Payments: Sep 14 Invite coupons available · Workspace open
KBMILL architecture Enter the plant

Diligence · Level 3

Architecture

In plain English: KBMill receives a bounded document pile, processes it in isolation, removes unreliable material from the answer path, records what happened, and returns a package you can use without KBMill.

KBMill performs corpus manufacture: messy source documents → residual-honest knowledge bricks ready for retrieval. Unreliable parts are listed, not silently quoted as fact.

What runs behind the hopper states there is a plant. This page is the artifact that kills “just a script” — system boundaries, real station names with failure modes attached, a closed fail path, and a snippet from a real package. It is not source code and not the closed magazine.

System boundary (public-safe)

[ Hopper / public door ]
        │
        ├─► Receive   — isolated per-job dir, idempotent, caps refuse before run
        ├─► Extract   — layout + table recovery + OCR trap handling
        ├─► Structure — brick plan → build → link (composition, not mega-melt)
        ├─► Package   — ZIP + residual board + SECURITY_REPORT + MILL_RECEIPT
        │
        └─► 72h download window · fail closed · no charge on fail
            = more copies of the same plant (commodity tower → heavier mirrors)

Host owns the clock. Kitchen work can detach. Terminal envelope: done | failed | timeout. Ship cue: ready to ship | warnings | blocked.

Three mouths (same plant)

  • Factory — manufactures the package (brick).
  • Constructor — federates a shelf: link related bricks; open set stays small on purpose — composition, not one melt.
  • Connect — serves one brick into your stack.

Vocabulary for diligence readers. Implementation details of cutters and the closed magazine stay internal.

Stations as systems (real names, real gates)

A script can do “extract PDF → chunk → embed.” A script does not carry years of trap handling and a closed fail class that still authors a residual board when the model times out.

  1. Receive File checks and nested-archive unpack happen in an isolated per-job directory. Ceiling on this public door: 50 files / 500 MB — refuse before processing, no charge. Idempotent: same job identity → same work root; no double-build, no double charge.
  2. Extract Dual-path layout work, table recovery, OCR trap handling. Image-heavy manuals are a known failure class — the plant is built to ship text-first packages plus a SECURITY_REPORT.md, not a silent dump of figure shells.
  3. Structure Plan bricks → build → link related packages. Multi-brick programs are composition with an intentional small open set — not melting the estate into one unreliable mega-index. See packages, not monoliths.
  4. Package Deterministic portable ZIP: primary markdown, bundle_manifest.json, chunks.jsonl (with per-chunk content_hash), embeddings + provenance when produced, craft_brief.md, residual / mute board (listed exclusions + muted chunk flags), SECURITY_REPORT.md, MILL_RECEIPT.md.

Failure path is the moat

“On failure you are not charged” is table stakes. The plant difference is a closed fail path: fail once, cleanly; ship cue and fail class are enumerated; the leave-behind still tells you what was muted or blocked. Failures on this door clear within 24 hours.

One public trap class: figure-shell / image-only extracts. They look like “content” to a naive chunker and poison retrieval or training feed. The plant mutes them and lists the mute — see the snippet below from a real package (public shelf lineage). That class of trap is why “PyMuPDF + generic chunker” is not a plant.

Compiler proof — real package snippets

Collapsed from a manufactured portable ZIP (US Army TM 11-666 lineage). Open full samples on the public shelf.

bundle_manifest.json (trimmed)

{
  "format": "vectorforge_portable_bundle_v1",
  "schema_version": "vectorforge_kb_v1",
  "kb_name": "TM_11_666_Antennas_Radio_Propagation",
  "kb_version": 1,
  "total_chunks": 530,
  "artifacts": {
    "primary_markdown": "TM_11_666_Antennas_Radio_Propagation.md",
    "kb_json": "kb.json",
    "chunks_jsonl": "chunks.jsonl",
    "embeddings": "embeddings.npy",
    "embedding_provenance": "embedding_provenance.json",
    "security_report": "SECURITY_REPORT.md",
    "craft_brief": "craft_brief.md"
  },
  "exported_at": "2026-08-19T20:12:14.007320"
}

Muted chunk flag (from chunks.jsonl)

{
  "chunk_id": "chunk-0004",
  "muted": true,
  "mute_reason": "Figure-shell / image-only extract — excluded from RAG",
  "page_number": 4,
  "source_document": "TM11-666.pdf",
  "content_hash": "b93a05d0c66e8462",
  "token_count": 8
}

When an architect sees manifest + per-chunk hash + listed mute, they are looking at a manufactured artifact — not a chat export.

What this page deliberately does not publish

  • Closed cutter magazine / enzyme internals
  • Full operator SOP and agent loop budgets
  • Hardware mirror counts and GPU SKU plans (NDA / data room)

We show enough that a serious buyer believes manufacturing exists. We do not show enough that a serious engineer can stand up a twin from the website. Naming stations with failure modes attached raises the diligence bar for wrapper competitors — it does not ship them our edge-case code.

Where to go next

Behind the hopper · Data boundaries · Sample packages · Enter the plant · /llms