Result indexes and tidy exports
Export a recorded run to a new directory:
metewand export --output-dir runs/example --format csv --destination tables/example
metewand export --output-dir runs/example --destination tables/example-jsonlThe destination's parent must exist. --destination authorizes creating the directory and its two files: results.csv or results.jsonl, and the adjacent lineage manifest export.json. JSONL is the default. Metewand prepares both files privately and publishes the directory without replacing an existing path, including an empty directory or a symbolic link. These rebuildable outputs do not carry the run store's durable-publication guarantee.
The global --output human|json|jsonl option controls the command receipt, not the table encoding. Machine receipts conform to export-output inside the existing machine-output envelope. A successful export exits with 0 even when the selected history contains failed attempts.
Authoritative evidence and the index
metewand export --output-dir runs/example --rebuild-index--rebuild-index explicitly authorizes replacing only index.jsonl at the top of the run directory. It can accompany --destination. The index always contains the complete run, regardless of export filters. An existing index must be a regular file; symbolic links and directories are rejected.
Each rebuild reads a private, committed SQLite snapshot and verifies registered artifact inventories and content hashes through the same reader used by status and explain. The reader does not initialize, recover, or checkpoint the source store, launch workers, or require the original repository. Unregistered directories never establish an observation or acceptance. An unavailable store or corrupted artifact fails the command.
The index is a disposable projection, never coordination state. Exports always read the database and finalized artifacts directly, even when an index is present. A missing, stale, or malformed index cannot hide history, establish acceptance, or authorize deletion. Rebuilding atomically replaces the index; unchanged evidence produces identical canonical JSONL bytes. Neither index creation nor export is an implicit side effect of inspection commands.
Records and rows
The index contains one versioned export-record per committed applicability decision, attempt, and observation. Both applicable and not-applicable decisions retain their original evidence. Current one-shot runs register unconditional applicability. Pending logical slots without allocated attempts are available through status, rather than manufactured as attempts in the index.
Attempts retain every retry, including accepted, invalid, censored, superseded, failed, and nonterminal attempts. attempt_state preserves the stored state. A failed attempt with failure_code: "invalid_result" additionally has outcome: "invalid" and valid: false. Other failed attempts have unknown validity unless they produced a separately registered observation. Censoring does not become a technical failure, and supersession does not invalidate an earlier result.
Observations retain independent valid, selected, and completion fields, as well as their producing attempt's current state and outcome. A valid, unselected observation remains visible after its attempt fails, is censored, or is superseded. No export averages repetitions, chooses a winning retry, filters warm-ups, or drops attempts that lack metrics.
Tidy tables use export-row: one row per evaluator metric coordinate on an observation, plus one row with a null metric coordinate for every entity without metrics. Metric coordinates are RFC 6901 JSON pointers. Object keys sort bytewise; arrays and empty nested objects remain intact JSON values rather than becoming invented repetitions. An actual null metric value remains distinguishable from an absent metric through its nonnull metric_name.
Both table formats expose numeric timed_wall_time_ns, implementation_time_ns, evaluation_time_ns, cpu_time_ns, and peak_rss_bytes columns. Wall time, CPU time, and resident-set size also carry complete, partial, or unavailable status columns. Partial resource values remain diagnostic evidence. The original timing, measurements, coverage reasons, and accounting scopes remain available in the nested evidence.
Rows preserve candidate, logical and resolved run, attempt slot, observation slot, attempt, and retry identities. provenance and evidence retain the original documents and coordination, including recorded component seeds, pairing and scheduling blocks, run order, warm-up reports, timeout limits, machine context, worker environments, and observed controls. Missing historical evidence stays null. Export does not resolve missing provenance on the current host. Schema references come from the recorded result and metric envelopes. The current one-shot contract does not record metric unit or direction annotations; their columns remain null.
JSONL uses canonical JSON followed by one newline per row. CSV uses UTF-8, fixed column order, CRLF record endings, and standard escaping for quotes, commas, and line breaks. Empty scalar cells mean unavailable values. The metric_value cell always contains canonical JSON, including null, quoted strings, and arrays. Nested timing, compatibility, provenance, and evidence cells likewise contain canonical JSON when available.
Compatibility and lineage
Every row carries an explicit compatibility object and an export-v1: grouping key. The key hashes that object's canonical bytes with SHA-256. It binds execution-semantics and measurement versions, timing scope, primary time, environment class, the complete execution policy, observed control requests and guarantees, and resource-accounting scope and completeness. Missing classification evidence binds the particular record ID so absence cannot establish compatibility between unrelated records.
execute_only remains distinct from prepare_and_execute and cold_end_to_end. Development evidence remains distinct from locked environments, and best-effort or unsupported controls remain distinct from enforced controls. The control summary reports the weakest recorded status; the full list retains each requested control and its scope. control_evidence distinguishes observed controls from declarations retained in preflight failure records. Declarations alone have unknown observed guarantees. Matching keys do not establish scientific comparability or equal hardware. Benchmark authors must also select compatible problems, datasets, parameters, machines, and measurement evidence before downstream analysis.
export.json contains the canonical selector, source candidate, attempt, and observation identities in deterministic order, recorded result and metric schema references, complete compatibility classes, export and tool versions, and the SHA-256 digest of the exact emitted table. The manifest does not hash itself. This local lineage receipt does not package schemas or upgrade the original guarantees; portable snapshots belong to the later snapshot slice.
Select exact identities with repeatable --id options:
metewand export --output-dir runs/example --id mw1-attempt-DIGEST \
--format csv --destination tables/selectedReplace the example ID with a complete ID from status or explain. Selection accepts candidate, applicability, logical or resolved run, logical or resolved attempt or observation slot, attempt, and observation IDs. Multiple selectors are unioned. Selecting an observation retains its producing attempt and that attempt's other observations; selecting any history retains its candidate's applicability decisions. Unknown selectors fail before writing outputs.
The export schemas begin at EXPORT_SCHEMA_VERSION = 1 and join the existing version-1 public catalog. Existing public schemas, identity transcripts, wire messages, execution semantics, and the run-store layout remain unchanged: PUBLIC_SCHEMA_VERSION = 1, IDENTITY_VERSION = 1, WIRE_PROTOCOL_VERSION = 1, EXECUTION_SEMANTICS_VERSION = 2, and RUN_STORE_SCHEMA_VERSION = 1. The canonical index vector lives beside the schemas in schemas/v1/fixtures/export-index-v1.jsonl.