Archive snapshot — 2026-07-28

A prebuilt snapshot of observatory.db — the queryable archive — for readers who want to look
without installing anything. Download, open in any SQLite browser, and query.

The database is DERIVED, and this file is a convenience, not a source of truth. The record is
the hashed JSONL in the repository. This snapshot can be deleted and regenerated at any time,
byte-identically, from those artifacts:

pip install hardmap
hardmap db build

at the commit this release targets (0ce3b2f) that rebuilds this database. The archive grows,
so a later checkout builds a later — and equally valid — one; pin to this tag if you want this content.

On comparing hashes. This file's sha256 is f84dd1fcaec5cbee…, and your rebuild will very likely
differ — that is expected and not a fault. SQLite stamps the version of the library that wrote a
database into its file header, so a different sqlite build produces a logically identical database with
different bytes. What is portable, and what you should actually compare, is in the attached
observatory_db_manifest.json: the 23 source hashes (of the JSONL, which is the record) and the
row counts. Those match across machines; the file hash is a within-environment fingerprint.

Every table carries the sha256 of the artifact it was compiled from, so the database and its sources can
be checked against each other rather than trusted — and if the two ever disagree, the JSONL is
right. The manifest is attached so you can verify all 23 source hashes without a checkout.

What's in it

problem rows, with charges from classification theorems	346
frames of measured region geometry along difficulty ramps	2,032
catalog cells (descriptors, versioned extraction rule)	446
candidate questions enumerated, screened and dispositioned	1,942
reserved rows — declared, never captured	16
maptrail records — what changed, in the domain's vocabulary	83

The unusual table is candidates: every question the data could have supported and why each was
screened, held, sealed or killed. A multiple-comparisons correction computed from an enumeration you
can audit, rather than a number you have to trust.

Getting started

hardmap query --list                 # eight worked queries, by name
hardmap query rejected-candidates
hardmap query --sql "SELECT COUNT(*) FROM catalog WHERE excess_ref IS NOT NULL"

Or in a SQLite browser, start with SELECT * FROM maptrail ORDER BY at — the biography of the
territory.

A snapshot, not a version. This is a dated data release, deliberately not tied to the package
version: the archive grows with every batch and the code that reads it moves on its own schedule.
Tags of the form archive-YYYY-MM-DD are data snapshots; vX.Y.Z are the software.

hardmap v1.0.0

First public release — reproducible evidence for the charge-atlas / proof-space program.

pip install hardmap && hardmap repro --all regenerates all eight paper-cited numbers (eightfold recomputed from the frozen atlas, proof-census from the committed proof checkpoint, foundry from the verified oracle matrices); hardmap verify runs the internal-coherence sweep.

• History-preserving consolidation of four research projects; every sealed preregistration resolves via docs/seal-chain.md.
• Sealed verification pass: docs/findings/H4-verification.md.
• Dual-licensed: Apache-2.0 (code) + CC-BY-4.0 (data/prose).