Hyphal Context Benchmark for Receipt Routing

Status

This page is website copy for a working paper draft. The benchmark result is HYPHAL_CONTEXT_FIXTURE_MATCH for one deterministic fixture only. It compares a full-context route against a gradient-envelope-and-receipt route over the frozen biology/network-intelligence corpus. Model answer quality is NOT_MEASURED. General route superiority is UNVERIFIABLE. A native BuildLang/buildc receipt is NOT_REPLAYED.

Claim ledger

Benchmark fixture

HYPHAL_CONTEXT_FIXTURE_MATCH. The hyphal route recovered the same required evidence classes and guardrails as the full-context route with fewer estimated prompt tokens.

Full-context route

MATCH. Sends all ten source bodies plus the source gate and seed note. Estimated prompt tokens: 123413.

Hyphal route

MATCH. Sends ten gradient envelopes and rehydrates six evidence cards by receipt. Estimated prompt tokens: 1338.

Guardrails

MATCH. Both routes block biological nervous-system equivalence, universal intentional common-mycorrhizal-network messaging, and overpromoting the protocol beyond the fixture.

Tooling receipts

MATCH / VERIFIED. Crucible assessed the bounded thesis as MATCH 3 / DRIFT 0 / UNVERIFIABLE 0. Learn reverified the prooflesson receipt as VERIFIED.

Scope

UNVERIFIABLE outside fixture. The result does not prove that hyphal routing wins generally or that a model answer improves.

Measured fixture

Token savings

122075 estimated prompt tokens.

Token savings ratio

0.9892 of the full-context estimate.

Evidence recall delta

0 across the six required evidence classes.

Guardrail delta

0 across the three bounded guardrails.

Required evidence

• fungal_signal
• plant_signal
• network_evidence
• overclaim_boundary
• source_availability
• architecture_seed

Benchmark relation

hyphal.required_class_recall == full.required_class_recall
hyphal.guardrail_recall == full.guardrail_recall
hyphal.estimated_prompt_tokens < full.estimated_prompt_tokens

The current receipt verifies this relation in JavaScript over one frozen corpus. The next version should replay the relation in BuildLang/buildc and add a model answer-quality phase.

Receipts

• Benchmark CLI: demo/hyphal-context-benchmark.mjs.
• Benchmark receipt: hyphal-context-benchmark-2026-07-02.json.
• Crucible run: twenty-second-wave-hyphal-context-run-2026-07-02.json.
• Crucible report: twenty-second-wave-hyphal-context-report-2026-07-02.md.
• Learn prooflesson: twenty-second-wave-hyphal-context-benchmark.prooflesson.json.
• Official copy: HYPHAL-CONTEXT-BENCHMARK-FOR-RECEIPT-ROUTING-2026-07-02.md.

Do not infer

• Do not infer that this fixture is a general context-routing theorem.
• Do not infer that estimated prompt tokens equal provider billing tokens.
• Do not infer that model answer quality has been measured here.
• Do not infer that the benchmark changes the biology source boundaries.
• Do not infer that this relation already has a native BuildLang/buildc receipt.