ManifoldMemory / Warrant

Earlier versions reported “MixK over full pool” numbers suggesting the latent processor could serve as a first-stage retriever. A later methodology audit found the harness was accidentally conditioned on the gold document latent at retrieval time. After removing the oracle, non-oracle first-stage P@1 collapses to 0.0430 at 7.77M — on par with NDN mean-pooling. The paper explicitly revises the claim: those original numbers are an oracle bi-encoder upper bound, not a shippable first-stage retriever.

The correction changes the product architecture. MixK is not the first-stage retriever. MixK is a reranker / navigator over a candidate set. The production-realistic system becomes:

Navigating the Manifold reports the shipping pipeline as BGE-large ∪ QNDN v0 → MixK, with stage-2 P@1 lifted from ~0.18 to ~0.34 and P@10 from ~0.20 to ~0.47 over the deployed BGE-small → MixK baseline (+83 % / +133 % relative). It states that no single retriever in the inventory matches the union.

The honest reading is not “SOTA memory”. It is: structured context delivery substantially improves a modest open-weight reader under a private retrieval setup. Public phrasing: “Warrant improves self-hosted open-reader memory performance by routing evidence, not by relying on a frontier full-context reader.”

This part is commercially strong but scientifically less mature than the retrieval work. Current audits are promising but small. A continuation team should commission larger adversarial refusal benchmarks before lean-pushing this as a SOTA refusal claim.

ManifoldMemory began as an attempt to compress language into dense latent memory and ask whether a model could reason over those latents instead of raw tokens. The first research phase falsified the obvious retrofit path: pretrained LLMs with cross-attention or prefix adapters did not read compressed latents, as shown by a paired wrong-latent diagnostic and gate-trajectory analysis across 46 experiments. The second phase discovered a more viable substrate: a frozen reconstruction-autoencoder latent manifold, navigated by a small latent-native processor, exhibits measurable content-addressable memory behavior for doc-anchored QA at million-candidate scale. The third phase turned that mechanism into a deployable retrieval architecture: a 30M Perceiver/MixK navigator, trained with a mixed-K curriculum, reranks candidate unions from standard dense retrieval and a latent-native QNDN retriever.

The central empirical discovery is that QNDN is not a better BGE-style retriever at rank 1. It is a different retrieval geometry. BGE-large is sharp-headed (strong at P@1 and shallow ranks); QNDN v0 is shallow-tailed (weak at P@1, but better at deep-K recall). Their union creates substantially higher candidate coverage than either alone; MixK then concentrates that coverage into top-k evidence for a reader. This makes the product commercially meaningful: Warrant can improve private-corpus evidence retrieval without asking customers to send data to external APIs.

The current project should be treated as a serious narrow-breakthrough candidate, not as a solved general retrieval system. Its strongest scientific claim is “retrieval as manifold geometry”; its strongest product claim is “private, auditable, evidence-first retrieval for regulated corpora.” The next step is independent reproduction, modern baseline bakeoff, cross-domain validation, and — if pursuing the deepest version — co-evolved latent-native models trained from scratch to operate inside the manifold rather than retrofitted onto language-token priors.