# Decision records Each record fixes one decision that would otherwise be reopened by accident. They rank above the implementation and below `AGENTS.md` in the precedence order that file states. Applying that order means knowing what each record decides, which is what this index is for: one line per record, saying what it settles rather than what it is about. | Record | What it decides | |---|---| | [0001](../decisions/0001-foundational-decisions.md) | The four decisions the project was founded on, in the words they were first written in. No shared core package; 0002-0004 are retired and never reused | | [0005](../decisions/0005-process-pool-for-distance-workers.md) | Distance shards run on a `ProcessPoolExecutor` with the `spawn` context, measured against threads | | [0006](../decisions/0006-object-source-axis.md) | Producing object bytes is one concern and it belongs to the normalizer, whether the objects come from a table or from files | | [0007](../decisions/0007-run-self-containment.md) | Adopted files are copied into the run directory, never referenced where they live | | [0008](../decisions/0008-named-object-encoding.md) | The encoding that produced the measured bytes is part of the result, recorded as a named, versioned `object_encoding` | | [0009](../decisions/0009-spreadsheet-engine.md) | Spreadsheets are read with `openpyxl` in read-only mode, not with `pandas.read_excel`, because bounded memory is the invariant | | [0010](../decisions/0010-cell-text-rule.md) | The rule that turns a typed spreadsheet cell into text is this project's, not the parsing library's | | [0011](../decisions/0011-dead-schema-fields-await-a-breaking-release.md) | Three informationless fields in the published 0.2 schemas stay until a breaking release removes them | | [0012](../decisions/0012-duplicated-helpers-stay-duplicated.md) | Small helpers duplicated across packages stay duplicated; the evidence ADR 0001 asserted without it is here | | [0013](../decisions/0013-the-public-signature-stays-flat.md) | `run` and `estimate` take the source axis as flat keyword arguments rather than a union a caller would have to construct | | [0014](../decisions/0014-dataset-partitioning.md) | Dataset partitioning lives upstream of the source axis, emits datasets, and its results are fixed by a named rule | | [0015](../decisions/0015-the-bibliography-is-fetched.md) | A work with a DOI is never typed: `docs/refs.json` is fetched from a resolver, and no mechanism here checks that a cited work supports the claim beside it | | [0016](../decisions/0016-branch-lengths-become-edge-weights.md) | The mapping from branch lengths to FastGreedy edge weights is this project's own and decides the partition | | [0017](../decisions/0017-the-neighbor-joining-tie-rule.md) | Q scores within a relative band are one tie, broken in lexicographic label order; neither published source states a tie rule | | [0018](../decisions/0018-method-provenance-in-the-manifest.md) | The manifest records which named rule was in force at each stage, separately from the parameters `config_hash` already covers | | [0019](../decisions/0019-an-attribution-declares-whether-it-was-read.md) | Every attribution to a published work declares whether anybody read it against this code, and the gate checks shape, never truth | | [0020](../decisions/0020-ppmd-is-the-default-compressor.md) | PPMd variant I with restart is the default compressor, its order and memory are configuration, and `pyppmd` is a required dependency | | [0021](../decisions/0021-bzip2-and-lzma-join-the-compressors.md) | bzip2 and raw LZMA2 join the compressors; each exposes only the settings that change its window, its stream memory is charged per worker, and its codec is named by a probe | | [0022](../decisions/0022-execution-settings-are-not-run-identity.md) | `config_hash` leaves out the execution settings, which `config` still records, so a run resumes and is reused on any machine and under any worker count | | [0023](../decisions/0023-the-compressor-window-is-a-correctness-gate.md) | A pair longer than the window its compressor declares is refused by preflight, under a factor that lifts it; PPMd declares none, because its model size is not a limit the measurement shows | ```{toctree} :maxdepth: 1 /decisions/0001-foundational-decisions /decisions/0005-process-pool-for-distance-workers /decisions/0006-object-source-axis /decisions/0007-run-self-containment /decisions/0008-named-object-encoding /decisions/0009-spreadsheet-engine /decisions/0010-cell-text-rule /decisions/0011-dead-schema-fields-await-a-breaking-release /decisions/0012-duplicated-helpers-stay-duplicated /decisions/0013-the-public-signature-stays-flat /decisions/0014-dataset-partitioning /decisions/0015-the-bibliography-is-fetched /decisions/0016-branch-lengths-become-edge-weights /decisions/0017-the-neighbor-joining-tie-rule /decisions/0018-method-provenance-in-the-manifest /decisions/0019-an-attribution-declares-whether-it-was-read /decisions/0020-ppmd-is-the-default-compressor /decisions/0021-bzip2-and-lzma-join-the-compressors /decisions/0022-execution-settings-are-not-run-identity /decisions/0023-the-compressor-window-is-a-correctness-gate ```