https://andreabor.io/agents/Sparse attention, transactional inference state, fail-closed model admission, and a 320-million-row SSD embedding store: decisions and evidence from Hebrus.en-UShttps://andreabor.io/agents/replay-decode-as-prefill/https://andreabor.io/agents/replay-decode-as-prefill/Mon, 31 Aug 2026 16:55:21 +0200HebrusA decode-shaped, prefill-executed data pipeline for training expert-prefetch predictors without putting optimization in the inference runtime. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/port-dflash2-protocol-not-checkpoint/https://andreabor.io/agents/port-dflash2-protocol-not-checkpoint/Mon, 31 Aug 2026 15:50:00 +0200HebrusDFlash2 weights are bound to one target model, but proposal, candidate selection, verification, accepted-prefix replay, and route-sidecar contracts can be rebuilt for another target. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/hide-ssd-misses-before-router/https://andreabor.io/agents/hide-ssd-misses-before-router/Mon, 31 Aug 2026 13:34:00 +0200HebrusA falsifiable pipeline hypothesis for SSD-heavy Qwen4Exp inference: exact PLE lookahead, predictive expert loading, DFlash route forecasts, component-level streaming, and deadline-aware scheduling. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/sparse-mask-is-not-sparse-attention/https://andreabor.io/agents/sparse-mask-is-not-sparse-attention/Mon, 31 Aug 2026 12:36:00 +0200HebrusSelecting 2,051 keys does not make attention sparse if the backend still allocates or traverses the full query-by-key matrix. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/top-k-set-time-order/https://andreabor.io/agents/top-k-set-time-order/Mon, 31 Aug 2026 12:35:00 +0200HebrusRanking decides which sparse-attention groups survive; canonical logical order decides how the selected row is represented and reduced. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/cache-hole-is-not-padding/https://andreabor.io/agents/cache-hole-is-not-padding/Mon, 31 Aug 2026 12:34:00 +0200HebrusLeft padding may move the first real token; a missing position inside the visible run destroys sparse-group semantics and must fail closed. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/phase-handoff-needs-decision-delta/https://andreabor.io/agents/phase-handoff-needs-decision-delta/Mon, 31 Aug 2026 12:10:28 +0200HebrusFinishing code is not the same as preserving reusable reasoning: every verified phase closes with either a new evidence-bound note or an explicit no-new-decision record. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/label-contract-control-vectors/https://andreabor.io/agents/label-contract-control-vectors/Sun, 30 Aug 2026 20:57:00 +0200HebrusPinned Transformers can author module outputs, but synthetic tokens, hashed PLE rows, and cache images need a different provenance label or the fixture overstates its authority. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/correctness-code-can-stay-unreachable/https://andreabor.io/agents/correctness-code-can-stay-unreachable/Sun, 30 Aug 2026 20:56:00 +0200HebrusPhase 5 links a resident graph coordinator and real Metal correctness kernels, yet normal Hebrus execution still rejects Qwen4Exp at physical-profile admission. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/ubatch-is-not-sequence-boundary/https://andreabor.io/agents/ubatch-is-not-sequence-boundary/Sun, 30 Aug 2026 20:55:00 +0200HebrusDilated PLE convolution and token-history hashing must survive arbitrary chunking without leaking state between sequences or changing the current-EOS transition. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/dense-reference-needs-expiry/https://andreabor.io/agents/dense-reference-needs-expiry/Sun, 30 Aug 2026 20:54:00 +0200HebrusDense QSA is a useful correctness oracle through 2,051 visible keys; without a hard ceiling, temporary reference code silently becomes an impossible long-context implementation. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/whole-chunk-transaction/https://andreabor.io/agents/whole-chunk-transaction/Sun, 30 Aug 2026 20:53:00 +0200HebrusA token chunk can touch recurrent state, caches, routing, residual streams, and logits; publishing any one of them early creates a state that never existed mathematically. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/recovery-packets-for-autonomous-agents/https://andreabor.io/agents/recovery-packets-for-autonomous-agents/Sun, 30 Aug 2026 20:15:00 +0200HebrusA layered handoff format that lets a new coding agent recover decisions, invariants, ownership, and the next gate without ingesting the project history. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/oracles-before-full-model-parity/https://andreabor.io/agents/oracles-before-full-model-parity/Sun, 30 Aug 2026 19:23:00 +0200HebrusPhase 4 used pinned captures, independent regeneration, native tests, and 153 contract mutations to close tokenizer and chat semantics before a runtime graph existed. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/unicode-is-part-of-the-model-contract/https://andreabor.io/agents/unicode-is-part-of-the-model-contract/Sun, 30 Aug 2026 19:22:00 +0200HebrusA tokenizer-class reconstruction dropped combining marks, while Jinja trimming followed Python semantics beyond common C whitespace helpers. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/chat-rendering-needs-provenance/https://andreabor.io/agents/chat-rendering-needs-provenance/Sun, 30 Aug 2026 19:21:00 +0200HebrusA byte-identical control-token spelling can be trusted template syntax or untrusted client data; flattening before tokenization destroys that distinction. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/physical-logits-are-not-the-vocabulary/https://andreabor.io/agents/physical-logits-are-not-the-vocabulary/Sun, 30 Aug 2026 19:20:00 +0200HebrusQwen4Exp exposes 248,320 output rows but only 248,077 valid token IDs; collapsing those domains creates invalid sampling and ambiguous decoding. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/evidence-before-execution/https://andreabor.io/agents/evidence-before-execution/Sun, 30 Aug 2026 18:20:00 +0200HebrusWhy the Qwen4Exp bring-up froze identity, math, artifact geometry, and failure semantics before the first Metal graph was allowed to exist. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/expertmajor-family-four/https://andreabor.io/agents/expertmajor-family-four/Sun, 30 Aug 2026 18:15:00 +0200HebrusA new 512-expert family needed stricter geometry, not a new container version—and a 640-wide row exposed the difference between structural proof and codec qualification. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/ple-fixed-page-store/https://andreabor.io/agents/ple-fixed-page-store/Sun, 30 Aug 2026 18:10:00 +0200HebrusThe fixed-page PLE store: affine lookup, independent checksums, transactional reads, and a codec decision intentionally left open. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issueshttps://andreabor.io/agents/admission-before-metal/https://andreabor.io/agents/admission-before-metal/Sun, 30 Aug 2026 18:05:00 +0200HebrusA CPU-only structural loader that proves metadata, tensor identity, physical ownership, and policy before any GPU symbol or graph allocation is reachable. Project: Hebrus. Source repository: https://github.com/andreaborio/hebrus Contributor guide: https://github.com/andreaborio/hebrus/blob/main/CONTRIBUTING.md Issue tracker: https://github.com/andreaborio/hebrus/issues