Hyperdistillation¶

Trace synthesis component — reasoning-to-code learning pipeline.

Cluster: ML & Optimisation | Type: component | MCP Tools: 29

Overview¶

Reasoning-to-code learning pipeline that captures LLM reasoning traces from JSONL event logs, synthesizes deterministic Python artifacts, verifies them, and maintains a searchable artifact library with Bayesian confidence tracking. Six core operations (capture, check_library, execute_artifact, list_artifacts, get_artifact, delete_artifact) cover the learn-store-reuse lifecycle that closes the G6 self-improvement loop.

When to use:

Converting successful solver or agent traces into reusable, verified Python functions
Checking whether a known-good artifact already solves a new problem before invoking the LLM
Running the learn loop that continuously grows G6's library of distilled deterministic solutions

Production behavior:

Artifact execution is statically checked before runtime
Runtime execution occurs in an isolated Python subprocess with a timeout
Imports, dunder traversal, dynamic evaluation, file access, and network access are rejected
Usage outcomes update artifact confidence through Bayesian tracking

Example:

from mvp.hyperdistillation import HyperdistillationBlock, DistillInput

block = HyperdistillationBlock(name="distill")
result = block.infer(DistillInput(
    op="capture",
    event_log=[{
        "event_type": "step_complete",
        "step_number": 8,
        "llm_model": "gpt-4o-mini",
        "tokens_used": 12,
        "latency_ms": 80,
        "goal": "Calculate 2+2",
        "data": {
            "status": "completed",
            "input_summary": "2+2",
            "output_summary": "4",
            "reasoning_trace": "add two and two",
        },
    }],
))

Works well with: evoskill, cegis, solver

Production Caveats¶

Artifact execution boundary

Hyperdistillation rejects unsafe Python syntax and executes artifacts in an isolated Python subprocess with a timeout. This prevents common artifact failures such as imports, file access, dynamic evaluation, dunder traversal, and runaway loops. It is still not a VM, container, or OS-level security boundary. Do not use it to run arbitrary untrusted code in a multi-tenant deployment without an outer container or VM sandbox.

Cache validity

A distilled artifact is evidence that a previous trace pattern worked for similar inputs. It is not proof that the artifact is correct for every future input, and it is not a substitute for domain tests, holdout evaluation, or human review on high-stakes workflows. Distribution shift should trigger fallback to fresh reasoning and re-validation.

Latency and feature flags

The default path prefers local deterministic synthesis before CEGIS. CEGIS and advanced template synthesis can be slower and should be monitored through the synthesis coverage ledger. Keep G6_ADVANCED_TEMPLATE_SYNTHESIS unset in production unless you have benchmark evidence for the target workload. For launch users, advanced template synthesis should remain an internal reliability primitive rather than a visible onboarding feature; only enable it when you can compare template_a/template_d outcomes against the default path and inspect failures.

Template synthesis verification boundary

Advanced template synthesis verifies fitted templates against supplied examples only. It requires at least three I/O samples and does not prove correctness for unseen inputs, distribution shifts, or high-stakes workflows. Treat any generated artifact as a candidate that still needs domain tests, holdout cases, and human review where consequences are material.

Data handling

Event logs, test cases, and stored artifacts can include summaries of user inputs and outputs. Treat the distillation database as sensitive application data: scope it per user/workspace, back it up intentionally, and avoid feeding secrets or regulated data into traces unless the deployment has the required data controls.

Public API¶