Reference

The command line and the library share one code path, so they never diverge.

Commands

Command	Does	Exit
`tacit-deid inspect <input>`	the preflight: formats and JSON shapes, sizes against the cap, refusals at load, keep values that match nothing, speakers seen, skipped files; `--stage0` adds entity counts from the local detectors; `--survey` lists what they left in place that looks like an identifier, by shape; `--config F` reads the cap, the detectors, and `custom_identifiers`; `--json`	2 when anything would be refused
`tacit-deid run <input> --config F`	one transcript to an artifact (`--out`, `-` for stdout), or a batch to `--out-dir`	0 crossed, 2 refused, 1 (batch) any error
`tacit-deid schema`	prints the JSON Schema of the artifact; `--out PATH` writes it to a file	0
`tacit-deid send <artifact.json>`	re-validates and posts the artifact; prints the token and the artifact’s state (Tacit stores it and does nothing else); `--tokenized-transcript PATH` also sends that file as its own call after the artifact	0
`tacit-deid pull <token>`	fetches the state (scenario, regeneration, fidelity); `--regenerate` creates the scenario if needed and starts a regeneration, `--wait` polls	0
`tacit-deid roundtrip <input> --config F`	run, send, create the scenario, regenerate, pull, and grade both hops	0 clean, 1 any survival

Flags shared by inspect, run, and roundtrip:

Flag	Meaning
`--format {plaintext,utterances,jsonl,vtt,srt,csv,tsv,docx,pdf}`	the input format when the extension does not say; required for `-` (stdin)
`--encoding`	the file’s encoding, default `utf-8`; a UTF-16 or UTF-32 byte order mark is honoured on its own
`--recursive`	a directory: also the files in the folders below it
`--text-column`, `--ref-column`, `--keep-column`	CSV and TSV columns, when the header does not hold exactly one known name
`--shape NAME`	JSON: force a registered shape instead of the first that fits
`--records-key`, `--text-key`, `--speaker-key`, `--ref-key`	JSON: paths (`a.b[0].c[]`) that map a shape no row knows; `--text-key` is the one required
`--speaker-label VALUE=LABEL` (repeatable)	render a speaker value as a label, e.g. `0=Doctor`
`--keep VALUE` (repeatable), `--keep-file PATH`	values that must survive verbatim

--shape takes one of utterance-list, utterances, transcript-utterances, messages, deepgram-utterances, deepgram-transcript, aws-transcribe-segments, aws-transcribe-transcript, azure-speech, google-stt, whisper, rev-ai, text. run and roundtrip also take --runs-dir (default runs/), --no-trace, and --env-file PATH (a KEY=VALUE file of credentials loaded first, never overriding a variable already set); run takes --tokenized-out (where the tokenized transcript goes when send_tokenized_transcript is on). send, pull, and roundtrip take --key-file and --tacit-url. pull and roundtrip take --max-turns, --timeout, --interval, and --stall. roundtrip takes --planted FILE (the path to a JSON file listing the case’s own identifiers, as plain strings or as {"type": "direct", "category", "value"} entries) and --out (the full verdict, owner-only).

Detectors on the command line

run, roundtrip, and inspect --stage0 or --survey build the stack named by detectors in the config, [presidio, gliner-pii-open] when it is unset (inspect without --config uses the same default). A GLiNER member without the gliner extra stops the command with the install line before any transcript is processed. The first use of gliner-pii-open downloads about 0.7 GB unless TACIT_DEID_GLINER_PII_OPEN_MODEL names a local directory. A run or roundtrip whose stack has no model member prints one warning to stderr that names the recall gap. See Configuration.

Library

Everything below is importable from tacit_deid.

Pipeline

Name	Role
`deidentify(text, deps, config, *, keep=(), traces=None) -> DeidResult`	the engine
`DeidResult`	`crossed`, `artifact`, `refusal_reason`, `audit_log`, `ledger`, `tokenized_transcript`, `coverage`
`EngineConfig`	`boundary_k`, `prevalence_floor`, `max_generalization_loops`, `mattr_window`, `max_transcript_chars`
`EngineDeps`	`detector`, `extractor_client`, `judges`, `provider`, `extractor_model`
`build_detector(names, custom=())`, `CompositeDetector`, `GLinerDetector`	the Stage-0 registry and its members (`presidio`, `gliner-pii-open`, `gliner-pii`, `gliner`)
`LLMJudge`, `LiteLLMClient`, `LLMConfig`	the judge and the gated model client
`create_base_rate_provider()`	the base-rate tables for the rarity controls
`TransformationLedger`, `TraceCollector`	the local record and the model-call sink
`TranscriptTooLargeError`, `TruncatedCompletionError`, `EndpointNotAllowedError`	raised before anything crosses

Inputs

Name	Role
`load_inputs(source, *, format=None, encoding="utf-8", options=...) -> list[TranscriptInput]`	one source (a `Source` or a path) into records
`input_sources(path, *, recursive=False) -> Sources`	`.sources` (a file, the members of a `.zip`, the files of a directory, or stdin for `-`) and `.skipped`
`Source`, `file_source(path)`	a name, a suffix, and a `read()` of the bytes
`TranscriptInput`	`text`, `ref`, `keep`, `source`, `shape`, `speakers`
`LoadOptions`	the column names, a forced or custom `JsonShape`, and the speaker labels
`JSON_SHAPES`, `JsonShape`, `custom_shape(...)`, `record_from_json(obj, *, source, shape, speaker_labels)`	the shape registry and one JSON value into one record
`TranscriptInputError`, `TranscriptEncodingError`, `EmptyTranscriptError`	refused at load
`inspect_inputs(sources, *, cap, skipped, keep, detector) -> PreflightReport`	the preflight
`run_batch(records, deps, config, *, keep_for, traces_for, on_outcome) -> list[BatchOutcome]`	many records through one set of deps; each `BatchOutcome` is one record’s result or error

Contract

Name	Role
`BoundaryArtifact`	the wire contract, schema 1.2
`seal_artifact(...)`, `verify_artifact_integrity(artifact)`	build with a digest; re-check one
`mint_artifact_token()`	`art_` plus 32 hex characters
`export_json_schema()`	the JSON Schema

Verification

Name	Role
`run_canary`, `check_artifact`, `default_canaries`, `Canary`, `CanaryResult`	the canary tripwire
`redetect_artifact`, `RedetectResult`	re-detection over the beats
`l_diversity`, `LDiversityResult`	the batch attribute-inference check
`clinical_rarity_flags`, `RarityFlag`	confirmed-rare conditions in beats
`export_defensibility_report(...)`	the per-batch pack
`qualify_model`, `QualificationResult`, `QualificationCheck`	the three-check model battery

Tacit transport

Name	Role
`TacitClient`, `TacitTransportError`	send, start a regeneration, fetch, wait
`run_roundtrip`, `roundtrip_oracle`, `RoundTripReport`, `RoundTripVerdict`	both hops graded on the client

Configuration

Name	Role
`load_config(path) -> SdkConfig`	the YAML file
`SdkConfig`	every field on the Configuration page; `engine_config()`, `llm_config()`

Getting started

How it works

Operating it

Reference

Reference

Reference

Commands

Detectors on the command line

Library

Pipeline

Inputs

Contract

Verification

Tacit transport

Configuration

​Reference

​Commands

​Detectors on the command line

​Library

​Pipeline

​Inputs

​Contract

​Verification

​Tacit transport

​Configuration

Reference

Commands

Detectors on the command line

Library

Pipeline

Inputs

Contract

Verification

Tacit transport

Configuration