CoRe/T2M), docs/design/verifier-stack.md (G0–G12 corpus gates + confidence tiers), docs/design/abstraction-house-style.md (L0–L5 determinacy, grounding, and canonical defaults), docs/design/verifier-guided-repair.md (PDDL-Instruct / verifier-repair applicability map), docs/design/quality-experiment-matrix.md (E0–E75 + X0–X15 matrices; E34 deferred), docs/design/speculative-denoising.md (V7 stability / dependency-cluster / survival / successor-cache decode), docs/design/dsl-native-tokenizer.md (V5 lexer alphabet), docs/design/grammar-fastpath.md, docs/design/grammar-backends.md, docs/design/dsl-pack-contract.md (F1 DSL-pack contract; OpenUI first pack), docs/design/structure-only-eval.md, docs/design/binding-aware-meaningful-v2.md (versioned binding-aware metric and gaming audit), docs/design/judge-independence-audit.md (EFS0-04 cross-family/human audit contract), docs/design/adversarial-review.md, docs/design/runtime-performance.md, docs/design/hf-jobs-train.md (HF Jobs full train — not ZeroGPU), docs/design/gpu-multi-farm-mcp.md, and docs/MODEL_CARD.md.

Final spectral program policy: null-calibrated spectral learning disposition (diagnostics only; no spectral training/default/promotion adoption).

Calculated arity, task rate, neural precision, and physical cost are kept distinct by the CAP0 contract.

Model card (summary)

Full card: docs/MODEL_CARD.md. Agents update both this summary and the full card whenever a checkpoint is created or promoted.

Current compatibility: output contract v2 is symbol-only. All pre-E714 checkpoints are provenance-only; current code refuses to load, serve, resume, promote, or evaluate them. E714 is the first compatible scratch baseline, but it fails semantic gates and is not promoted. See the contract.

Spectral disposition: the final NCS audit adopts only fail-closed diagnostics. It rejects or blocks raw-alpha quality claims, spectral training controls, causal spectral retention, and WW-PGD/trace-log projection. No checkpoint, roster, training default, champion, or promotion status changes.

Role Checkpoint Where Claim
CAP1 cached zero-parameter typed solver 8 solver manifests; no neural checkpoint outputs/experiments/cap1-cached-zero-parameter-solver-successor-20260729/ (local, explicit no-sync) Score/prediction-identical to the uncached 0-param solver, but immutable frame/schema memoization cuts elapsed 27.802→15.134s (−45.6%; −59.0% vs trained deterministic). Schema arms remain identical and n=32 underpowered; retained runtime baseline, CERT_CAP1 rejected, not learned/ship (results)
CAP1 zero-parameter typed solver 8 solver manifests; no neural checkpoint outputs/experiments/cap1-zero-parameter-solver-successor-20260729/ (local, explicit no-sync) Corrected v2 suite with 0 params/steps/train rows/forwards. Schema scores exactly match the trained deterministic arm (relevant/invariant/CAP0 1.0, canonical .800) in 27.8s vs 36.9s, but all schema arms remain identical and n=32 underpowered. Smallest sufficient fixture baseline; CERT_CAP1 rejected, not learned/ship (results)
CAP1 deterministic typed candidate successor 8 matched scratch scorer checkpoints outputs/experiments/cap1-deterministic-typed-candidate-successor-20260729/ (local, explicit no-sync) Corrected v2 suite; schema arms reach 1.0 relevant/invariant/CAP0 and +.3125 vs NL (p=.00195) through 24 fact + 4 exact + 4 accepted-set bypasses and 8 guard abstentions, with zero model forwards. All schema arms are identical, so trained params are non-load-bearing; CERT_CAP1 rejected and checkpoints must never be reused/promoted/ship (results)
CAP1 typed-fact candidate successor 8 matched scratch scorer checkpoints outputs/experiments/cap1-typed-fact-candidate-successor-20260729/ (local, explicit no-sync) First valid cap1_two_pack_freeze/v2 run; same 34,913 params/arm. Schema arms reach 1.0 relevant sensitivity/invariance and +.3125 vs NL (p=.00195) with 27 exact fact bypasses and 5 forwards, but filtered-multi equals unfiltered, paired n=32 is underpowered, and filtered-single CAP0 is .75. CERT_CAP1 rejected; never reuse/promote/ship (results)
CAP1 counterfactual-supervision candidate successor 8 matched scratch scorer checkpoints outputs/experiments/cap1-counterfactual-candidate-successor-20260729/ (local, explicit no-sync) Invalidated v1-suite fixture: mini-flow prompts asserted the forbidden mode. Historical relevant sensitivity 0 and filtered-single −.250 are provenance only, not comparable to corrected v2. CERT_CAP1 rejected; never reuse/promote/ship (results)
CAP1 schema-guarded candidate successor 8 matched scratch scorer checkpoints outputs/experiments/cap1-schema-guarded-candidate-successor-20260729/ (local, explicit no-sync) Invalidated v1-suite fixture: mini-flow prompts asserted the forbidden mode. Historical .667 sensitivity and −.250 single-arm effect are provenance only, not comparable to corrected v2. CERT_CAP1 rejected; never reuse/promote/ship (results)
CAP1 pack-generic candidate successor 8 matched scratch scorer checkpoints outputs/experiments/cap1-pack-generic-candidate-successor-20260729/ (local, explicit no-sync) Invalidated v1-suite fixture: mini-flow prompts asserted the forbidden mode. Historical .55–.725 canonical range is provenance only, not comparable to corrected v2. CERT_CAP1 rejected; not reusable/promoted/ship (results)
Deadline-classification callout diagnostic parser-fork-smoke-callout/last.pt outputs/autoresearch/… (local, no-sync) Same 1,610,114-param deterministic checkpoint SHA; timeout laundering was removed: smoke/held-out/OOD became explicit runtime timeouts with no fallback output or grammar dead end. The safety repair is retained, but callout meaning and overall gates failed, AgentV was 0/5—not reusable/promoted/ship (results)
Parser-fork button diagnostic parser-fork-smoke-button/last.pt outputs/autoresearch/… (local, no-sync) Same 1,610,114-param deterministic checkpoint SHA; button compiler was 14,684.025 ms and total 16,276.909 ms with parse 1.0 and zero fallback. The locked smoke endpoint passed, but held-out/RICO fell back, adversarial timed out, AgentV was 0/5, and compound/overall decisions failed—not reusable/promoted/ship (results)
Parser-fork aggregate diagnostic parser-fork-smoke-hero/last.pt outputs/autoresearch/… (local, no-sync) Same 1,610,114-param deterministic checkpoint SHA; exact parser forks cut hero compiler time 23.0% to 21,650.370 ms and total to 23,195.864 ms. Locked runtime endpoint passed, but five subsets were only n=1, OOD used certified fallback, AgentV 0/5, and overall gates failed—not reusable/promoted/ship (results)
Incremental-lex aggregate diagnostic incremental-lex-smoke-hero/last.pt outputs/autoresearch/… (local, no-sync) Same 1,610,114-param deterministic checkpoint SHA; hero compiler fell another 555.999 ms and total cleared the deadline at 29,812.095 ms with zero fallback. Locked diagnostic endpoint passed, but five subsets were only n=1, OOD timed out, AgentV 0/5, and overall gates failed—not reusable/promoted/ship (results)
Lexer-cache aggregate diagnostic lexer-cache-smoke-hero/last.pt outputs/autoresearch/… (local, no-sync) 16-step / 1,610,114-param CPU scratch; hero compiler cost fell 1.84% but total remained 30,003.055 ms with certified fallback, five subsets were only n=1, AgentV 0/5 — locked hypothesis rejected, not reusable/promoted/ship (results)
SLM-313 AbstractPlan local checkpoint slm313_local_plan_1k_v2/last.pt outputs/runs/… (local, no-sync) 9-step / 1,006-token CPU scratch plan-head + learned connector; complete locked 6,102-row matrix gives zero meaningful-v2/binder-F1 effect vs destructive controls—rejected, not reusable, promoted, or ship (evidence)
SLM-322 AP-027 Pareto screening checkpoint slm322_ap027_scratch_v1/last.pt outputs/runs/… (local, no-sync) 8-step connector-free CPU scratch checkpoint, trained only to load the current symbol_only/v2 output contract; 1-seed screening decode at refinement rounds 1-2, rounds 4/8 and connector-dependent arms pending—wiring only, not promoted or ship (evidence)
E1211 seed-7 topology-dose control e1211_v273_e937_batch4_seed7_lr1e4_binder_topology_quarter/last.pt outputs/runs/… (local, no-sync) 395 × 4 CPU scratch draws; strict held n=5 exactly matches E1182/E1200 (parse/meaning .4, strict .2, fidelity .28, structure .2852, recall .3333, reward .3388, 3 timeouts)—neutral/rejected, not promoted or ship (results)
SLM-298 local factorial diagnostics 20 local d32/d64 scratch checkpoints outputs/runs/slm298_local_factorial*/cells/ (local, no-sync) 520 strict rows; locked n=1; all completed constrained cells syntax 1.0 but strict meaningful/binder F1 0.0; four d32 seed-2 cells cap-censored — rejected, not promoted or ship (results)
SLM-287 five-seed locked baseline slm287-trained-local-v13-20260725 (10 cells) outputs/runs/… (local, no-sync) CPU float32 Choice TwoTower; 97-record strict train snapshot, 5k-token budget, five seeds × scratch design off/on, 226-record locked holdout and AgentV per shard. Meaning-v2/binder F1 stay 0; absolute-probability MDE 2.0 pp. Diagnostic only, not promoted or ship
SLM-230 bounded recurrence diagnostic slm230_bounded_recursive_r4_r2/last.pt outputs/runs/… (local) R=4 scratch checkpoint: SLM-230 is stagnant; SLM-231 is expansive_unstable; SLM-232 finds z0 rank 2.1054 but rank 0 after context/position removal and vacuous bounded ablations (unstable). Rejected, not ship
E735 full-head root-arity diagnostic e735-symbol-only-root-arity-fullhead140-r1/last.pt outputs/runs/… (local) Removes impossible class-41 tail prediction, but weight 0/1 smoke quality remains identical and strict-v2 0.0 — fix retained, checkpoint rejected
E733 invalid lexer root-identity attempt e733-symbol-only-root-identity140-r1/last.pt outputs/runs/… (local) Proposed lever has zero reachable decode applications; config now rejects lexer identity before artifacts — checkpoint invalidated
E731 lexer root-arity diagnostic e731-symbol-only-root-arity140-r1/last.pt outputs/runs/… (local) Lexer-native head is executable, but weights 0/1/2 change no choices; smoke strict-v2 0.0 — checkpoint rejected
E714 symbol-only baseline e714-symbol-only-scratch600-r1/last.pt outputs/runs/… (local) First v2-compatible CPU scratch checkpoint; 600 steps / 48.72s, strict meaning 0.0 and AgentV 0/5 — diagnostic only, not ship
E720 component-inventory diagnostic e720-symbol-only-component-inventory600-r1/last.pt outputs/runs/… (local) Inventory head learned (top-k recall 0.6875), but smoke parse/strict meaning remained 0.0 and weight-4 decode timed out 3/3 — rejected, not ship
E721 role/count plan diagnostic e721-symbol-only-component-plan190-r4/last.pt outputs/runs/… (local) Smoke parse 1.0, but strict meaning 0.0 and plan weight 1 is identical to weight 0; local 190-step syntax diagnostic only, rejected
E722 component-edge diagnostic e722-symbol-only-component-edge150-r1/last.pt outputs/runs/… (local) Parse 1.0 / structure 0.2861 / recall 0.5, but strict meaning 0.0 and edge on/off identical — rejected, not ship
E723 slot-owner diagnostic e723-symbol-only-slot-owner140-r1/last.pt outputs/runs/… (local) Causal smoke + held-out gains; smoke meaning-v1 0.6667 / structure 0.5614, but strict-v2 0.0 — lever retained, checkpoint rejected
E725 cumulative inventory diagnostic e725-symbol-only-component-inventory130-r1/last.pt outputs/runs/… (local) Inventory head learned, but weight 1/0 decode is identical and smoke meaning-v1/strict-v2 0.0 — rejected, not ship
E726 invalid root-arity attempt e726-symbol-only-root-arity140-r1/last.pt outputs/runs/… (local) Choice-only arity lever was unavailable on lexer; tensors match E723 exactly — invalidated, never evaluate/sync/serve
E727 binder-arity diagnostic e727-symbol-only-binder-arity140-r1/last.pt outputs/runs/… (local) Arity head learned, but weights 1/2 change no smoke or held-out choices and strict-v2 remains 0.0 — rejected
E729 binder-topology diagnostic e729-symbol-only-binder-topology140-r1/last.pt outputs/runs/… (local) Topology weights 0.25/1 regress smoke meaning 0.6667→0.3333 and structure 0.5614→0.4642 — rejected
Playground demo playground_demo/last.pt src/slm_training/resources/checkpoints/playground_demo/ (git) E497 clean-revision honest smoke: parse/meaningful/fidelity 0.0, structure 0.2203, AgentV 0/5; wiring only
Restructure CPU verify restructure_cpu_scratch_v0/last.pt outputs/runs/… (local) Fixture scratch train OK; smoke parse 0.0 — not ship
Local DirectML verify local_directml_adreno_20260714/last.pt outputs/runs/… (local) Adreno GPU train/checkpoint OK; 5-step wiring run, not evaluated or ship
Overnight retrain overnight_retrain_200/last.pt /tmp/slm-training-overnight/outputs/runs/… (local) 200-step CPU scratch; honest parse 0.0, not ship
Overnight retrain extended overnight_retrain_1000/last.pt /tmp/slm-training-overnight/outputs/runs/… (local) 1,000-step CPU scratch; smoke parse 0.0, not ship
E120 singleton diagnostic e120_unsandboxed/last.pt outputs/runs/iter-e120-unsandboxed-20260715/… (local) 8-step CPU scratch; guarded singleton decode verified, rico_held n=1 parse 0.0 — not ship
E121 judged-corpus E53 iteration qx_e53_honest_v5_champion/last.pt outputs/runs/iter-e121d-e53-judged-20260715/… (local) 405 judge-approved records; bounded smoke parse 0.0 with decode timeout — not ship
E123 judged-corpus 32-step iteration e123_judged_32step_b/last.pt outputs/runs/iter-e123b-judged-20260715/… (local) 405 judge-approved records; loss 10.97 but smoke parse 0.0 with fallback/canvas cap — not ship
E127 schema/slot-contract iteration e127_judged_schema_slots/last.pt outputs/runs/iter-e127-schema-slots-20260715/… (local) 405 judged records; placeholder validity 0.55 / normalized fidelity 0.25, but parse 0.0 — not ship
E128 schema/slot 64-step iteration e128_judged_schema_slots_64/last.pt outputs/runs/iter-e128-schema-slots-20260715/… (local) Higher LTR/fidelity weights regressed placeholder signals and parse remained 0.0 — not ship
E129 schema/slot 64-step low-weight control e129_judged_schema_slots_64_lowweights/last.pt outputs/runs/iter-e129-schema-slots-20260715/… (local) Lower-weight control also had placeholder/parse 0.0; longer training not justified — not ship
E130 schema/slot seed-1 control e130_judged_schema_slots_seed1/last.pt outputs/runs/iter-e130-schema-slots-20260715/… (local) Seed-1 control had parse and placeholder signals 0.0; E127 not reproducible — not ship
E132 generation-focused mixture e132_generation_focus/last.pt outputs/runs/iter-e132-generation-focus-20260715/… (local) Three-prompt smoke parse/placeholder 0.0; task reweighting rejected — not ship
E133 no-fused-LTR path e133_no_fuse_ltr/last.pt outputs/runs/iter-e133-no-fuse-ltr-20260715/… (local) Three-prompt smoke parse/structure 0.0 with one timeout; fused LTR retained — not ship
E135 HF context control e135_hf_context_control/last.pt outputs/runs/iter-e135-hf-context-20260715/… (local) HF context improves structural/placeholder signals but parse 0.0 with one timeout — not ship
E136 HF context 32-step control e136_hf_context_32/last.pt outputs/runs/iter-e136-hf-context-20260715/… (local) Longer HF run regressed structure/placeholder to 0.0; checkpoint selection next — not ship
E137 HF context 16-step midpoint e137_hf_context_16/last.pt outputs/runs/iter-e137-hf-context-20260715/… (local) Placeholder validity 0.40 and structure 0.2142, parse 0.0; non-monotonic checkpoint trajectory — not ship
E138 HF context seed-1 8-step control e138_hf_context_seed1_8/last.pt outputs/runs/iter-e138-hf-seed1-20260715/… (local) Same recipe as E135 but seed 1: placeholder validity 0.0 and structure 0.1683, parse 0.0 — not ship
E139 HF context seed-2 8-step control e139_hf_context_seed2_8/last.pt outputs/runs/iter-e139-hf-seed2-20260715/… (local) Same recipe as E135 but seed 2: placeholder validity/structure/parse 0.0 with two timeouts — not ship
E173 schema-context 32-step control e173-schema-context-32step/last.pt outputs/runs/e173-schema-context-32step/… (local) Schema/slot context enabled; bounded syntax probe 1.0 but meaningful parse 0.0 — not ship
E174 unfrozen-context 8-step control e174-unfrozen-context-8step/last.pt outputs/runs/e174-unfrozen-context-8step/… (local) Unfrozen context regressed bounded syntax to 0.0; rejected control — not ship
E175 retrieval 8-step control e175-retrieval-8step/last.pt outputs/runs/e175-retrieval-8step/… (local) Retrieval k=4 regressed bounded syntax/parse to 0.0; rejected control — not ship
E176 broad-corpus 8-step control e176-broad-corpus-8step/last.pt outputs/runs/e176-broad-corpus-8step/… (local) 1,417-record corpus regressed bounded syntax/parse to 0.0; rejected control — not ship
E177 semantic-judge 32-step control e177-semantic-judge-32step/last.pt outputs/runs/e177-semantic-judge-32step/… (local) 496 published judge-gated records; E180 bounded decode reaches syntax 1.0 but meaningful parse 0.0 — not ship
E181/E184/E191 compiler-alignment diagnostics e181-semantic-balanced-32step, e184-compiler-aligned-32step, e191-full-compiler-aligned-32step outputs/runs/… (local) Balanced mixture did not improve quality; component alignment recovered the root, all-branch alignment regressed it; no meaningful parse or promotion — not ship
E195/E196 stratified-alignment diagnostics e195-stratified-compiler-aligned-32step, e196-stratified-compiler-aligned-matched-32step outputs/runs/… (local) E195 invalid (mixture unset); matched E196 reaches syntax 1.0 after parser-state fixes but meaningful parse 0.0 — not ship
E201 generated-role diagnostic e201-role-stratified-compiler-aligned-32step outputs/runs/… (local) Grammar/schema role constraints improve component and placeholder signals, but recursive children hit the token cap with parse 0.0 — not ship
E205 Lark-terminal diagnostic e205-lark-terminal-stratified-32step outputs/runs/… (local) Terminal-derived alignment and schema enum paths restore syntax 1.0 without fallback, but empty bound stacks leave meaningful parse 0.0 — not ship
E208/E210/E212 contextual-decision diagnostics e208-list-occupancy-stratified-32step, e210-list-scope-occupancy-stratified-32step, e212-contextual-decision-stratified-32step outputs/runs/… (local) Contextual root-child supervision recovers a populated root and fidelity signal, but required schema semantics still fail and meaningful parse remains 0.0 — not ship
E214/E215 overfiltered schema-judge diagnostic e215-schema-role-judged-32step outputs/runs/e215-schema-role-judged-32step/… (local) E214 falsely rejected 27 legal optional-null records; E216 syntax 1.0 but meaningful parse 0.0; superseded by E218 — not ship
E218/E219 corrected schema-admission diagnostic e219-schema-normalized-32step outputs/runs/e219-schema-normalized-32step/… (local) Restores 33 valid records and fixes future producers; E220 syntax 1.0, component recall 0.25, meaningful parse 0.0 — not ship
E221 task-balanced exposure diagnostic e221-canonical-task-balanced outputs/autoresearch/e221-task-balanced-exposure-v4/runs/… (local) 32 CPU steps on canonical E218; effective exposure 29.68/128; strict eval failed 9 gates, AgentV 1/5 — not ship
E222 capacity-aware exposure diagnostic e222-capacity-aware-matched outputs/autoresearch/e222-capacity-aware-exposure/runs/… (local) Effective exposure rose to 83.59/128, but strict smoke parse regressed to 0.0 and 10 gates failed — not ship
E223 quota-capacity exposure diagnostic e223-quota-capacity-matched outputs/autoresearch/e223-quota-capacity-exposure/runs/… (local) Task quotas and syntax are deterministic, but semantic metrics are 0.0 and 12 gates failed — not ship
E224–E226 semantic alignment + honest tree eval e224-semantic-exhaustive-matched outputs/autoresearch/e224-semantic-exhaustive-alignment/runs/… (local) Deterministic tree reaches syntax 1.0 on all suites with honest fidelity, but meaningful-program quality fails 5 gates — not ship
E227 legal-candidate alignment e227-candidate-set-matched outputs/autoresearch/e227-candidate-set-alignment/runs/… (local) Candidate loss optimizes, but empty-layout collapse fails 12 gates and AgentV 0/5 — rejected, not ship
E228 legal-candidate margin e228-candidate-margin-matched outputs/autoresearch/e228-candidate-margin-alignment/runs/… (local) Best diagnostic: syntax/contract 1.0, failures reduced to 4, but AgentV 1/5 — not ship
E229 64-step margin continuation e229-margin-64step outputs/autoresearch/e229-margin-continuation/runs/… (local) Syntax restored to 1.0 after generalized literal-frame fix, but the same 4 gates fail — duration rejected, not ship
E230 diverse judged roots e230-diverse-roots-32step outputs/autoresearch/e230-diverse-judged-roots/runs/… (local) Published 126 judge-passed generation roots and verified RICO/human exposure; same 4 gates fail and adversarial regresses — data fix retained, checkpoint rejected, not ship
E231 component inventory e231-component-inventory-32step outputs/autoresearch/e231-component-inventory/runs/… (local) Inventory target learns, but bias-off metrics/component choices are identical; 6 thresholds fail, AgentV 1/5 — rejected, not ship
E232 role component plan e232-role-component-plan-32step outputs/autoresearch/e232-role-component-plan/runs/… (local) Root/count targets learn and improve one adversarial case, but 4 frontier thresholds still fail; stronger calibration has no aggregate gain — rejected, not ship
E233 resolved-AST component edges e233-component-edges-32step outputs/autoresearch/e233-component-edges/runs/… (local) Edge target learns, but edge on/off suite aggregates are identical and 4 thresholds fail — rejected, not ship
E234 edge decision alignment e234-edge-decision-alignment-32step outputs/autoresearch/e234-edge-decision-alignment/runs/… (local) Legal-decision accuracy learns and changes 5 choices, but on/off aggregates are identical and 4 thresholds fail — rejected, not ship
E235 binder-instance plan e235-binder-instance-plan-32step outputs/autoresearch/e235-binder-instance-plan/runs/… (local) Full binder supervision changes 4 legal choices, but on/off aggregates are identical and 9 thresholds fail — rejected, not ship
E236 binder topology e236-binder-topology-32step outputs/autoresearch/e236-binder-topology/runs/… (local) Topology objective fails to learn, changes 0/38 applied choices, and collapses semantic metrics; 12 thresholds fail — rejected, not ship
E237 detached topology e237-detached-topology-32step outputs/autoresearch/e237-detached-topology/runs/… (local) Detaching already-frozen context is a no-op and exactly reproduces E236; 12 thresholds fail — rejected, not ship
E238 binder arity (invalidated) e238-binder-arity-32step outputs/autoresearch/e238-binder-arity/runs/… (local) Optional-head RNG shifted matched training draws; ten thresholds fail and the run is confounded — not ship
E239 isolated binder arity e239d-binder-arity-fully-isolated-32step outputs/autoresearch/e239-binder-arity-corrected/runs/… (local) 104/104 shared tensors match the control; 29 changed choices do not produce meaningful programs; 11 thresholds fail — rejected, not ship
E249 exact-event CE plus margin qx_e249_local_ce_margin outputs/autoresearch/e249-local-ce-margin/runs/… (local) Held-out lexical wins improve sharply, but structure/reward regress on every suite and AgentV is 0/5 — rejected, not ship
E252 verifier-backed set FTPO qx_e252_local_ftpo_set outputs/autoresearch/e252-ftpo-set/runs/… (local) Syntax remains 1.0, but fidelity collapses to 0, structure/reward regress everywhere, and AgentV is 0/5 — rejected, not ship
E263 broad gold-AST set FTPO qx_e262_broad_gold_ast_ftpo_set outputs/autoresearch/e262-broad-gold-ast-ftpo/runs/… (local) Emitted as E262 before ID reconciliation; syntax/fidelity match E248, but held-out loss worsens, structure regresses everywhere, and AgentV is 0/5 — rejected, not ship
E264 guarded gold-AST set FTPO qx_e264_guarded_gold_ast_ftpo_set outputs/autoresearch/e264-guarded-gold-ast-ftpo/runs/… (local) No trained step passed the held-out Pareto guard; restored checkpoint is bit-identical to E228 and current parent control reproduces all metrics — no model gain, not ship
E265 safe gold-AST set FTPO qx_e265_safe_gold_ast_ftpo_set outputs/autoresearch/e265-safe-gold-ast-ftpo/runs/… (local) 3/30 backtracked proposals improve aggregate exact-state metrics, but per-kind regressions are masked and semantic quality falls on most suites — rejected, not ship
E266 stratified safe set FTPO qx_e266_stratified_safe_gold_ast_ftpo_set outputs/autoresearch/e266-stratified-safe-gold-ast-ftpo/runs/… (local) Per-decision-kind guard rejects all 30 global FTPO proposals; parent is restored exactly, while batched validation is 37.7× faster — no model gain, not ship
E267 block-coordinate safe set FTPO qx_e267_block_stratified_safe_gold_ast_ftpo_set outputs/autoresearch/e267-block-stratified-safe-ftpo/runs/… (local) Averaging gradients within each decision kind still yields 0/30 safe proposals; parent is restored exactly — no model gain, not ship
E268 projected safe set FTPO qx_e268_projected_stratified_safe_gold_ast_ftpo_set outputs/autoresearch/e268-projected-stratified-safe-ftpo/runs/… (local) PCGrad projects 2,220 conflicting task pairs but still yields 0/30 safe proposals; parent restored exactly, 38m59s CPU stage — rejected, not ship
E269 MGDA safe set FTPO qx_e269_mgda_stratified_safe_gold_ast_ftpo_set outputs/autoresearch/e269-mgda-one-step-final/runs/… (local) One-step MGDA certifies common train descent, but all five scales regress held-out decision kinds; full 30-step run rejected, parent restored — not ship
E272 MGDA plus SGD preflight qx_e272_mgda_sgd_stratified_safe_gold_ast_ftpo_set outputs/autoresearch/e272-mgda-sgd-one-step/runs/… (local) Collinear SGD improves aggregate held-out loss, but all scales regress per-kind probability/margin guards; parent restored, no full run — not ship
Matrix honest champion V6 E53 family outputs/runs/ + matrix docs Scratch + limited rico_held — not production HF ship
P13 matched E50 controls fixture + integrated E50 /tmp/slm17-e50-*-honest/ (local scratch) Integrated fidelity +0.04 held / +0.0333 RICO; parse 0.0, not ship
Frozen X2 baseline gx_x2_codec seeds 0/1/2 /tmp/slm-training-fixed-baseline/outputs/topology_baseline/ Fixed-canvas comparison scored zero on all suites; not ship
Topology v2 smoke grammar_diffusion_overfit pytest temporary checkpoint n=2 parse/fidelity 0.5, topology composite 0.482; wiring only, not ship
Topology X9/X14 confirmation 6 seed checkpoints /tmp/slm-training-grammar-topology/outputs/topology_confirm_4bf964d/ 200-step CPU scratch; all fail multi-suite gates, no promotion/sync
ScopeDiff X18/X21 confirmation 6 seed checkpoints outputs/runs/gx_x{18,21}_*_confirm_200/ (local) 200-step CPU scratch; all-suite median parse/fidelity 0.0, all fail gates, no promotion/sync
EFS0-04 X22 reproduction gx_x22_kapur_tree_edit_s0/last.pt outputs/runs/gx_x22_kapur_tree_edit_s0/… (local) 80-step seed-0 audit-material replay; SHA a9cfb450…02ff6; syntax 1.0 but meaningful parse 0.333/0.2/0/0/0.667 on bounded suites; gates fail, no sync/promotion (results)
B3 five-minute lexer control capacity_lexer_v1__d64_h2_c1_dn2_t5000_x1__s0/last.pt outputs/ladders/b3-matched-5m-e287-r2/… (local) 53-step / 5,004-token CPU scratch; five-suite parse/meaningful 0.0, AgentV 0/5 — not promoted or ship
B3 five-minute choice arm capacity_choice_v1__d64_h2_c1_dn2_t5000_x1__s0/last.pt outputs/ladders/b3-matched-5m-e287-r2/… (local) E288 frozen eval: deterministic parse 1.0 on all suites, but meaningful/fidelity 0.0 and AgentV 0/5 — not promoted or ship
E289 cached choice arm capacity_choice_v1__d64_h2_c1_dn2_t5000_x1__s0/last.pt outputs/ladders/e289-choice-state-cache/… (local) Same checkpoint SHA as E288; exact symbolic-state cache preserves parse 1.0 and cuts p50 2.65×–5.86×, but semantic metrics and AgentV remain zero — not promoted or ship
E290 direct-candidate choice arm capacity_choice_v1__d64_h2_c1_dn2_t5000_x1__s0/last.pt outputs/ladders/e290-choice-direct-candidates/… (local) Same checkpoint SHA; exact grammar-derived candidates improve p95 1.14×–1.19× but regress p50, while semantic metrics and AgentV remain zero — not promoted or ship
E291 completion-cached choice arm capacity_choice_v1__d64_h2_c1_dn2_t5000_x1__s0/last.pt outputs/ladders/e291-choice-completion-cache/… (local) Same checkpoint SHA; exact completion caching improves p50 1.29×–1.99× and p95 1.51×–1.93× vs E290, but semantic metrics and AgentV remain zero — not model-promoted or ship
E292 complete-loss choice arm capacity_choice_v1__d64_h2_c1_dn2_t5000_x1__s0/last.pt outputs/ladders/e292-choice-loss-suite-complete-r2/… (local) Same checkpoint SHA; all five frozen loss categories now complete (weighted NLL 7.2265), but honest meaningful rate is 0.0 and AgentV is 0/5 — not promoted or ship
E293 choice-native component plan e293-choice-component-plan-r3/last.pt outputs/runs/e293-choice-component-plan-r3/… (local) Plan target learns and legal bias reduces failures 17→13, but matched no-DESIGN meaningful rate is 0.0 and AgentV 0/5 — not promoted or ship
E294 no-DESIGN choice control e294-choice-no-design-control-r1/last.pt outputs/runs/e294-choice-no-design-control-r1/… (local) No-plan control exactly matches E293 bias-off; meaningful 0.0, AgentV 0/5, 17 failures — not promoted or ship
E295 DESIGN-dropout choice arm e295-choice-design-dropout-r1/last.pt outputs/runs/e295-choice-design-dropout-r1/… (local) 50% deterministic context dropout yields adversarial meaningful 0.25 and AgentV 1/5, but four suites remain 0.0 and 14 gates fail — not promoted or ship
E396 durable diagnostic checkpoint e396-balanced-type-head-continuation-r1/last.pt hf://buckets/TKendrick/OpenUI/checkpoints/e396-balanced-type-head-continuation-r1/ Exact SHA feefa056…c2f2eee0; bucket verified. E498 restores current-main loading and learned-head application (smoke structure 0.27057), but semantic gates and AgentV remain red. Diagnostic, not champion or ship
E499 bounded strict-corpus checkpoints e499-*-r4/r6/last.pt outputs/runs/e499-…/ (local) Matched strict-r4 and document-only r6 both regress smoke structure 0.1542→0.0375 and recall 0.25→0.0; AgentV 0/1, no sync or promotion. All seven checkpoint SHAs are in the full card
E500 documentized-expression checkpoints e500-*-r1/r2/r3-5k/r4-5k/last.pt outputs/runs/e500-…/ (local) The 260-row projected corpus is clean and diverse, but both matched 1k/5k pairs have structure 0.0375, semantic metrics zero, and AgentV 0/1. Four exact SHAs are in the full card; no sync or promotion
E501 E396→E500 warm-start checkpoints e501-e396-e500-*/last.pt outputs/runs/e501-…/ (local) Explicit new-corpus initialization works, but 5k arms forget parent structure; the 1k arm reaches structure 0.2317 with semantic metrics still zero. Three exact SHAs are in the full card; no sync or promotion
E502 prior-retention checkpoints e502-e396-e500-*/last.pt outputs/runs/e502-…/ (local) Preserving checkpoint serving priors raises 1k structure to 0.3169 with recall 0.0833, but 5k collapses and all semantic gates remain zero. Four exact SHAs are in the full card; no sync or promotion
E503 initialized-weight retention checkpoints e503-e396-e500-retention*-5k/last.pt outputs/runs/e503-…/ (local) Retention cuts RMS drift up to 74% and restores structure to 0.2029, but recall falls to zero and semantic gates remain red. Four exact SHAs are in the full card; no sync or promotion
E504 parent-replay checkpoints e504-e396-e500-replay*-5k/last.pt outputs/runs/e504-…/ (local) 50% exact E357 replay raises structure to 0.2469 and cuts drift 10.46%, but semantic gates remain zero; replay plus retention regresses structure. Five exact SHAs are in the full card; no checkpoint sync or promotion
E505 replay-loss attribution checkpoint e505-e396-e500-replay050-loss-attribution-r1-5k/last.pt outputs/runs/e505-…/ (local) E511 component-plan weight 4 reaches aggregate meaningful 0.3846 and fidelity 0.6718 across 13 records. E512 rejects slot weight 8; strict semantic and AgentV gates remain red, with no promotion
E513 durable slot-role checkpoint e513-e396-e500-replay050-slotrole4-focal2-r3-5k/last.pt hf://buckets/TKendrick/OpenUI/checkpoints/e513-e396-e500-replay050-slotrole4-focal2-r3-5k/ Bucket-verified SHA 59253c67…a88a9548; 5,000 target tokens in 79.6s under the three-minute cap. Matched OOD meaningful 0.0, fidelity 0.4917, structure 0.2750, AgentV 0/1; durable diagnostic, rejected for promotion
E515 focal-zero slot-role checkpoint e515-e396-e500-replay050-slotrole4-focal0-r1-5k/last.pt hf://buckets/TKendrick/OpenUI/checkpoints/e515-e396-e500-replay050-slotrole4-focal0-r1-5k/ Bucket-verified SHA 97f2e426…24721c1b; 5,000 target tokens in 105.8s under the three-minute cap. Matched OOD meaningful 0.25, fidelity 0.6583, structure 0.3213, AgentV 0/1; focal 2 rejected, checkpoint not promoted
E517 slot-loss-1 context checkpoint e517-e396-e500-replay050-slotrole1-context-r1-5k/last.pt hf://buckets/TKendrick/OpenUI/checkpoints/e517-e396-e500-replay050-slotrole1-context-r1-5k/ Bucket-verified SHA 2b572a04…e24b60e3; 5,000 target tokens in 130.7s under the three-minute cap. Matched OOD meaningful 0.0, fidelity 0.4083, structure 0.2250, AgentV 0/1; rejected
E519 honest slot-context checkpoint e519-e396-e500-replay050-slotrole1-honest-context-r1-5k/last.pt hf://buckets/TKendrick/OpenUI/checkpoints/e519-e396-e500-replay050-slotrole1-honest-context-r1-5k/ Bucket-verified SHA d82155b0…6c91805f; 5,000 target tokens in 103.2s from clean harness v7. Exact E517 quality parity (meaningful 0.0, fidelity 0.4083, structure 0.2250, AgentV 0/1); honest path retained, checkpoint rejected
E522 visible-inventory checkpoint e522-e396-e521-replay050-slotrole1-honest-context-r2-5k/last.pt hf://buckets/TKendrick/OpenUI/checkpoints/e522-e396-e521-replay050-slotrole1-honest-context-r2-5k/ Bucket-verified SHA 97cb10f4…bf420ce; 5,059 target tokens in 120.7s. E523 fidelity 0.8667 and recall 0.2708 improve, but meaningful stays 0.0, structure falls to 0.1955, and AgentV is 0/1; rejected
E525 visible-component checkpoint e525-e396-e524-replay050-slotrole1-honest-context-r2-5k/last.pt hf://buckets/TKendrick/OpenUI/checkpoints/e525-e396-e524-replay050-slotrole1-honest-context-r2-5k/ Bucket-verified SHA dbd11811…e55e4b9; 5,059 target tokens in 76.7s. E526 recall rises to 0.4167, but fidelity falls to 0.4667, structure to 0.1452, meaningful stays 0.0, and AgentV is 0/1; rejected
E528 visible-component-types checkpoint e528-e396-e527-replay050-slotrole1-honest-context-r1-5k/last.pt hf://buckets/TKendrick/OpenUI/checkpoints/e528-e396-e527-replay050-slotrole1-honest-context-r1-5k/ Bucket-verified SHA 6a2180d7…306976d5; 5,059 target tokens in 146.8s. E529 meaningful recovers to 0.25 and reward to 0.5778, but structure falls to 0.1136, strict meaning remains 0.0, and AgentV is 0/1; rejected
E616 object-frame slot-bias replay (80-step) e616-object-property-slot-bias-scratch80-20260720/last.pt outputs/runs/e616-object-property-slot-bias-scratch80-20260720/ (local) Fresh 80-step CPU scratch loop on E530, loss 26.5243; matched OOD n=4 eval now parses 4/4 but stays byte-identical because Gallery's array closes empty before any item opens — not ship
E620 required-slot coverage replay (800-step) e620-required-slot-coverage-scratch800-20260720/last.pt outputs/runs/e620-required-slot-coverage-scratch800-20260720/ (local) 800 CPU scratch steps in 80.96s, loss 4.0680; OOD treatment fidelity 0.5500, structure 0.4886, strict-v2 0.0, AgentV 0/1. Lower loss regressed E619 generalization — rejected, not ship
E548 fresh TwoTower loop e548_training_loop_twotower_scratch_20260720/last.pt outputs/runs/e548_training_loop_twotower_scratch_20260720/ (local) Fresh 8-step CPU scratch loop on E530, loss 39.4267; no eval/sync, wiring only — not ship
E547 fresh TwoTower loop e547_training_loop_twotower_scratch_20260720/last.pt outputs/runs/e547_training_loop_twotower_scratch_20260720/ (local) Fresh 7-step CPU scratch loop on E530, loss 35.7431; no eval/sync, wiring only — not ship
E546 fresh TwoTower loop e546_training_loop_twotower_scratch_20260720/last.pt outputs/runs/e546_training_loop_twotower_scratch_20260720/ (local) Fresh 6-step CPU scratch loop on E530, loss 40.3390; no eval/sync, wiring only — not ship
E545 fresh TwoTower loop e545_training_loop_twotower_scratch_20260719/last.pt outputs/runs/e545_training_loop_twotower_scratch_20260719/ (local) Fresh 5-step CPU scratch loop on E530, loss 42.1226; no eval/sync, wiring only — not ship
E544 fresh TwoTower loop e544_training_loop_twotower_scratch_20260719/last.pt outputs/runs/e544_training_loop_twotower_scratch_20260719/ (local) Fresh 4-step CPU scratch loop on E530 after missing prior local checkpoint, loss 42.3848; no eval/sync, wiring only — not ship
E543 resumed TwoTower loop e543_training_loop_twotower_resume_scratch_20260719/last.pt outputs/runs/e543_training_loop_twotower_resume_scratch_20260719/ (local) Resumed E542 full-state to step 3 on E530, loss 39.7476; no eval/sync, wiring only — not ship
E542 resumed TwoTower loop e542_training_loop_twotower_resume_scratch_20260719/last.pt outputs/runs/e542_training_loop_twotower_resume_scratch_20260719/ (local) Resumed E541 full-state to step 2 on E530, loss 43.6742; no eval/sync, wiring only — not ship
E541 TwoTower training-loop iteration e541_training_loop_twotower_scratch_20260719/last.pt outputs/runs/e541_training_loop_twotower_scratch_20260719/ (local) One-step CPU scratch TwoTower loop on E530, loss 36.9158; no eval/sync, wiring only — not ship
E540 training-loop sentinel e540_training_loop_scratch_20260719/last.pt outputs/runs/e540_training_loop_scratch_20260719/ (local) One-step CPU scratch stub loop check on E530, loss 0.5; no eval/sync, wiring only — not ship
E531 visible-semantic-role checkpoint e531-e396-e530-replay050-slotrole1-honest-context-r1-5k/last.pt hf://buckets/TKendrick/OpenUI/checkpoints/e531-e396-e530-replay050-slotrole1-honest-context-r1-5k/ Bucket-verified SHA 6b8c1abc…74a6154; 5,059 target tokens in 99.72s. E532 structure improves slightly to 0.1431, but meaningful falls to 0.0, fidelity to 0.4667, reward to 0.3685, strict meaning stays 0.0, and AgentV is 0/1; rejected
E542 learned root-arity checkpoint e542-e531-root-reference-arity1-r1-24s/last.pt outputs/runs/e542-e531-root-reference-arity1-r1-24s/ (local) 24-step scratch continuation in 52.93s, SHA 2d5cd4b3…6854c5d8; OOD n=4 meaningful 0.50 / fidelity 0.5917 / structure 0.3019, but learned weight 1 is quality-neutral, strict meaning 0.0, AgentV 0/1; no sync or promotion
E543 bounded root-arity checkpoint e543-e531-root-reference-bounded-r1-24s/last.pt outputs/runs/e543-e531-root-reference-bounded-r1-24s/ (local) 24-step scratch continuation in 37.17s, SHA c6be3791…51d7f90; bounded loss improves calibration, but OOD n=4 decisions and quality exactly match E542, strict meaning 0.0, AgentV 0/1; no sync or promotion
E544 root-identity checkpoint e544-e543-root-identity1-r2-24s/last.pt outputs/runs/e544-e543-root-identity1-r2-24s/ (local) 24-step scratch continuation in 40.96s, SHA 3b6e3c00…474f20c; rank-only identity decode raises OOD n=4 meaningful 0.00→0.25, structure 0.1250→0.1688, and recall 0.1458→0.2708, but strict meaning 0.0 and AgentV 0/1; no sync or promotion
E545 matched negative-weight checkpoints e545-e544-root-identity-neg{1-control,4}-r*/last.pt outputs/runs/e545-…/ (local) Matched 24-step scratch continuations in 30.64s / 28.64s, SHAs 9e54d470…76fa1 / 14dd4404…61ae; weight 4 slightly improves sparse late negative accuracy, but predictions and OOD n=4 metrics are identical, both regress from E544, and AgentV is 0/1; no sync or promotion
E546 matched strict-subset checkpoints e546-e544-strict-subset{1-control,5}-r*/last.pt outputs/runs/e546-…/ (local) Multiplier 5 raises strict-negative exposure 7→22 rows and improves OOD n=4 fidelity 0.4250→0.6083, structure 0.1494→0.2038, reward 0.5078→0.8120, and AST edge F1 0→0.0417, but recall falls 0.2083→0.0625, meaning remains 0, AgentV 0/1; no sync or promotion
E547 moderate strict-subset checkpoint e547-e544-strict-subset2-r1-24s/last.pt outputs/runs/e547-e544-strict-subset2-r1-24s/ (local) 24-step multiplier-2 scratch run in 36.48s, SHA 37002bfd…0fc57; OOD n=4 structure 0.2248 and AST node F1 0.3270 lead the 1/2/5 ladder while recall stays 0.2083, but fidelity falls to 0.2583, meaning remains 0, AgentV 0/1; no sync or promotion
E551 no-lexeme-prior checkpoint e551-e544-strict-subset2-no-lexeme-r1-24s/last.pt outputs/runs/e551-e544-strict-subset2-no-lexeme-r1-24s/ (local) 24-step scratch run in 41.85s, SHA e7921e66…dac32fc6; fidelity improves to 0.3000, but structure falls to 0.1594 and recall to 0.1250; meaning 0, AgentV 0/1; no sync or promotion
E552 half-strength lexeme-prior checkpoint e552-e544-strict-subset2-lexeme05-r1-24s/last.pt outputs/runs/e552-e544-strict-subset2-lexeme05-r1-24s/ (local) 24-step scratch run in 34.75s, SHA 49a9c111…a151fc04; fidelity 0.1333, structure 0.2181, recall 0.1250, reward 0.3435; meaning 0, AgentV 0/1; no sync or promotion
E553 corpus-local proportional-prior checkpoint e553-e544-prior-proportional-r3-24s/last.pt outputs/runs/e553-e544-prior-proportional-r3-24s/ (local) 24-step scratch run in 34.48s, SHA 510e55cf…e75399d; fidelity 0.3000, structure 0.1244, recall 0.0625, reward 0.5453; meaning 0, AgentV 0/1; no sync or promotion
E554 next-slot-context checkpoint e554-e544-slot-next-context-r2-24s/last.pt outputs/runs/e554-e544-slot-next-context-r2-24s/ (local) 24-step scratch run in 39.91s, SHA af3cbce7…c67b579; fidelity 0.2583, structure 0.1594, recall 0.1250, reward 0.5328; meaning 0, AgentV 0/1; no sync or promotion
E555 slot-pair-interaction checkpoint e555-e544-slot-pair-interaction-r2-24s/last.pt outputs/runs/e555-e544-slot-pair-interaction-r2-24s/ (local) 24-step scratch run in 50.29s, SHA af53e161…addf19e; fidelity 0.3000, structure 0.1594, recall 0.1250, reward 0.5453; Pareto lever retained, meaning 0, AgentV 0/1; no sync or promotion
E556 combined-slot-context checkpoint e556-e544-slot-context-combined-r1-24s/last.pt outputs/runs/e556-e544-slot-context-combined-r1-24s/ (local) 24-step scratch run in 68.42s, SHA 139c670c…5831f0a; fidelity 0.2167, structure 0.1594, recall 0.1250, reward 0.5203; combination rejected, meaning 0, AgentV 0/1
E557 full-balance checkpoint e557-e544-slot-pair-balance1-r1-24s/last.pt outputs/runs/e557-e544-slot-pair-balance1-r1-24s/ (local) 24-step scratch run in 70.09s, SHA 438d9871…b97db05; metrics exactly match E555; no sync or promotion
E558 owner-coverage engineering trial e558-e544-owner-coverage-r1-24s/last.pt outputs/runs/e558-e544-owner-coverage-r1-24s/ (local) 24-step scratch run in 43.31s, SHA 8a572738…de85382; dirty-tree trial persisted but excluded from decisions
E558 owner-coverage checkpoint e558-e544-owner-coverage-r2-24s/last.pt outputs/runs/e558-e544-owner-coverage-r2-24s/ (local) 24-step scratch run in 43.74s, SHA a45909df…381ede; fidelity 0.4250 but structure/reward regress and AgentV fails; no sync or promotion
E559 twofold owner-coverage checkpoint e559-e544-owner-coverage2-r1-24s/last.pt outputs/runs/e559-e544-owner-coverage2-r1-24s/ (local) 24-step scratch run in 31.14s, SHA 1d11926d…9aac861; fidelity 0.4417 and recall 0.2708, but reward 0.1643 and AgentV fails; no sync or promotion
E560 narrow owner-coverage checkpoint e560-e544-owner-threshold4-r1-24s/last.pt outputs/runs/e560-e544-owner-threshold4-r1-24s/ (local) 24-step scratch run in 42.26s, SHA dae11cee…d7686a3; structure 0.2181 and AST-node F1 0.3389, but semantic gates fail; no sync or promotion
E561 midpoint owner-coverage checkpoint e561-e544-owner-threshold7-r1-24s/last.pt outputs/runs/e561-e544-owner-threshold7-r1-24s/ (local) 24-step scratch run in 41.47s, SHA 35a4fe6d…3a127f9; fidelity 0.5750, structure 0.2419, reward 0.5753, but meaning/AgentV fail; no sync or promotion
E568 design-context continuation checkpoint e568-e561-cont48-r1-48s/last.pt outputs/runs/e568-e561-cont48-r1-48s/ (local) 48-step scratch run in 116.24s, SHA 8dcc0804…0283a12b; reward 0.6920 but fidelity/structure regress to 0.2583/0.1375 and meaning/AgentV fail; no sync or promotion
E569 matched continuation checkpoint e569-e561-matched-cont48-r1-48s/last.pt outputs/runs/e569-e561-matched-cont48-r1-48s/ (local) 48-step scratch run in 75.20s, SHA 8254fcf7…c6535f73; meaning-v1 0.25, recall 0.3333, reward 0.6920, but strict meaning/AgentV fail; no sync or promotion
E572 fidelity-loss checkpoint e572-e569-fidelity2-r1-48s/last.pt outputs/runs/e572-e569-fidelity2-r1-48s/ (local) 48-step scratch run in 84.26s, SHA bb6a58ff…cc29efa2; fidelity 0.6500 and reward 0.8170, but meaning-v1/v2 0 and AgentV fails; no sync or promotion
E573 midpoint fidelity checkpoint e573-e569-fidelity1-r1-48s/last.pt outputs/runs/e573-e569-fidelity1-r1-48s/ (local) 48-step scratch run in 109.72s, SHA ff21fc0c…cf59070d; meaning-v1 0.25, fidelity 0.4750, reward 0.7570, but strict meaning/AgentV fail; no sync or promotion
E574 slot-loss checkpoint e574-e569-slotloss2-r1-48s/last.pt outputs/runs/e574-e569-slotloss2-r1-48s/ (local) 48-step scratch run in 76.23s, SHA 649cf512…3810b7c2; aggregates exactly match E573 and strict meaning/AgentV fail; no sync or promotion
CAP5 evidence package cap5-03-evidence docs/design/calculated-arity-adaptive-precision-results.md Reproducible evidence package for CAP0–CAP4 exact calculations and controlled fixtures; not a checkpoint or ship claim (results)
Production HF ship (none yet) HF Bucket TKendrick/OpenUI checkpoints/<run_id>/ Register here after first full HF sync + --ship-gates

CAP2 capability certificate: CERT_CAP2 is not issued. The terminal SLM-385 ledger records compiler-contract-only symbolic transform and merge support, rejects learned discrete-token action benefit, and leaves the CAP1-dependent and conditional branches unavailable or unrun. DSH4 action distillation is closed; no checkpoint or ship claim changed. See the disposition.

Load demo: python -m scripts.serve_playground · Full train sync: set HF_TOKEN, then train_model --context-backend hf (auto-uploads). Details, eval tables, and history live in the model card.

Quick start

# Node.js 20-22 is required for the locked bridge and browser dependencies.
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,hf]"

# Official OpenUI parser + DESIGN.md bridges
cd src/apps/openui_bridge && npm ci && cd ../..
cd src/apps/design_md_bridge && npm ci && cd ../..

# optional MCP server deps
pip install -e ".[mcp]"
# optional live RICO download
pip install -e ".[rico]"

Quick start (train / disjoint test)

Every pipeline phase is also reachable through the unified slm CLI (slm list shows the full command map; slm guide <phase> prints the matching operating reference from .agents/skills/autotrain/references/). The python -m scripts.<name> forms below remain the direct equivalents.

# High-quality versioned corpus (default: all sources + quality synthesizer)
python -m scripts.build_train_data --source all --version v1 --synthesizer quality

# Fast fixture-only rebuild
python -m scripts.build_train_data --source fixture --version v0 --synthesizer quality

# Test suites with strict leakage checks against the train manifest
python -m scripts.build_test_data --source both --version v1 \
  --train-manifest outputs/data/train/v1/manifest.json

# Full HF-context trains sync checkpoints to the OpenUI bucket
# (https://huggingface.co/buckets/TKendrick/OpenUI). Requires HF_TOKEN.
export HF_TOKEN=hf_...   # or: hf auth login
python -m scripts.train_model \
  --train-dir outputs/data/train/v1 \
  --model twotower \
  --context-backend hf \
  --steps 200 \
  --run-id twotower_v1
# → hf://buckets/TKendrick/OpenUI/checkpoints/twotower_v1/

python -m scripts.evaluate_model \
  --test-dir outputs/data/eval/v1 \
  --model twotower \
  --run-id twotower_v1 \
  --ship-gates

Canonical lever registry

ModelBuildConfig is the single source for user-facing model, data, training, decode, and evaluation lever defaults. Discover the complete machine-readable set without searching scripts:

python -m slm_training.levers
python -m slm_training.levers --category decode

The catalog identifies intentional checkpoint-vs-harness default differences, each decode lever's executable tokenizer/compiler configurations, and the training objective that must exist in a checkpoint before a learned decode head can be enabled. Invalid, untrained, or inert combinations fail during config construction or checkpoint override, before run artifacts are made. The repository-wide run cap is the sole policy lever owned directly by src/slm_training/levers.py; changing MAX_RUN_MINUTES updates every Python consumer. Local compute is the default experiment path. Remote CI and managed jobs are optional last-resort execution surfaces and are not part of this local lever registry.

Evaluation uses the AgentEvals JSONL/YAML contract and the pinned AgentV SDK. Run npm ci before Python eval commands; shared model, loss, task, and diagnostic eval paths automatically write AgentV bundles beside their domain JSON under <run-dir>/agentv/. The existing honest OpenUI ship gates remain authoritative. See the AgentV evaluation contract.

Local-only / CI scratch: add --no-sync-checkpoints (matrix scripts default to scratch and stay local). Manual sync: python -m scripts.sync_checkpoints --run-dir outputs/runs/<id> --ensure-bucket. See docs/design/checkpoint-bucket.md.

Checkpoint provenance is fail-closed: each sync emits a verified CheckpointReferenceV1, and frontier/ship_candidate citations must resolve from a fresh clone or CI fails (python -m scripts.verify_checkpoint_references --check). See docs/design/checkpoint-provenance.md.

Honest ship path (V4 inventory-in-prompt / V6 stacked champion):

python -m scripts.run_quality_matrix --matrix v4 --only E35,E36 \
  --steps 40 --device cpu --context-backend scratch --no-design-md-context \
  --scratch-control

# V6: CoRe remask + slot-aware trust + honest V5 alphabet
python -m scripts.run_quality_matrix --matrix v6 --only E53 \
  --steps 80 --device cpu --context-backend scratch --no-design-md-context \
  --scratch-control

Train artifacts land in outputs/data/train/<version>/; eval, preference, annotation, trajectory, ProgramSpec, and mixture data use sibling typed roots. Use slm-data list, slm-data resolve train <version>, and slm-data verify train <version> instead of memorizing paths. Selected immutable snapshots publish to Git with slm-data publish train <version>.

Every new run writes outputs/runs/<id>/trace.json and OTLP JSONL signals under outputs/traces/<trace-id>/. Set OTEL_EXPORTER_OTLP_ENDPOINT for an optional remote OTLP mirror; detailed domain traces remain local and linked by trace ID. Set LANGSMITH_TRACING=true and provide LANGSMITH_API_KEY to export an additional, best-effort aggregate trace to the slm-training LangSmith project (LANGSMITH_PROJECT, LANGSMITH_ENDPOINT, and LANGSMITH_WORKSPACE_ID are optional overrides). This exports run/suite metrics, version stamps, gate verdicts, and AgentV summaries only—never prompts, targets, completions, checkpoints, raw logs, or credentials. Local evidence and AgentEvals remain authoritative when LangSmith is unavailable. Install the harness-side client with pip install -e '.[observability]'; the shared trace boundary loads the repository's ignored .env without overriding environment variables provided by CI or a shell. Standard OTLP settings are honored as well: OTEL_SERVICE_NAME, OTEL_RESOURCE_ATTRIBUTES, signal-specific endpoints, and OTEL_EXPORTER_OTLP_TIMEOUT (milliseconds; bounded to five seconds).

The flush pipeline remains: curated seeds + RICO + Awwwards → deterministic quality synth → per-record DESIGN.md + OpenUI validate → quality gates → stable sort by id + content fingerprint.

Eval uses meaningful parse (rejects empty stacks, missing placeholders, and low gold component-type recall), strict placeholder_fidelity for ship gates, structural_similarity, and composite reward_score (does not credit gold DESIGN.md lint). Suites: smoke/held_out (fixtures), rico_held, adversarial, ood. Soft placeholder_validity is diagnostic only.

Fixture demo vs ship: a tiny upsample + scratch + smoke-only fail-under is wiring only. Readiness requires --ship-gates on the full scoreboard (see adversarial review).

Expand rico_held with 1500 additional HF RICO screens (cached under src/slm_training/resources/rico/hf_test_cache.jsonl):

python -m scripts.build_test_data \
  --source both --version v1 \
  --train-manifest outputs/data/train/v1/manifest.json \
  --rico-hf-split test --rico-limit 2600 --target-records 1500
# Lightweight unit/integration suite (iterative model training is excluded)
pytest

# Only suites affected by staged + unstaged local changes
.githooks/check-changed

# Repository layout, skill mirrors, and tracked-artifact policy
python -m scripts.repo_policy

# Explicit, compute-intensive model-training tests
pytest -m training

Enable the tracked pre-commit hook once per clone with git config core.hooksPath .githooks (a Claude Code SessionStart hook arms it when it is unset). That hook is what runs the changed-file checker.

Agent hooks are narrower and are certified identical across harnesses by python -m scripts.verify_agent_surfaces:

Harness Config Blocks raw mv Post-edit parity, version, and case checks
Claude Code .claude/settings.json yes yes
Codex .codex/hooks.json yes yes
Copilot CLI .github/hooks/ yes yes
Cursor, Gemini CLI — (no hook mechanism configured) no no

Agents on a harness without hooks run python -m scripts.repo_policy and .githooks/check-changed themselves. CI remains authoritative either way. See docs/repository-organization.md and docs/design/agent-harness-parity-audit.md.

OpenUI Lang

Fixtures and validation use official openuiLibrary syntax, e.g.:

root = Stack([hero], "column")
hero_title = TextContent(":hero.title")
hero_body = TextContent(":hero.body")
hero = Card([hero_title, hero_body])

Content props must be placeholder strings. Parsing/serialization/prompt generation come from @openuidev/lang-core + @openuidev/react-ui — see src/apps/openui_bridge/.

DESIGN.md conditioning + linter: src/apps/design_md_bridge/ and src/slm_training/resources/design_md/.

Mission Control dashboard

serve_playground serves a control-plane + observability SPA at / — one pane of glass over the whole lifecycle (data → experiments → smoke → checkpoints/promotion) — including the annotate playground at /playground.

pip install -e ".[dev,torch,web]"
python -m scripts.serve_playground --port 8765        # full control plane (local)
python -m scripts.serve_playground --no-enable-jobs   # read-only observability
# For network exposure, set SLM_ANNOTATION_TOKEN and add --public.
# open http://127.0.0.1:8765

Surfaces (React 19 + Vite SPA, dark-first "mission control" design system):

Route What
/ Overview Live jobs, experiment scoreboard, checkpoint roster, corpus health, system status, remote dispatches
/data Navigate + generate versioned corpora (build_train_data / build_test_data)
/experiments Quality / grammar / perf / phase matrices; run run_*_matrix; dispatch bounded GPU checkpoint smokes (hf_jobs_train / remote_train); drill into any run
/smoke Smoke canary + perf & telemetry; launch wiring runs
/checkpoints Roster + live configurable ship gates + promote / deploy + blinded A/B
/runs/<id> Per-run detail — gate matrix, telemetry spans, train_summary metrics, durable-checkpoint link
/playground Full annotate UI (React): staged generation, browser fallback/review, DSL repair, and feedback

Read vs execute. Observability views are pure reads (work on a fresh checkout and on read-only Vercel, falling back to committed docs/design/*.json / MODEL_CARD.md / src/slm_training/resources/, tagged with provenance). Generate/run/promote actions execute an allowlisted set of scripts as tracked background jobs with live SSE logs — only when served locally (--enable-jobs, default on); Vercel degrades to read-only automatically. Gate math (POST /api/gates/evaluate) is pure, so the threshold editor stays live even read-only. Backend: src/slm_training/web/{observability,jobs,capabilities,routes}.py; SPA source in src/apps/dashboard/ (built bundle committed under web/static/app/, like the preview lib).

Compiled ↔ interpreted (dogfooding OpenUI). The sidebar has a ◈ Compiled / ◇ Interpreted toggle. Compiled is the hand-written React above. Interpreted renders each page from a committed OpenUI Lang program (src/slm_training/web/static/openui/<slug>.openui) run live through the official @openuidev <Renderer> — same components, live /api data via a tool provider, working nav, reactive selectors, launchers, and the live gate editor — so the app is the DSL. The two are kept at parity (scripts/validate_page_dsl.py + tests/test_web/test_page_dsl.py + the dashboard-openui-parity skill); interpreted-mode source lives in src/apps/dashboard/src/interpret/.

Annotate playground (/playground)

python -m scripts.serve_playground --port 8765
# open http://127.0.0.1:8765/playground

/playground is the React annotate UI inside the SPA shell (shares the dark design system). It owns the complete annotation flow: bounded server attempts, browser review/fallback, editable and validated DSL corrections, annotator/model identity, bearer-token support, activity history, keyboard/swipe grading, and the diffusion progress canvas. The retired /playground/classic URL redirects here. If both model paths are unavailable, the page shows a clearly labeled wiring fallback so the renderer/editor/annotation flow remains testable; uncorrected fallback feedback is excluded from derived training data.

The demo checkpoint lives in src/slm_training/resources/checkpoints/playground_demo/ (committed last.pt + tokenizer + meta). To regenerate it:

python -m scripts.bootstrap_playground --force

If last.pt is missing after a sparse checkout, run the bootstrap command above before starting the playground. Annotate mode (default UI): auto-generated prompts, prefetch 1–2 samples ahead, and a live OpenUI visual preview (same @openuidev/react-lang Renderer path as openui.com/demo).

Input Action
Thumbs up (persist, stay on sample)
Thumbs down (persist, stay on sample)
/ Previous / next sample
typing Focus optional note
swipe Mobile: horizontal navigate, vertical grade

Annotations append to outputs/data/annotation/feedback.jsonl. Invalid model outputs are quarantined to outputs/data/annotation/bad_outputs.jsonl (never shown in the app). Thumbs-up rows promote into src/slm_training/resources/annotations/human_train.jsonl (merged by build_train_data). Opposite ratings on the same prompt also write outputs/data/preference/human_pairs.jsonl.

python -m scripts.export_annotations status
python -m scripts.export_annotations export

Rebuild the OpenUI preview bundle

npm run preview:install
npm run preview:build
# writes src/slm_training/web/static/preview/{preview.js,preview.css}

Rebuild the dashboard bundle

npm run dashboard:install
npm run dashboard:build
# writes src/slm_training/web/static/app/ (built SPA, committed like the preview lib)

Playwright visual / e2e

npm ci
npx playwright install chromium
# optional agent skills (already in .agents/skills + discovery mirrors)
playwright-cli install --skills
npm run test:e2e

MCP (Cursor): .cursor/mcp.json launches @playwright/mcp.

  • Context tower: scratch TokenEncoder or frozen HF model (--context-backend hf, default HuggingFaceTB/SmolLM2-135M)
  • Denoiser tower: MaskGIT-style masked token prediction with cross-attention to context (Chang et al. 2022; adapted)
  • Grammar decode: DFA force-emit + MaskGIT hole-admit + LTR certify so constrained samples stay valid OpenUI (research lineage). Constrained decoding is the product, not a switch — see decode invariants; --unconstrained-control (formerly --no-grammar) is a diagnostic control arm whose output is never certified or shipped
  • Output tokenizer: dual-mode — default compositional OpenUITokenizer, or V5 lexer / DSL-native DSLNativeTokenizer (output_tokenizer=lexer; see dsl-native-tokenizer.md)
  • Eval: syntax parse_rate, separate meaningful_program_rate, placeholder fidelity, and canonical tree match — no hidden gold channel at generate time
# Optional HF context (requires: pip install -e ".[hf]")
python -m scripts.train_model --model twotower --context-backend hf \
  --hf-model HuggingFaceTB/SmolLM2-135M --steps 200 --run-id twotower_hf --fast-train

Hugging Face Jobs (full GPU train)

ZeroGPU Spaces are for short demos only. Full trains use managed Jobs:

python -m scripts.hf_jobs_train --dry-run --run-id twotower_jobs_v1 --steps 200
# submit: export HF_TOKEN=… && python -m scripts.hf_jobs_train --run-id … --steps 200

Details: docs/design/hf-jobs-train.md.

GPU multi-farm MCP

cp .env.example .env
pip install -e ".[mcp]"
GPU_MULTI_FARM_MODE=mock python -m scripts.multi_farm_mcp

Agent instructions

All coding agents (Cursor, Claude Code, Codex, Gemini, Copilot / GHCP, …) must follow AGENTS.md. Canonical skills live in .agents/skills/ (mirrored under .claude/skills/, .cursor/skills/, and .grok/skills/).

Iron law: after any train / eval / bench / profile / telemetry / matrix / reproduction (or decision-informing ad-hoc) run, update docs/design/ JSON and the matching measured-results markdown. Full trigger list and recipe checklist: AGENTS.md (skill: documenting-experiment-results). Do not leave results only under outputs/.

All eval entrypoints also publish standard AgentEvals cases and AgentV SDK artifacts. Do not add evaluator-specific envelope formats; extend src/slm_training/evals/agentv.py.

Token-efficiency stack

Repo ships ponytail, caveman, headroom, and rtk under .agents/skills/ (plus RTK.md, Cursor rules, and GHCP .github/copilot-instructions.md). Details and refresh commands: AGENTS.md — Token-efficiency stack.

# RTK binary (once per machine) — must pass `rtk gain`
curl -fsSL https://raw.githubusercontent.com/rtk-ai/rtk/refs/heads/master/install.sh | sh

OpenWiki (code mode)

Repository wiki for agents lives under docs/openwiki/ (start at docs/openwiki/quickstart.md). Setup uses langchain-ai/openwiki code mode: AGENTS.md / CLAUDE.md OpenWiki snippets and .github/workflows/openwiki-update.yml.

npm install -g openwiki@0.1.2
# needs OPENAI_API_KEY (preferred) or OPENROUTER_API_KEY
python -m scripts.update_openwiki --update --print

Add repo secret OPENAI_API_KEY to enable scheduled OpenWiki update PRs. The workflow falls back to OPENROUTER_API_KEY when OpenAI is unavailable and fails clearly when neither secret exists. LANGSMITH_API_KEY enables optional tracing.

Hugging Face CLI + skills

Agents use the official hf CLI and the huggingface/skills pack (skill: hf-cli plus datasets / papers / trainers / Spaces / … under .agents/skills/). Cursor also gets the Hugging Face MCP server via .cursor/mcp.json.

curl -LsSf https://hf.co/cli/install.sh | bash
hf skills add --force
hf skills update
hf skills add --claude --force
hf skills add --dest=.cursor/skills --force

Optional Cursor UI: marketplace — Hugging Face. CLI docs: huggingface_hub CLI. Tokens: settings/tokens.

Serena MCP

Semantic code tools via Serena (not marketplace installs). Project is initialised under .serena/; Cursor / Claude / VS Code MCP configs are wired in-repo. See AGENTS.md — Serena MCP.

uv tool install -p 3.13 serena-agent
serena init
serena project health-check

Layout

AGENTS.md              # cross-tool agent instructions (required reading)
RTK.md                 # Rust Token Killer usage (shell output compression)
docs/MODEL_CARD.md     # checkpoint roster + eval (README holds a summary)
docs/repository-organization.md # tracked-file placement + move policy
.agents/skills/        # canonical agent skills
src/slm_training/
  dsl/                 # OpenUI adapter + design_md + grammar/{backends,fastpath}
  harnesses/           # train_data, test_data, model_build, rl, preference,
                       # distill, quality(+retrieval), experiments, annotations
  models/              # TwoTower, grammar_diffusion, tokenizers, remask
  data/                # RICO / Awwwards adapters + leakage fingerprints
  evals/               # loss suites / denoising NLL
  runtime/             # accel, telemetry, compression, cactus
  web/                 # mission-control API (observability + jobs) + annotate playground + SPA
src/gpu_multi_farm/    # FastMCP server + farm adapters
src/apps/openui_bridge/   # @openuidev/lang-core Node sidecar
src/apps/design_md_bridge/
src/apps/openui_preview/
scripts/               # CLIs
src/slm_training/resources/              # seed pairs + RICO semantic slices
docs/design/           # architecture + research lineage + contracts
tests/
  test_dsl/            # parser, grammar, design_md
  test_harnesses/      # mirrors harnesses/* (rl is its own suite)
  test_runtime/        # accel / cactus / compression
  test_models/ test_data/ test_web/ ...

Releases

Packages

Used by

Contributors

Languages