supabrain-x · context allocation research · trusted-local Product Alpha

The problem is not longer context.

The problem is deciding what deserves context.

Supabrain studies evidence allocation before answer generation: how systems select, rank, structure, and preserve the material an LLM should see.

01 · Problem

Memory is not storage alone. It is an allocation policy over time.

LLM context is finite, costly, and easy to waste. Larger windows help, but they do not decide what should be placed inside them.

Supabrain focuses on the layer before answering: did the right evidence reach the working context, with enough structure and provenance to inspect it?

Claim boundary: this is not a claim that memory is solved.

02 · Method

evidence allocation

Context packages before final answers

Supabrain separates context allocation from answer generation. The measured question is whether labelled evidence reaches an inspectable context package.

The current supported v2 candidate combines BM25 retrieval, bi-encoder pre-ranking, Cross-Encoder reranking, Routed Context Policy, and Multi-Resolution packing.

retrieve candidatesBM25
↓
pre-rank and rerankbi-encoder + Cross-Encoder
↓
allocate context formRouted Context Policy
↓
pack within budgetMulti-Resolution
↓
emit context packageevidence + provenance

03 · Evidence

Evidence-level metrics, not final answer-quality claims.

Evidence hit rate measures whether labelled evidence reached the context package. It does not prove final answer quality.

The current v2 candidate improves evidence hit rate and ranked R@3 under the measured LongMemEval-S protocol.

v2 evidence hit
0.886
v2 ranked R@3
0.6688
BEIR mini hit
1.000
answer field
False
Method Ranked R@3 Evidence Hit Rate Claim
Cross-Encoder baseline 0.6604 0.834 Baseline evidence-level path.
Routed Context Policy default 0.6604 0.876 Higher evidence coverage without ranked R@3 loss.
Two-Tier Reranker v2 candidate 0.6688 0.886 Current supported v2 evidence-level candidate.

The BEIR-style mini check is synthetic and local: corpus, queries, qrels, evidence_hit_rate 1.000, recall@3 1.000, missing_queries none. It is not real BEIR validation or a generalisation claim.

04 · Product Alpha

Trusted-local context packages

Product Alpha builds local context packages from stored memory. It exposes evidence, provenance, diagnostics, run logs, and snapshots.

It is local operator tooling, not a product UI and not a hosted runtime.

Local checks and tooling

  • Local operator check script.
  • Fixture evidence check.
  • Local operator console with CLI fallback.
  • BEIR-style mini retrieval sanity check.
  • Export, redacted export, metadata-only export, restore dry-run, safe reset, and backup verification.

05 · Limits

What this page claims

  • Evidence-level context allocation research.
  • Trusted-local Product Alpha for context packages.
  • Inspectable evidence, provenance, diagnostics, and snapshots.
  • Synthetic BEIR-style mini harness sanity check.

What it does not claim

  • No memory-solved claim.
  • No final answer-quality improvement claim.
  • No answer generation in Product Alpha.
  • No production, hosted, cloud, or public runtime approval.
  • No private/shared beta readiness.
  • No customer, regulated, or production data readiness.
  • No OAuth/RBAC or multi-user access claim.
  • No Research v3 validation.
  • No real BEIR generalisation claim.

06 · Whitepaper

Intelligent Context Windows: Evidence Allocation Before Answer Generation

Draft technical note

The whitepaper draft explains the context allocation problem, evidence-level method, current metrics, Product Alpha local context packages, and the boundaries that remain deliberately unclaimed.

The whitepaper is not a production, private beta, answer-quality, or broad benchmark-generalisation claim.