Skip to content

Hard data & reports

The right-hand end of the pipeline: stamp a table so it can travel, verify it later, assemble a board pack, and read back the ledger of everything the tools layer did. This is the one-way rule — soft data → tools → hard data — made mechanical.

The audit ledger

Every tools-layer call is recorded as it happens: the function, its arguments (frames summarised by shape and hash), the input and output content hashes, and the wall time.

sc.audit()            # the session ledger, most recent last
sc.audit(5)           # just the tail
sc.clear_audit()      # a new deliverable, a clean slate
sc.enable_audit(False)  # off (and back on) for tight loops

sc_audit()
sc_audit(5)
sc_clear_audit()
sc_enable_audit(FALSE)

                           at        fn  ...              in       out    ms
0   2026-08-23T20:04:42+00:00      load  ...            None  52a022c2…   2.1
1   2026-08-23T20:04:42+00:00  triangle  ...      52a022c2…   1b774e0b…   3.4
2   2026-08-23T20:04:42+00:00      mack  ...      1b774e0b…   18cc78c4…   5.0

The hashes chain: triangle's output hash is mack's input hash. That chain is what hard seals into a table.

Stamping

t = sc.hard(sc.mack(tri).table, assumptions={"tail": 1.0})
t.provenance
# {'sha256': '18cc78c4c8676f59…', 'at': '2026-08-23T20:04:57+00:00',
#  'scelo': '0.1.0', 'python': '3.13.12', 'pandas': '3.0.5',
#  'rows': 8, 'columns': [...],
#  'trail': [{'fn': 'load', …}, {'fn': 'triangle', …}, {'fn': 'mack', …}],
#  'assumptions': {'tail': 1.0}}
sc.verify(t)          # True — and False the moment any cell is edited

t <- sc_hard(sc_mack(tri)$table, assumptions = list(tail = 1.0))
sc_provenance(t)
sc_verify(t)          # TRUE — FALSE the moment any cell is edited

hard stamps the content hash (SHA-256 of the values in Python; MD5 of the CSV rendering in R — each verified by its own verify), the UTC timestamp, the library and runtime versions, the shape, the last twelve audit entries as the call trail, and any assumption set you attach. The stamp survives subsetting and travels with .md / .html exports; editing a single cell breaks verify, which is the point.

Board packs

sc.report(reserves, sc.life_table(), "## Method\n\nProse goes here.",
          title="Q3 reserving pack", summary="IBNR up 4 % on Q2.",
          author="A. Denewade", to="pack.html")

sc_report(reserves, sc_life_table(), "## Method\n\nProse goes here.",
          title = "Q3 reserving pack", summary = "IBNR up 4 % on Q2.",
          author = "A. Denewade", to = "pack.html")

report takes any mix of Tables, plain frames, model results and Markdown strings, stamps anything not yet hard, and writes one document: title and generation line, executive summary, every table with its basis, notes and hash, and the audit trail at the back. .html gets the IDE's cream-and-ink styling with no external dependency; anything else is returned as Markdown. Writes are atomic — a crash never leaves half a pack.

Exports and snapshots

sc.export(t, "reserves.csv")      # CSV / TSV / parquet / xlsx / json / md / html
sc.snapshot(df, "post_cleaning")  # a named, hashed copy under ~/.scelo/snapshots
sc.restore("post_cleaning")
sc.snapshots()                    # name, at, rows, cols, hash

sc_export(t, "reserves.csv")
sc_snapshot(df, "post_cleaning")
sc_restore("post_cleaning")
sc_snapshots()

Snapshots live under $SCELO_HOME/snapshots (default ~/.scelo), each with a JSON sidecar recording when it was taken, its shape and its hash — and the two languages read each other's snapshots, so a frame snapshotted in Python restores in R.

Function list

hard provenance verify export report snapshot restore snapshots · audit clear_audit enable_audit content_hash — each with the sc_ twin in R.