public alpha intent · docs-first · not a crawl

Johnny · CHIPPERS · CHAINS

Files you already have. Figure out what they are, hash the twins, keep a receipt. Do not call the receipt a court seal.

View on GitHub v0.2.0 fast-models v0.2.0 · MIT · public

the trio

Classify. Hash. Custody. Three crates, one job.

Most AI data platforms want to crawl the open web. This sketch does the opposite: materials already in your hands.

classify

Johnny Appleseed

Figure out what kind of thing a file is before you spend energy parsing it. Operator-owned materials only.

hash · inspect

C.H.I.P.P.E.R.S.

Content Hashing, Inspection, Parsing, Provenance, Extraction, Retrieval, Simplification. Unknown binaries stay metadata-only.

receipts

C.H.A.I.N.S.

Custody History Assurance Integrated Network System. Tamper-evident bookkeeping — not blockchain cosplay, not identity proof.

contract

Audit invariants

Append-only. Parent digest on every child. Quarantine and reject are first-class. “Verified” is reproducible, not legally true.

CHAINS_AUDIT_CONTRACT.md →

what it looks like

Seeds in the top crate. Gears in the middle. Receipts on a chain.

Watercolor crates: Johnny Appleseed seeds, CHIPPERS gears and index cards, CHAINS receipts on a chain
CLASSIFY · HASH · CUSTODY — conceptual crate stack
Three bands: classified papers and boxes, a glowing hash cube splitting identical files, wax-sealed custody envelopes
Classify the pile. Hash the twins. Seal the receipt.

do not conflate

Three reclaim layers. A catalog is not a fourth.

File-hash reclaim is not bees. Structure clutter is not byte proof. The Johnny index is a call graph, not a delete plan.

L1  structure     dual install roots, husks, path chaos
L2  content-hash  size → 4 MiB sample → SHA-256 (only on collisions)
L3  bees          Btrfs extent sharing — occupancy is not “disk full”

catalog plane     classify → hash envelope → custody receipt
                  (different job; do not renumber L1–L3 into it)

lab checks · 2026-08

The internal tool learned a few sharp edges. The public notes now say them.

Process truth only. No private hosts, no care data, no “we shipped extractors.”

walks

No unbounded find

Agents read catalogs. Structure probes are allowlisted, depth-capped, no follow-mount. A client-side walk of a storage pool is the anti-pattern.

venvs

Skip callable envs on L2

L2 hash audits skip .venv / venv by default. Darwin and Linux environments are not interchangeable. Never share them across OS families.

gate

sha256 + a real manifest

Apply / purge needs verified=true and all_hosts_readable=true. Sample match is not enough. A staging filter is not an apply.

schema

Reshape before apply

Staged JSON and the apply tool used different field names and looked similar enough to be dangerous. Reshape, or do not run it.

libs

.so is not just .so

Vendor shared libraries come as .so, .so.12, .dylib. A naive suffix check misses the ones that matter.

mount

zstd is not a cron job

Pool compression is a filesystem / mount property. bees occupancy is not disk-full. Same-filesystem mv does not free space until rm.

tools

Probe, don’t assume

Record {tool_id, version, manager, host_role, probed_at}. No silent brew upgrade. “latest” is not provenance for a weight.

stage

Download on the storage host

Land in a stage prefix. Verify size (and hash when known). Promote. Laptop-through-NFS is how you get two copies and a lie.

rendition contract

Originals are never overwritten.

00_original → 10_assured → 20_extracted → 30_cleaned → 40_simplified →
45_markdown → 50_accessible → 60_index → 70_exports

batch = IngestionEngine().ingest(materials)
verified = IngestionEngine().verify(batch)
Callers acquire bytes and choose scope. The engine returns artifacts, portable indexes, and optional receipts — no direct network, filesystem, or database writes inside the module. Adapters consume return values.

siblings

This is the custody sketch. The other pages are the house around it.

the pool

fast-models

Btrfs + bees + host NFS. L3 lives there. This repo tells you not to confuse it with a file-hash report.

the1truedan.github.io/fast-models → the door

ai-gateway

Headroom → LiteLLM. Outside the file custody chain. Same audit spine is allowed as another event class — policy, not a second product.

the1truedan.github.io/ai-gateway → the launchers

grok-tua / tok-tua

Dual-pane CLIs that show whether the door is up. They do not classify your PDFs.

the1truedan.github.io/grok-tua-tok-tua → the pictures

mok-tua

Local Comfy storyboard box. The crate art on this page is conceptual; the origin comics live there.

the1truedan.github.io/mok-tua →

how it came to be

Care work produces paper. Then the same file shows up twice.

Notices, letters, scans, exports, model weights, tool installs. Early M.A.N.A.G.E.R. design chats kept returning to one problem: how do you classify and hash what you already hold, keep a plain custody trail, and never pretend a receipt log is a court seal?

This repo is the design sketch for that pipeline, split out so strangers can poke holes in it.

13 April 2026 — mission pivot. June 2026 — monorepo coding. Late July 2026 — modular public sketch. 6 August 2026 — alpha-intent narrative. 16 August 2026 — this page, plus the lab checks written down.

Independent. Not affiliated with any package ecosystem or download manager named in the catalog. © 2026 M.A.N.A.G.E.R. LLC — prepare for the care when we cannot be there.