Colab notebooks
11
Hugging Face token
set
Kaggle credentials
missing
Zero-cost preflight
not run/blocked
Bijective gate
25/25
Operator bus
ready
Operator Agent Bus
The new bus is the practical coordination lane for bounded agent work: task packet in, local/private routing first, verification receipt out. It is not a public autonomy promise; it is the dispatch surface that keeps Codex, Claude, Cursor, Hugging Face, Colab, and local checks from stepping on each other.
| Bus Layer | What It Does | Buyer/Product Value |
|---|---|---|
operator_agent_bus | Routes narrow work packets into training, review, docs, release, and support lanes. | Turns the repo from experiments into repeatable workflows with receipts. |
| Constrained decoding gate | Forces closed-vocabulary helper calls through the canonical schema before evaluation. | Reduces support risk because the model must emit usable tool calls, not vague prose. |
| Verification receipts | Records the command, artifact, gate result, and next action for each run. | Makes delivery and troubleshooting inspectable after a customer buys. |
Latest Training Proof
Local constrained bijective gate rerun: 25/25 passing across AV, RU, CA, UM, and DR on CPU with no new GPU spend. Receipt: artifacts/bijective_tongue/local_constrained_1778133356.json. Use this as a production-hardening proof for the shim, not as a general raw-model capability claim.
Daily Stack
| Step | Command | Cost |
|---|---|---|
| refresh consolidated local buckets | python scripts/system/consolidate_ai_training.py | local |
| verify zero-cost profile inputs | npm run training:preflight:zero-cost | local |
| open the safest remote compute lane | npm run training:surfaces -- --surface colab | Colab or existing subscription |
| review current candidates before new GPU time | npm run training:eval-matrix && npm run training:run-ledger | local |
| run reversible-buffer leakage tests | npm run experiment:info-leakage-buffer | local |
| build Jupiter-ring feedback rows | npm run training:jupiter-ring | local |
| publish this operator view | npm run training:hub | local/Vercel static hosting |
Highest-Value Run Queue
| Run | Triage | Platform | Lane | Next Step |
|---|---|---|---|---|
polly-coding-approval-metrics-v2/checkpoint-100 | EVALUATE_NEXT | kaggle | coding_approval | Run missing gates: frozen_perplexity, functional_benchmark. |
polly-coding-approval-metrics-v2/checkpoint-60 | EVALUATE_NEXT | kaggle | coding_approval | Run missing gates: frozen_perplexity, functional_benchmark. |
polly-coding-approval-metrics-v2/checkpoint-90 | EVALUATE_NEXT | kaggle | coding_approval | Run missing gates: frozen_perplexity, functional_benchmark. |
brick3_instruct_live_20260424/checkpoint-30 | EVALUATE_NEXT | local | general_training | Run frozen perplexity plus functional coding-agent gate; do not merge until executable behavior passes. |
brick3_instruct_live_20260424/checkpoint-60 | EVALUATE_NEXT | local | general_training | Run frozen perplexity plus functional coding-agent gate; do not merge until executable behavior passes. |
brick3_local/checkpoint-400 | EVALUATE_NEXT | local | general_training | Run frozen perplexity plus functional coding-agent gate; do not merge until executable behavior passes. |
brick3_local/checkpoint-450 | EVALUATE_NEXT | local | general_training | Run frozen perplexity plus functional coding-agent gate; do not merge until executable behavior passes. |
brick3_local/checkpoint-500 | EVALUATE_NEXT | local | general_training | Run frozen perplexity plus functional coding-agent gate; do not merge until executable behavior passes. |
20260426T013255Z | EVALUATE_OR_RERUN_NEXT | huggingface-job-packet | aligned_foundations | Confirm remote artifact exists, pull/push adapter, then run the standard eval matrix. |
20260425T010044Z | EVALUATE_OR_RERUN_NEXT | huggingface-job-packet | binary_hex | Confirm remote artifact exists, pull/push adapter, then run the standard eval matrix. |
20260426T091757Z | EVALUATE_OR_RERUN_NEXT | huggingface-job-packet | dsl_synthesis | Confirm remote artifact exists, pull/push adapter, then run the standard eval matrix. |
20260426T091811Z | EVALUATE_OR_RERUN_NEXT | huggingface-job-packet | dsl_synthesis | Confirm remote artifact exists, pull/push adapter, then run the standard eval matrix. |
Remote Compute Lanes
| Notebook | Category | Use |
|---|---|---|
| scbe-finetune-free | training | Free T4 LoRA fine-tuning lane for SCBE governance data. |
| qlora-training | training | QLoRA training notebook for compact Colab model adaptation. |
| coder-code-primaries | training | QLoRA notebook for the SCBE coding-primaries lane across Python, JavaScript, Rust, Mathematica, Haskell, and Markdown. |
Evidence Receipts
These paths are local repo receipts, not public download promises. Buyer delivery should go through the package URLs and manuals above.
| Path | Status | Use |
|---|---|---|
artifacts/bijective_tongue/local_constrained_1778133356.json | present | 25/25 constrained gate receipt |
artifacts/training_run_ledger/latest/ledger.md | present | training run ledger |
artifacts/training_evaluation_matrix/latest.md | present | evaluation matrix |
products/packaged/SCBE_AI_Governance_Toolkit_v1.zip | ignored local artifact | paid Toolkit packet |
products/packaged/SCBE_AI_Security_Training_Vault_v1.zip | ignored local artifact | paid Training Vault packet |