---

title: SZL Second Brain emoji: 🧠 colorFrom: indigo colorTo: green sdk: docker app_port: 7860 pinned: false license: apache-2.0 short_description: Governed retrieval and review-gated continuous frontier memory. tags:

• retrieval
• hybrid-search
• governance
• fail-closed
• evidence
• szl-holdings

---

SZL Second Brain

A governed memory plane for SZL inference: public handles and digests outside; authorized content hydration inside the trusted controller.

The public projection contains 575 governed chunks. The private 9,464-node graph is not published, is not queried by this package, and is never admitted to gradients. The index is data, not model weights. Lambda uniqueness remains Conjecture 1.

Installed-mode guarantee

Version 1.3 packages the public corpus, schemas, and the review-gated frontier candidate index into the wheel. Installing szl-second-brain in a clean environment can build the 575-chunk index, run governed hybrid retrieval, inspect the exact frontier state, search review candidates by handles, and perform controller-only authorized hydration without a source checkout.

The wheel contains only public material. It contains no private graph rows, credentials, model weights, signing keys, or execution authority. CI builds the wheel, installs it into an isolated environment, changes outside the repository, and proves corpus discovery, frontier discovery, retrieval, digest parity, Living Anatomy feed generation, and authorized hydration.

An operator may still provide a different governed projection through SECOND_BRAIN_CORPUS or AYLLU_BRAIN_CORPUS. Explicit corpus paths remain controller decisions and must pass the same per-row digest checks.

Continuous frontier memory

The hourly Continuous frontier memory workflow is a bounded discovery loop, not autonomous retraining. It reads only seven fixed public source contracts:

• the exact szl-formulas formula/quant atlas;
• the canonical szl-ouroboros bounded-loop kernel README;
• the canonical szl-kernels governed suite and MiniEmbed truth card;
• the Living Anatomy README;
• the A11oy public-estate manifest;
• the Forge production-controller contract;
• the Nemo witness README.

Each source is resolved to the latest exact commit that changed its admitted path, fetched from immutable raw GitHub, scanned for secret-like material, and converted into content-addressed candidates. The committed state records the exact candidate count and digest, including 30 attributed formulas, 21 executable formulas, nine quant domains, the active kernel/model truth cards, the A11oy public topology, Forge controller contracts, and source-document sections.

A changed candidate set creates one content-addressed review branch and attempts to open its pull request. When organization policy blocks Actions from opening pull requests, the receipt reports BRANCH_READY_EXTERNAL_PR_REQUIRED and an authorized founder must open that exact branch without rewriting it. The workflow cannot force-push, merge, train, publish weights, reveal candidate content through the public API, load the private graph, or mutate a provider. Candidate state is always DISCOVERED_REVIEW_REQUIRED until a separate reviewed process acts.

This is the operational meaning of “the Brain keeps learning” here:

fixed public sources
       ↓ exact path revisions + SHA-256
bounded candidate extraction
       ↓
handles-only review index
       ↓
content-addressed review PR
       ↓
human-governed acceptance or rejection

It is continuous evidence acquisition, not silent model self-modification.

Formula and quant authority

The frontier memory consumes szl-holdings/szl-formulas as the active executable kernel and its source-attributed formula/quant atlas. The atlas preserves all 30 attributed records and nine quant domains while keeping three different concepts strictly separate:

• per-obligation PROOF_STATUS on the 21 executable functions;
• source-reported status on the attributed formula corpus;
• membership in the locked-proven set of exactly eight formulas.

No status string promotes a formula. The F-number-to-executable mapping remains UNKNOWN_NOT_INFERRED, and Lambda remains CONJECTURE_1_OPEN_ADVISORY_ONLY. Empirical and conjectural records are evidence or review inputs, never execution authority.

Retrieval modes

HybridSecondBrain is an engine-neutral retrieval coordinator:

BM25 candidates
      +
optional revision-pinned dense candidates
      ↓
reciprocal-rank fusion
      ↓
optional reranker
      ↓
source-diversity limit
      ↓
handles + source/content digests only

The public Space deliberately has no opaque dense provider configured, so it reports BM25_ONLY. A trusted Forge deployment can inject a qualified dense provider and reranker. If a provider fails, the response either blocks or explicitly reports BM25_FALLBACK_DENSE_UNAVAILABLE; it never claims a hybrid run that did not occur. Similarity and ranking are never represented as correctness.

Authorized hydration

AuthorizedHydrator and AuthorizedFrontierHydrator are library-only controller components. They require:

• a non-empty principal ID and tenant ID;
• an immutable policy revision;
• an explicit per-node authorization callback;
• matching node identity, source, and SHA-256;
• an untampered public corpus or frontier-candidate row.

Any denial, provider error, duplicate, unknown handle, source mismatch, or digest mismatch fails the whole hydration request closed. Hydrated text is never exposed by the public FastAPI application.

from second_brain import (
    AuthorizedFrontierHydrator,
    AuthorizedHydrator,
    HybridSecondBrain,
    frontier_search,
)

retriever = HybridSecondBrain(
    dense_provider=my_dense_provider,
    reranker=my_reranker,
)
context = retriever.context("locked formula authority", k=6)

hydrator = AuthorizedHydrator(my_authorizer)
hydrated = hydrator.hydrate(
    context["handles"],
    principal_id="principal-123",
    tenant_id="tenant-abc",
    policy_revision="<full immutable revision>",
)

review = frontier_search("formula quant anatomy ouroboros", k=12)
review_hydrator = AuthorizedFrontierHydrator(my_frontier_authorizer)
review_content = review_hydrator.hydrate(
    review["handles"],
    principal_id="reviewer-123",
    tenant_id="tenant-abc",
    policy_revision="<review policy revision>",
)

Public surfaces

GitHub is the source of truth; the Hugging Face Space is a deployed public surface. Apache-2.0. Doctrine v11.