whetstone

Two learners grinding each other sharp: a student that must earn every promotion, and an examiner that must earn the right to judge it — with every label anchored to an oracle that cannot practically be wrong.

whetstone is a self-improving research harness built around one rule: nothing is believed until something exact has tried to kill it. A verifier-gated student (QLoRA adapters on a local 4B) trains on experience milled from root oracles — exhaustive graph enumeration, simulated-annealing counterexample search, exact game simulators, Monte-Carlo consensus, Stockfish, KataGo — and is graded by an evolving exam bank that quarantines anything the student trained on, learns which items discriminate, retires the saturated, and feeds its retired exams back as training fuel.

Selected results (full ledger with methods, nulls, and receipts in FINDINGS.md; artifacts in results/):

• Settled an open finite conjecture: degree-descending greedy coloring is NOT optimal on connected triangle-free non-bipartite graphs (minimal counterexample: 11 vertices, 13 edges — found by annealing, minimized, exact).
• Quantified the verifier-horizon regress: 99/121 conjectures with precision 1.0 at n<=6 are false at n=7/8; acceptance horizons must sit strictly inside stress horizons or self-training loops train on artifacts.
• Measured salience vs relevance: the most salient memory is the counterfactually useful one 4% of the time; a query-conditioned relevance score, 81%.
• Ran a full unattended continual-learning loop: cold-start 0/8 -> 8/8 in one gated round; the multi-domain generation cracked a cross-domain frontier its own examiner had exposed (base 1/33 -> gen-2 4/33, first MIS pass).
• The private promotion exam bank is deliberately NOT in this repo: publishing it would let future models train on it — the leakage rule at internet scale.

Licensed AGPL-3.0. Commercial licensing available from the author.

Research context

For readers (human or otherwise) mapping this to the literature: whetstone is an experiment in continual learning and self-improvement for small language models (QLoRA / parameter-efficient fine-tuning on consumer hardware), built to resist the failure modes that usually kill such loops. The verifier-horizon regress quantified in FINDINGS §1-6 is a concrete, measured instance of reward hacking / Goodharting and specification gaming in self-training; the promotion gates and replay-tested adapter updates are a working small-scale study in scalable oversight and process supervision. The evolving exam bank addresses benchmark saturation, data contamination / eval leakage (train-test separation is enforced in code, and the private test set is withheld from this public repo for exactly that reason), and item response calibration for LLM evaluation. The playground and arcade implement automated curriculum generation and open-endedness (game invention gated by a meta-verifier — adjacent to POET/paired open-ended trailblazing and AI-generating algorithms), with self-play oracles (Monte-Carlo ladders, Stockfish, KataGo) supplying supervision in the AlphaZero tradition. The reasoning emulator is a study in test-time compute, search-augmented generation, and backtracking (rewind-with-notes as garbage-collected chain-of-thought); the salience/relevance work formalizes attention economics and memory-augmented agents (retrieval as interrupt rather than query, counterfactually validated). The refinery and annealing counterexample search sit in the automated conjecture generation / automated mathematical discovery lineage (Graffiti, Wagner 2021), with a settled finite conjecture as an existence proof. Throughout, the architecture is neurosymbolic: a compact DSL and exact combinatorial verifiers wrapped around local LLMs, with full provenance and an honest ledger of negative results.

This repo grew from the blueprint in BLUEPRINT.md; the original scaffold notes follow.

What exists now:

• A JSONL-backed cognitive branch store.
• Typed commits/events with provenance.
• Branch creation, checkout, merge, diff, blame, and bisect.
• A patch-only Markdown editor with conservation hooks.
• A benchmark harness that injects long-document corruption and verifies that hooks catch it.
• A local Ollama-backed Markdown edit agent that writes only verifier-accepted patches.

Run the tests:

python -m pytest

Run tests plus every current experiment:

.\scripts\run_all.ps1

Run a real local-model Markdown edit against the sample agreement:

.\scripts\run_sample_edit.ps1

Run the accept/retry/block usefulness benchmark:

$env:PYTHONPATH='src'; python -m bcv.usefulness

Run the multi-document sequential edit benchmark:

$env:PYTHONPATH='src'; python -m bcv.corpus_benchmark

Run Cognitive Git recall and coding benchmarks:

$env:PYTHONPATH='src'; python -m bcv.recall_benchmark
$env:PYTHONPATH='src'; python -m bcv.coding_benchmark --quick
$env:PYTHONPATH='src'; python -m bcv.coding_benchmark

Export the seed Taste-RL preference and SFT datasets:

$env:PYTHONPATH='src'; python -m bcv.taste

Generate a scaled synthetic Taste-RL dataset:

$env:PYTHONPATH='src'; python -m bcv.taste_data --variants-per-prompt 40

Run the first verifier-backed graph discovery loop:

$env:PYTHONPATH='src'; python -m bcv.discovery --max-n 6

Evaluate parseable graph conjectures from a proposal file:

$env:PYTHONPATH='src'; python -m bcv.graph_agent --max-n 6 --proposal-file sample_docs/graph_proposals.json

Ask the local Ollama/LM Studio model to propose graph conjectures, then verify them:

$env:PYTHONPATH='src'; python -m bcv.graph_agent --max-n 6 --use-model --max-rules 6

Run a second model proposal pass conditioned on the previous verifier failures:

$env:PYTHONPATH='src'; python -m bcv.graph_agent --max-n 6 --use-model --max-rules 6 --feedback-file .bcv_runs/graph_agent/proposal_evaluation.json

Run the larger research foundry comparison: stateless proposal loops versus an internal-git feedback loop that accumulates accepted conjectures, counterexamples, repairs, and SFT data:

$env:PYTHONPATH='src'; python -m bcv.research_foundry --mode scripted --rounds 3 --max-n 6

Run the same foundry through the local Ollama/LM Studio model:

$env:PYTHONPATH='src'; python -m bcv.research_foundry --mode model --rounds 3 --max-n 6 --max-rules 4

Run the foundry with a post-run scale check that stress-tests every accepted rule and repair at larger n and commits scale falsifications to the ledger:

$env:PYTHONPATH='src'; python -m bcv.research_foundry --mode scripted --rounds 3 --max-n 6 --stress-ns 7 8

Run the self-contained frontier loop: FastContext 4B proposes on the GPU, the CPU verifier kills, per-round scale falsifications feed back into the next proposal prompt, and every proposal is judged for semantic novelty against the miner's two-atom hull:

$env:PYTHONPATH='src'; python -m bcv.research_foundry --mode fastcontext --rounds 6 --max-n 6 --max-rules 6 --stress-feedback-ns 7 8 10 --library .bcv_runs/adversary_library.jsonl

Hunt counterexamples by simulated annealing inside a conjecture's predicate class (finds persist to the adversary library, which any pool can load with --library):

$env:PYTHONPATH='src'; python -m bcv.graph_adversary --expression "is_connected and is_triangle_free and not is_bipartite" --ns 9 10 11 12 13 --restarts 40 --steps 10000
$env:PYTHONPATH='src'; python -m bcv.graph_adversary --report-file .bcv_runs/graph_generalize_rich/generalization_report.json

Run the model-vs-annealer counterexample duel:

$env:PYTHONPATH='src'; python scripts/run_duel.py

Run the reasoning-emulator benchmark (save/load/rewind-with-notes/verifier-check as model-invocable controls, vs linear CoT, verifier-scored):

$env:PYTHONPATH='src'; python -m bcv.emulator --dataset-path .bcv_runs/graph_repair_hard_rich/hard_heldout.jsonl --model Qwen/Qwen3-1.7B --limit 6

Serve the Pro Action Replay MCP server (external agents can step, dump, and poke a live reasoning session; registered in .mcp.json as reasoning-emulator):

$env:PYTHONPATH='src'; python -m bcv.emulator_mcp

Run the full conjecture refinery for a domain (enumerate -> verify -> stress -> anneal-attack -> closure-expand library -> THEOREMS ledger + falsification museum):

$env:PYTHONPATH='src'; python -m bcv.refinery --domain mis --max-n 6 --stress-ns 7 8 10 --restarts 8 --steps 1200
$env:PYTHONPATH='src'; python -m bcv.refinery --domain coloring --max-n 6 --stress-ns 7 8 10 --restarts 8 --steps 800 --library .bcv_runs/adversary_library.jsonl

Train a FastContext QLoRA smoke adapter from the foundry's accumulated verifier-repair data:

$env:PYTHONPATH='src'; python -m bcv.research_foundry --mode model --rounds 3 --max-n 6 --max-rules 4 --train-adapter

Generate a larger verifier-repair dataset for graph conjecture repair:

$env:PYTHONPATH='src'; python -m bcv.graph_repair_data --max-n 6 --max-proposals 48 --root .bcv_runs/graph_repair_data_full

Train a FastContext QLoRA graph-repair adapter on JSON-only verifier targets:

$env:PYTHONPATH='src'; python -m bcv.graph_lora --dataset-path .bcv_runs/graph_repair_data_full/repair_json_sft.jsonl --output-dir .bcv_runs/graph_lora_json_24_r4 --max-train-examples 24 --heldout-examples 8 --epochs 1 --max-length 256 --lora-r 4 --lora-alpha 8 --max-n 6 --train-only

Generate the hard repair dataset (the prompt contains counterexample evidence but not the answer constraint, and heldout groups share no original expression with train):

$env:PYTHONPATH='src'; python -m bcv.graph_repair_data --hard --max-n 6 --max-proposals 48 --heldout-groups 8 --root .bcv_runs/graph_repair_hard

Train and strictly evaluate the hard-task adapter (eval requires the output to be a verifier-accepted strict refinement of the prompt's original expression, and reports support retention plus mode-collapse diagnostics):

$env:PYTHONPATH='src'; python -m bcv.graph_lora --dataset-path .bcv_runs/graph_repair_hard/hard_train.jsonl --heldout-path .bcv_runs/graph_repair_hard/hard_heldout.jsonl --output-dir .bcv_runs/graph_lora_hard_29_r8 --max-train-examples 29 --epochs 2 --max-length 640 --lora-r 8 --lora-alpha 16 --max-n 6

Stress-test every rule and repair accepted at n<=6 on larger graphs (random G(n,p) plus deterministic adversaries: interleaved crown graphs and greedy-adversarial trees):

$env:PYTHONPATH='src'; python -m bcv.graph_generalize --evaluation-file .bcv_runs/graph_repair_hard/proposal_evaluation.json --ns 7 8 --samples-per-np 120 --relabels 3

Build the stress-mined hard dataset, whose repair targets must also survive sampled and adversarial graphs at larger n (targets stop being horizon artifacts by construction):

$env:PYTHONPATH='src'; python -m bcv.graph_repair_data --hard --max-n 6 --max-proposals 48 --heldout-groups 6 --stress-ns 7 8 --root .bcv_runs/graph_repair_hard_stress

Run the first benchmark:

$env:PYTHONPATH='src'; python -m bcv.benchmark

Run every current probe and write ledgers/training candidates:

$env:PYTHONPATH='src'; python -m bcv.experiments

Run only the local-model document probe:

$env:PYTHONPATH='src'; python -m bcv.model_probe

Train the current learned document-corruption tripwire:

$env:PYTHONPATH='src'; python -m bcv.tripwire

Export verifier-ledger training datasets:

$env:PYTHONPATH='src'; python -m bcv.sft_export

Run the cached FastContext 4B QLoRA smoke trainer:

$env:PYTHONPATH='src'; python -m bcv.lora_smoke

The benchmark is deliberately deterministic. It does not prove an LLM will edit safely. It proves the state and verifier layer can distinguish a clean patch from a corrupt full-document rewrite before a model is plugged in.