Limitations

This library is reference IP for adoption, not a control operating in production. Read these limits before relying on any part of it.

Claim layer

The five primitives are real, tested reference patterns. The seven adviser controls are implemented and tested reference controls. None of them is a deployed system at any firm, and none has been examined by a regulator.
Nothing here is legal advice. The regulatory characterizations are engineering summaries; confirm every obligation against the primary source and a qualified compliance/legal function before relying on it.

The deployer supplies the judgment. The MNPI control does not infer what is or is not material nonpublic information — the compliance function curates the restricted/watch lists; the control enforces the consequence (a restricted-name order is blocked). The best-execution control does not measure execution quality for you — it gates on deployer-supplied factors and a deployer-set benchmark and tolerance. The marketing control gates on structured attributes a reviewer asserts, not on reading advertisement prose.
Allocation fairness surfaces a statistical asymmetry (favored vs. other accounts receiving favorable fills); it does not adjudicate intent, and a flag is a signal for review, not a finding of cherry-picking.
Independence is attested, not detected. The effective-challenge harness enforces challenger != primary in code, but vendor-family and prompt-template independence are an operator attestation — no detector is fabricated. A false attestation is the operator's risk, recorded to the chain.

The hash chain is internally consistent by construction: it detects after-the-fact mutation and chain-splice on replay, within the trust boundary. It does not prevent a privileged in-process actor from rewriting the store, and it does not, by itself, make end-to-end regeneration detectable.
End-to-end regeneration (rebuilding the whole chain with internally consistent hashes) is only caught when chain heads have been anchored to an external witness register the deployer does not control alone (OpenTimestamps, a transparency log, a regulator-side log). Production mode requires one and fails closed without it. The bundled InMemoryWitnessRegister is a reference for tests only — it is not durable and is not external.

The default advisory mode is backward-compatible and labeled advisory — it records honestly (authenticated=false, mode="advisory") but does not fail closed. Only production mode enforces the fail-closed contracts (a wired Authorizer / attestation verifier / witness register). Choosing advisory mode in a real deployment is the deployer's risk.

These are reference controls at reference depth. Known shallowness an examiner would extend, by design for 0.1.0:

Best execution is reviewed per order against a deployer benchmark — a pre-trade/at-trade TCA check, not the periodic-and-systematic, across-venue review the regular-and-rigorous obligation also requires. Treat the per-order gate as one input to a periodic review the deployer runs.
MNPI surveillance matches on the normalized symbol of the named instrument. It does not expand an issuer to its instrument family (bonds, options, ADRs, single-name CDS, index/ETF constituents); a deployer who needs family coverage supplies the expanded restricted set.
Allocation fairness models trade-level block-fill cherry-picking. PE/credit conflicts that turn on deal/co-investment allocation across funds and LPs, and on fee-and-expense allocation (broken-deal, monitoring, expense shifting), are not modeled here and are the deployer's to extend.
Marketing review gates the rule's headline conditions (hypothetical- performance policies, testimonial disclosure, net-of-fees). It does not cover every 206(4)-1 sub-condition (predecessor / related / extracted performance, third-party ratings, per-period gross-and-net).
Controls record and surface; they do not orchestrate enforcement. A BLOCKED screen or an unapproved review is recorded — wiring that result to a sovereign veto (so the order/communication is actually stopped) is the deployer's integration, not an automatic side effect.

This library models the investment-adviser fiduciary regime only. It does not model the obligations of any other regulated actor, and it is not a substitute for a firm's compliance program, Form ADV, or its written §206(4)/Rule 206(4)-7 policies and procedures.
The DEFCON thresholds are illustrative examples, not calibrated values. Calibrate them to the program's strategy, capital base, and risk appetite.