QuantumComputinglib learn · inspect · formalize
Checked on this commit 4,524 public declarations commit c681192368c2 Build record

ASPBE harness

Search, prove, validate, and retain evidence

Both generations keep a frozen mathematical contract and hard verification. The previous Harness routed work through fixed strategist / Lean-tree / worker / reviewer layers; the current Harness keeps the proof DAG but lets generalist Workers own substantive frontier advances end to end under a Frontier Master.

Previous → current

Same proof discipline, less cognitive role confinement

The old hierarchy made handoffs explicit. The current system keeps the hard gates and durable theorem graph, but parallel Workers may cross source, math, Lean, diagnostics, and exposition whenever that moves the root theorem.

Previous Harness

Layered roles

Previous ASPBE hierarchical Harness comic

Upper strategist → Middle Lean-tree manager → focused workers → reviewer.

Current Harness

Proof-frontier search

Current ASPBE Frontier Master and generalist Workers comic

Frontier Master → parallel generalist Workers → hard proof gates → certified outputs.

Candidate generation, scoring, verification, and acceptance editable Mermaid source
flowchart LR
  subgraph OLD["Previous layered Harness"]
    direction TB
    O0["Frozen task contract"]
    O1["Upper strategist<br/>choose objective · allocate budget"]
    O2["Middle Lean-tree manager<br/>decompose proof DAG · route feedback"]
    O3["Lower focused workers ×N<br/>solve bounded source / math / Lean leaves"]
    O4["Reviewer / verifier<br/>source · semantics · Lean · resources"]
    O5["Accepted theorem or typed failure"]
    O0 --> O1 --> O2 --> O3 --> O4 --> O5
    O4 -- "diagnostic feedback" --> O2
  end

  subgraph NEW["Current Frontier Master + Generalist Workers"]
    direction TB
    N0["Frozen State Preparation / Block Encoding contract"]
    N1["Frontier Master<br/>maintain global theorem graph · select real bottlenecks"]
    N2["Generalist Workers ×N<br/>own substantive objectives end to end<br/>source ↔ math ↔ Lean ↔ diagnostics ↔ exposition"]
    N3["Local integration<br/>compare evidence · preserve cross-layer ideas"]
    N4["Hard acceptance gates<br/>source fidelity · circuit semantics · Lean · integration · readable proof"]
    N5["Certified frontier advance<br/>or sharp typed obstruction"]
    N6["Researcher-facing outputs<br/>LaTeX proof · Lean · circuit · Qiskit/OpenQASM"]
    N0 --> N1 --> N2 --> N3 --> N4 --> N5 --> N6
    N5 -- "update theorem graph" --> N1
  end

  O4 -. "hard verification retained" .-> N4
  O2 -. "proof-DAG memory retained" .-> N1
  O3 -. "fixed role walls removed" .-> N2

  classDef old fill:#fff3d9,stroke:#b78638,color:#4b3517,stroke-width:1.5px;
  classDef oldGate fill:#ffe6d8,stroke:#b85d3b,color:#5a2c1c,stroke-width:1.6px;
  classDef master fill:#e8f3ff,stroke:#4778a6,color:#19364e,stroke-width:1.7px;
  classDef worker fill:#f0eaff,stroke:#7960a5,color:#30244a,stroke-width:1.5px;
  classDef gate fill:#e3f6ea,stroke:#3b805a,color:#173b29,stroke-width:1.7px;
  class O0,O1,O2,O3,O5 old;
  class O4 oldGate;
  class N0,N1,N3,N6 master;
  class N2 worker;
  class N4,N5 gate;

1. Formal target

Translate the mathematical request into a state-preparation or block-projection contract.

2. Memory retrieval

Retrieve compatible completed declarations, route cards, and known obstructions.

3. Candidate population

Generate distinct constructions and retain provenance, assumptions, and resource tuples.

4. Resource ranking

Apply the declared lexicographic score without promoting an invalid candidate.

5. Lean obligations

Split unitarity, dimensions, register order, projection, normalization, and approximation goals.

6. Adaptive exploration

The upper layer may add agents, recombine routes, or relax epsilon only from recorded feedback.

7. Acceptance gate

Require the Lean build, tests, no-new-proof-hole policy, and source-linked documentation.

8. Executable validation

Run applicable circuit or matrix exports, including Qiskit checks, after the formal certificate gate.

Evidence boundary. Harness declarations formalize controller data and acceptance policy. They do not prove that an external agent run was efficient. Run logs, token accounting, and Qiskit outputs remain engineering evidence attached after the Lean gate.