Refresh architecture and configuration documentation for current runtime behavior

This commit is contained in:
2026-05-23 18:24:06 +00:00
parent 56f9b28f4b
commit f790c1441c
8 changed files with 467 additions and 1271 deletions

14
docs/architecture.md Normal file
View File

@@ -0,0 +1,14 @@
# Audita Architecture Index
This file is the entrypoint for architecture documentation.
Core architecture overview:
- [Architecture Overview](./architecture/architecture.md)
Focused architecture contracts:
- [Public Contract](./architecture/public-contract.md)
- [Diagnostics](./architecture/diagnostics.md)
- [Structured LLM](./architecture/structured-llm.md)
- [Validators](./architecture/validators.md)
- [Prompts](./architecture/prompts.md)
- [Output Schemas](./architecture/output-schemas.md)

View File

@@ -1,861 +1,153 @@
# Audita Architecture # Audita Architecture
## Scope and intent ## Scope
This document describes: This document describes the production architecture implemented in this repository today.
- the architecture used in production today.
Historical rewrite details live in `docs/rewrite-notes.md`. Audita is a single-process Go CLI that:
- loads effective runtime configuration;
- reads transcript and glossary inputs;
- normalizes and sections transcripts;
- runs a built-in module pipeline with validator chains;
- writes transcript output and run diagnostics.
## Current implementation status ## Runtime entrypoints
Implemented today: Primary CLI commands:
- Go CLI entrypoint and `audita process` wiring. - `audita process <transcript.json> --glossary <glossary.yaml> [flags]`
- Config defaults, env loading, CLI override precedence, and validation. - `audita config validate --config <config.yml>`
- Transcript and glossary parsing/validation. - `audita config print-effective [--config <config.yml>]`
- Deterministic transcript normalization.
- Deterministic token estimation and transcript chunking.
- Per-run diagnostics directory creation plus process-level artifacts.
- Process report JSON output with diagnostics artifact references.
- Framework foundation packages for contracts and proposal application.
- Production runner orchestration package with deterministic sequential module execution.
- Module-level report structures with applied/skipped change records.
- Runtime validator models and deterministic validators.
- Deterministic validator-chain execution in the runner with cardinality enforcement.
- Module-level validator decision/rejection reporting.
- Internal structured LLM client contract plus an Audita-owned OpenAI-compatible structured LLM adapter package.
- Bounded FIFO LLM scheduler infrastructure with context-aware permit handling.
- Runtime primary/validation LLM effective-config resolution helpers with validation inheritance.
- Generic JSON prompt/response diagnostics writer primitives with secret redaction.
- LLM-backed validator models, prompt builders, batching, and runtime execution.
- Runner wiring for LLM validators via the internal structured LLM abstraction and scheduler hooks.
- LLM validator diagnostics artifacts and report-level decision metadata paths.
- Shared LLM proposal-generation helper with structured correction-set parsing.
- Deterministic proposal-index assignment and enriched proposal mapping for shared generation.
- Proposal-generation diagnostics artifacts with secret redaction.
- Production module registry with known-key recognition and explicit unsupported-module errors.
- Production `grammar` module implementation in `internal/modules/grammar`.
- Production `glossary` module implementation in `internal/modules/glossary`.
- Production `homophones` module implementation in `internal/modules/homophones`.
- Production `spoken_word` module implementation in `internal/modules/spoken_word`.
- Explicit runtime support for `--modules grammar` through the production runner path.
- Explicit runtime support for `--modules glossary`, including repeated stages such as `--modules glossary,glossary`.
- Explicit runtime support for `--modules homophones` through the production runner path.
- Explicit runtime support for `--modules spoken_word` through the production runner path.
Current reality: Command ownership lives in `internal/cli/run.go`.
- all production modules exist and are wired into the default runtime path.
- a normal `audita process` run without `--modules` now executes the full sequence:
- `glossary`
- `homophones`
- `glossary`
- `spoken_word`
- `grammar`
- repeated glossary stages are deterministic and reported distinctly as `glossary_1` and `glossary_2`.
## Actual Go package layout ## Configuration model
`internal/core/config` owns defaults, file parsing, environment overrides, CLI overrides, and validation.
```text Effective-config loading for `process` and `config print-effective` is centralized in:
cmd/audita/ - `ResolveConfigPath`
main.go - `LoadEffectiveConfig`
internal/cli/ Effective precedence for `audita process`:
run.go 1. defaults
2. config file
internal/core/config/
config.go
env.go
flags.go
redaction.go
validation.go
internal/core/schema/
transcript.go
glossary.go
errors.go
internal/core/io/
files.go
internal/core/normalization/
normalize.go
tokens.go
internal/core/chunking/
sections.go
summary.go
tokens.go
internal/core/diagnostics/
run_dir.go
internal/core/reporting/
report.go
internal/framework/contracts/
contracts.go
internal/framework/proposals/
proposal.go
policy.go
preview.go
apply.go
internal/framework/runner/
observability.go
runner.go
internal/framework/proposal_generation/
generate.go
internal/framework/modules/
registry.go
internal/modules/grammar/
module.go
prompt.go
internal/modules/glossary/
module.go
prompt.go
internal/modules/homophones/
module.go
prompt.go
internal/modules/spoken_word/
module.go
prompt.go
internal/framework/validators/
models.go
deterministic.go
llm_models.go
llm_prompt_builders.go
llm_batching.go
llm_validators.go
internal/framework/warnings/
warnings.go
internal/validators/
metadata/
metadata.go
registry.go
chains.go
proposal_shape/
validator.go
confidence_threshold/
validator.go
original_text_presence/
validator.go
non_empty_corrected_text/
validator.go
no_effect/
validator.go
protected_terms/
validator.go
spoken_form_plausibility/
validator.go
meaning_reversal_review/
validator.go
editorial_review/
validator.go
grammar_review/
validator.go
spoken_word_review/
validator.go
internal/prompts/
registry.go
render.go
assets/
shared/
modules/
validators/
internal/framework/llm/
openai_compatible_client.go
scheduler.go
effective_config.go
diagnostics.go
internal/framework/responseschema/
registry.go
registry_test.go
internal/cli/
review_artifacts.go
parity_test.go
release_fixtures_test.go
testdata/
parity/
release/
```
## Current CLI behavior
Primary commands:
```sh
audita process <transcript.json> --glossary <glossary.yaml> [flags]
audita config validate --config <config.yml>
audita config print-effective [--config <config.yml>]
```
Current runtime flow (`internal/cli/run.go`):
1. Build runtime config from:
- defaults;
- file config source (`--config`, `AUDITA_CONFIG`, or default search paths when present: `/usr/local/etc/audita/config.yml`, then `/etc/audita/config.yml`);
- environment overrides;
- CLI overrides.
2. Parse flags and apply CLI overrides.
3. Validate transcript positional argument and required `--glossary`.
4. Create per-run diagnostics directory.
5. Read transcript and glossary files.
6. Parse/validate transcript and glossary.
7. Write source transcript artifacts.
8. Normalize transcript.
9. Write normalized transcript and normalization summary artifacts.
10. Chunk normalized transcript and compute chunk summaries.
11. Write chunking summary artifact.
12. Execute runner modules sequentially:
- default run path uses configured default sequence (`glossary,homophones,glossary,spoken_word,grammar`);
- explicit `--modules` overrides the default sequence;
- test/injected module factory path remains available for deterministic runtime tests.
- each module recomputes chunks from the current working transcript, runs chunk proposal work concurrently, aggregates deterministically, validates, and applies approved proposals once.
13. Output working transcript to `--output` file or stdout.
14. Build process report metadata.
15. Optionally write `--report-json`; always write run-dir `report.json`.
16. Apply work-dir retention.
Config command behavior (`internal/cli/run.go`):
- `audita config validate --config <path>`:
- loads and validates a versioned YAML config file;
- does not require transcript or glossary inputs.
- `audita config print-effective [--config <path>]`:
- builds effective config from defaults + file config + env overrides;
- prints redacted JSON to stdout;
- does not require transcript or glossary inputs.
Parity fixture status:
- representative Python-parity fixture coverage exists under `internal/cli/testdata/parity`;
- parity tests use fake structured LLM responses for deterministic behavior, including default full-pipeline shape assertions;
- parity comparisons intentionally ignore nondeterministic metadata (timestamps, run IDs, temp paths, token usage) and remain strict for deterministic contract fields (transcript content, module order/instance naming, applied/skipped/rejected counts, and status).
- intentional Python-vs-Go differences and open parity gaps are documented in `docs/python-parity.md`.
Important behavior details:
- Glossary is validated and is used for explicit glossary/grammar/homophones/spoken_word module correction paths.
- Default production CLI behavior now executes the full production module sequence unless `--modules` override is supplied.
- Explicit `--modules grammar`, `--modules glossary`, `--modules homophones`, and `--modules spoken_word` continue to run production module paths with LLM-backed proposal generation and validator-chain execution.
- Default runs (without explicit module selection) perform LLM calls through production module and validator paths.
- Success path is generally quiet on stderr.
- Malformed module-stage LLM payloads degrade to validator rejections and module warnings instead of aborting the run.
- Source IDs are preserved into a canonical transcript before normalization; normalization then reassigns output IDs sequentially from `1`.
## Implemented data contracts
### Transcript input
Accepted top-level forms:
- bare JSON array of segments
- object with `segments` array
Source segment contract:
- `id` optional integer
- `speaker` non-empty string
- `start` finite non-negative number
- `end` finite non-negative number with `end >= start`
- `text` non-empty string
- `categories` optional array of non-empty strings
Additional checks:
- duplicate explicit source IDs are rejected.
### Transcript output
Transcript output is selected through an output schema registry (`internal/core/outputschema`).
Supported output schemas:
- `bare-segments` (default):
- top-level JSON array of normalized segments;
- each segment includes `id`, `speaker`, `start`, `end`, `text`, optional `categories`.
- `audita-v1`:
- top-level object with:
- `schema: "audita-v1"`
- `version: "v1"`
- `segments: [...]` (same normalized segment payload).
Current status:
- `seriatim-intermediate` is not implemented yet; selecting it fails clearly as an unsupported output schema.
Selection behavior:
- CLI: `--output-schema <name>`
- file config: `output.schema: <name>`
- precedence remains runtime-wide defaults -> file config -> env -> CLI.
Both stdout transcript output and `--output` file output use the same selected output encoder.
### Glossary input
YAML with `glossary` entries. Required fields per entry:
- `name`, `category`, `summary`
Optional:
- `aliases`, `plural`
## Implemented config/env/flag behavior
Precedence for `audita process`:
1. defaults (`config.Default()`)
2. config file (if resolved from `--config`, `AUDITA_CONFIG`, or default path)
3. environment overrides 3. environment overrides
4. CLI flags (`ApplyCLIOverrides`) 4. CLI overrides
File-config source behavior: `audita config validate` is intentionally file-only validation:
- explicit `--config <path>`: - load versioned file;
- required to exist, otherwise process fails clearly. - apply onto defaults;
- `AUDITA_CONFIG` (when `--config` is not provided): - validate;
- required to exist, otherwise process fails clearly. - do not apply environment overrides.
- default paths `/usr/local/etc/audita/config.yml`, then `/etc/audita/config.yml` (when neither explicit source is provided):
- first existing path in that order is used;
- both missing is silently ignored.
Versioned file-config behavior (`internal/core/config/file_config.go`): Supported module and output-schema keys are validated through shared catalogs:
- supported version: `version: 1`; - module keys: `internal/core/modulecatalog`
- missing version fails; - output schemas: `internal/core/outputschema`
- unsupported version fails;
- strict unknown-field rejection is enabled.
`api_key_env` behavior: ## Pipeline and module orchestration
- file config can declare API key environment variable names for proposal/validation LLM settings; The built-in module sequence is configured in runtime config and executed by `internal/framework/runner` through resolved module specs.
- runtime resolves those names from the process environment during config application;
- no direct API-key value field is supported in file config.
Redaction behavior: Current default sequence:
- effective config artifacts and `audita config print-effective` both use the same redaction path (`Config.Redacted()`), so API keys are not emitted in plaintext.
Implemented config surfaces include:
- module list
- primary and validation LLM settings
- total/proposal/validation LLM concurrency controls
- transcript description context (`--transcript-description`)
- section token controls and target sections
- confidence thresholds
- normalization controls
- work-dir and retention mode
Current caveat:
- LLM/module-related settings are active for default and explicit module-run paths.
- compatibility environment variables and lower-level CLI tuning flags remain available while the preferred config-driven surface is adopted.
Transcript description behavior:
- `--transcript-description` is a process-flag input for optional user-supplied background context.
- runtime config stores this value in `Config.TranscriptDescription` after CLI trimming and length validation.
- default value is empty; empty values produce no prompt context section.
- this value is intentionally non-secret and appears in effective config and invocation metadata artifacts.
## Implemented transcript description prompt context
Transcript description context is wired through production prompt paths:
- proposal prompts for `glossary`, `homophones`, `spoken_word`, and `grammar`;
- LLM-backed validator prompts for spoken-form plausibility, meaning reversal, editorial review, grammar review, and spoken-word review.
Prompt guardrail semantics are consistent across modules and validators:
- transcript description is labeled as "background context only";
- it may help interpret ambiguous terms;
- it must not override transcript content;
- the model must not invent corrections, facts, names, events, motivations, or speaker intent from this description.
Generated transcript descriptions remain deferred and are not implemented in the current runtime.
## Implemented embedded prompt assets
Prompt assets are now built-in embedded Markdown files under `internal/prompts/assets`:
- `assets/modules/*` for production module proposal prompts;
- `assets/validators/*` for LLM-backed validator prompts;
- `assets/shared/prompt_hardening.md` for shared prompt-injection hardening text.
Prompt source behavior:
- built-in embedded prompts are the only supported source in current runtime;
- filesystem prompt overrides and prompt-source selection flags are not implemented.
`internal/prompts` registry responsibilities:
- register stable prompt IDs;
- register prompt version and source metadata;
- load embedded assets;
- compute deterministic SHA-256 prompt source hashes;
- render system/user prompts with `text/template` using missing-key errors.
Prompt metadata fields:
- `prompt_id`
- `prompt_version`
- `prompt_source` (`builtin`)
- `embedded_path`
- `sha256`
Prompt rendering flow:
- module proposal builders construct typed template data (section JSON, glossary JSON, transcript-description block) and render via `internal/prompts`;
- validator prompt builders construct typed template data (validation payload JSON, transcript-description block) and render via `internal/prompts`.
Shared prompt hardening:
- the same centralized hardening fragment is included in every proposal and LLM-validator prompt;
- hardening text enforces untrusted transcript handling, no instruction-following from transcript content, and no invented facts/corrections.
Prompt metadata diagnostics flow:
- proposal-generation diagnostics request metadata includes prompt metadata;
- LLM-validator diagnostics request metadata includes prompt metadata;
- detailed prompt metadata is diagnostics-scoped today and is not yet expanded into broad report-level prompt registries.
## Implemented structured LLM infrastructure
`internal/framework/contracts` now defines a typed structured-completion contract:
- `StructuredLLMClient.CompleteStructured(ctx, req, out)`
- caller-owned typed decode target via `out` pointer.
- caller-selected structured response schema metadata via `StructuredCompletionRequest.ResponseSchema`.
`internal/framework/llm` provides `OpenAICompatibleClient`, a direct `net/http` adapter over OpenAI-compatible chat completions:
- configurable `base_url`, model, optional API key, retries, HTTP client, and request timeout;
- OpenAI-compatible endpoint behavior (for example OpenAI/OpenRouter/local-compatible base URLs);
- request message translation from `contracts.LLMMessage` to chat-completions messages;
- strict `response_format.type = json_schema` with registered structured response schemas (`strict: true`, schema name, and schema body);
- response metadata mapping (provider/model/token usage) into Audita-owned response types;
- API-key redaction in adapter-returned errors;
- context cancellation and timeout propagation through request contexts and HTTP client timeouts;
- bounded retry behavior for transient request failures and malformed retryable structured responses.
Structured response schemas are owned by Audita in `internal/framework/responseschema` and currently include:
- key `correction_set`:
- id `audita.correction_set`
- version `v1`
- name `audita_correction_set_v1`
- sha256 `05f8ff3fa04f68115c0cb1859d2656f51aa5c0bae8ff2470b2d4f6f531953195`
- key `validator_decision_set`:
- id `audita.validator_decision_set`
- version `v1`
- name `audita_validator_decision_set_v1`
- sha256 `b73f4790b98fbb955f0aec5496dd8ce9a8fe14aa2f35c700b4b4e5634f106fd5`
Provider-level structured output is treated as a guardrail, not a trust boundary:
- the adapter decodes assistant message content into caller-owned structs;
- proposal-generation and validator layers continue local validation (shape, cardinality, confidence bounds, and proposal-index semantics) before changes can be applied.
Current runtime boundary:
- the default CLI runtime path (without explicit module selection) instantiates the full production module sequence.
- LLM calls are exercised in production in both default full-pipeline runs and explicit `--modules` runs, and in tests when fake/injected clients are used.
- normal `go test ./...` does not require real LLM credentials or Python dependencies.
`internal/framework/llm` also provides:
- a bounded FIFO `Scheduler` for controlled concurrent LLM calls with reliable permit release on success, error, and cancellation;
- primary/validation effective-config resolution helpers, including validation inheritance fallback to total LLM concurrency settings;
- generic interaction diagnostics primitives that write machine-readable JSON artifacts for request metadata, request payload, response payload, and optional error payload with secret redaction.
Structured LLM diagnostics behavior:
- proposal-generation and validator diagnostics include structured response schema metadata (`id`, `version`, `name`, `sha256`) when schema-driven calls are made;
- API keys and bearer tokens are redacted from request/response/error diagnostics artifacts and surfaced errors.
Dependency posture:
- the runtime no longer depends on `instructor-go`;
- structured LLM behavior is implemented through Audita-owned code paths behind `StructuredLLMClient`.
LLM concurrency runtime behavior:
- `total` concurrency bounds all proposal and validation LLM calls.
- `proposal` concurrency adds a proposal-only sub-cap, composed with total.
- `validation` concurrency adds a validation-only sub-cap, composed with total.
- legacy `llm-concurrency` inputs remain compatibility aliases for total concurrency.
- modules execute serially, chunk proposals run concurrently within each module, and approved proposals are applied once per module in deterministic order.
## Implemented normalization behavior
Normalization (`internal/core/normalization`) currently:
- sorts by segment start time;
- merges adjacent same-speaker segments when constraints pass;
- uses gap-based joiners:
- gap `< ellipsis_gap` -> single space join
- gap `>= ellipsis_gap` -> `... ` join
- enforces merged duration and token-limit constraints;
- reassigns output IDs sequentially from `1`;
- returns `NormalizationSummary` with merge and skip counters.
Note: merged categories are concatenated (not deduplicated).
## Implemented chunking behavior
Chunking (`internal/core/chunking`) currently provides:
- deterministic heuristic token estimation;
- contiguous sectioning with section metadata;
- max/min section token validation;
- optional `target_sections` override for section-count planning;
- summary and detailed summary generation.
Current behavior details:
- if a single segment exceeds max tokens, it is emitted as its own section (not hard-failed);
- default section count is planned from `ceil(total_tokens / max_section_tokens)`;
- section sizing targets `ceil(total_tokens / section_count)` with a deterministic forward pass;
- sections remain contiguous and ordered, and segments are never split.
## Implemented proposal/replacement infrastructure
`internal/framework/proposals` provides deterministic proposal composition logic:
- `CorrectionProposal` and `EnrichedCorrectionProposal` models;
- replacement policies: `require_unique`, `replace_all`;
- safe preview (`PreviewProposalForSegment`) with stable skip reasons;
- deterministic apply (`ApplyProposals`) in ascending `proposal_index` order;
- applied/skipped change records suitable for reporting.
`internal/framework/contracts` provides interfaces and run-spec metadata scaffolding, including deterministic repeated module instance naming (`ResolveModuleRunSpecs`).
These primitives are wired into the production runner and report model. The grammar, glossary, homophones, and spoken_word modules are implemented.
## Implemented validator runtime infrastructure
`internal/framework/validators` provides deterministic validator infrastructure:
- runtime validation request/result models;
- stable validator reason codes;
- cardinality enforcement for validator decisions:
- missing proposal indexes fail
- duplicate proposal indexes fail
- unknown proposal indexes fail
- deterministic validators:
- confidence threshold by module key/config threshold
- original-text presence against current working transcript
- non-empty corrected text
- identical/no-effect rejection
- conservative protected glossary-term guard for non-glossary modules
`internal/framework/runner` executes module pipelines with deterministic boundaries:
- modules still execute serially over the working transcript;
- section proposal work is launched promptly and can run concurrently;
- section-level validator-chain work starts as section proposals become available (deterministic validators before LLM-backed validators);
- proposal-generation and LLM-validator calls can overlap under composed scheduler limits;
- approved proposals are still applied once per module after section work settles.
Validator rejections are reported distinctly from proposal-application skips.
Validator composition is now explicit and registry-backed through `internal/validators`:
- built-in validator registry with stable keys and lookup/build failure for unknown keys;
- built-in chain definitions per production module key;
- production modules resolve validator chains from those built-in definitions.
Package ownership boundary:
- `internal/validators/<validator_key>` owns built-in validator construction and stable key identity.
- `internal/framework/validators` remains shared runtime machinery:
- request/result models;
- decision cardinality enforcement;
- protected-vocabulary helpers;
- generic LLM-backed validator runtime, batching, and diagnostics glue.
Validator execution classification metadata:
- `internal/validators/metadata` defines execution class markers:
- `deterministic`
- `llm_backed`
- runner ordering uses this metadata interface rather than concrete framework validator type assertions.
- validators without classification metadata default to deterministic ordering.
`protected_terms` construction ownership:
- `internal/validators/protected_terms.New()` builds the general (non-glossary-stage) variant.
- `internal/validators/protected_terms.NewGlossaryStage()` builds the glossary-stage variant used by glossary chains.
- both variants preserve the stable key `protected_terms`.
Stable built-in validator keys:
- deterministic:
- `proposal_shape`
- `confidence_threshold`
- `original_text_presence`
- `non_empty_corrected_text` (historical key name; current semantics reject empty resulting segment text)
- `no_effect`
- `protected_terms`
- LLM-backed:
- `spoken_form_plausibility`
- `meaning_reversal_review`
- `editorial_review`
Built-in module chains:
- `glossary`:
- `proposal_shape`
- `no_effect`
- `original_text_presence`
- `confidence_threshold`
- `protected_terms`
- `non_empty_corrected_text`
- `spoken_form_plausibility`
- `meaning_reversal_review`
- `homophones`:
- `proposal_shape`
- `no_effect`
- `original_text_presence`
- `confidence_threshold`
- `protected_terms`
- `non_empty_corrected_text`
- `spoken_form_plausibility`
- `meaning_reversal_review`
- `spoken_word`:
- `proposal_shape`
- `no_effect`
- `original_text_presence`
- `confidence_threshold`
- `protected_terms`
- `non_empty_corrected_text`
- `editorial_review`
- `meaning_reversal_review`
- `grammar`:
- `proposal_shape`
- `no_effect`
- `original_text_presence`
- `confidence_threshold`
- `protected_terms`
- `non_empty_corrected_text`
- `editorial_review`
- `meaning_reversal_review`
1.0 boundary:
- validator chains are built-in and not user-configurable from config/CLI.
- existing threshold and batching knobs remain configurable.
## Implemented LLM-backed validator infrastructure
`internal/framework/validators` now includes LLM-backed validator support:
- typed request/response models for structured LLM validation;
- prompt builders for:
- spoken-form plausibility
- meaning reversal detection
- editorial review
- grammar review
- spoken-word review
- deterministic batching by `validation_max_prompt_tokens`;
- strict cardinality validation of synthesized validator decision sets;
- malformed validator payloads reject only the affected batch with warnings;
- oversized single-proposal validator inputs reject only the affected proposal;
- transport/provider/runtime LLM call failures remain fatal.
`internal/framework/runner` wires LLM validators into existing validator chains using:
- the internal structured LLM client abstraction (`contracts.StructuredLLMClient`);
- bounded scheduler hooks for validator call execution;
- diagnostics writer hooks for machine-readable prompt/response artifacts with secret redaction.
## Implemented shared proposal-generation infrastructure
`internal/framework/proposal_generation` provides a reusable, prompt-agnostic helper for future real modules:
- structured request model including module key/instance, replacement policy, working transcript context, optional section metadata, glossary, config, and diagnostics context;
- structured correction-set response model (`corrections`) mapped into existing `proposals.CorrectionProposal` and `proposals.EnrichedCorrectionProposal` models;
- deterministic proposal-index assignment through a caller-provided `start_index`;
- structured LLM calls through `contracts.StructuredLLMClient` only (no direct provider calls);
- optional bounded execution through scheduler hooks (`contracts.LLMScheduler`);
- prompt/response diagnostics artifact writing via the generic `internal/framework/llm` diagnostics primitives with redaction of API keys/secrets.
This helper only produces candidate proposals; validator-chain execution and proposal application remain runner responsibilities.
## Implemented production module-registry scaffolding
`internal/framework/modules` now provides a production registry scaffold:
- recognizes intended module keys:
- `glossary` - `glossary`
- `homophones` - `homophones`
- `glossary`
- `spoken_word` - `spoken_word`
- `grammar` - `grammar`
- supports explicit constructor registration with dependency injection for:
- run spec
- config
- glossary
- proposal/validation structured LLM clients
- proposal/validation schedulers
- diagnostics directory context
- returns explicit errors for unknown keys (`unsupported_module`).
The `grammar`, `glossary`, `homophones`, and `spoken_word` module keys are now registered and constructible. Execution behavior:
- modules execute serially over the working transcript;
- section proposal work can run concurrently within a module;
- validator execution happens on generated proposals before application;
- approved proposals are applied once per module in deterministic proposal-index order.
## Implemented grammar production module Production modules remain separate packages:
`internal/modules/grammar` now provides the first production module: - `internal/modules/glossary`
- prompt builder faithfully constrained to punctuation/capitalization/spacing/article cleanup; - `internal/modules/homophones`
- explicit guardrails against meaning-changing rewrites, style rewrites, summarization, and invention; - `internal/modules/spoken_word`
- proposal generation through `internal/framework/proposal_generation` and `contracts.StructuredLLMClient`; - `internal/modules/grammar`
- scheduler-aware proposal calls through existing `contracts.LLMScheduler` hooks;
- replacement policy `require_unique` (current runtime policy);
- validator chain integration using existing deterministic + LLM-backed validators;
- grammar confidence threshold enforcement through existing validator/config infrastructure;
- module-level reporting and diagnostics capture through existing runner/reporting paths.
## Implemented glossary production module ## Proposal generation and prompt context
`internal/modules/glossary` now provides the second production module: Shared proposal plumbing is centralized in `internal/framework/proposal_generation`.
- prompt builder aligned to Python glossary-module intent, constrained to glossary-backed domain/acoustic corrections;
- prompt context includes glossary names, aliases, categories, summaries, and plural forms where available;
- guardrails against broad style rewriting and against replacing unrelated terms simply because they appear in glossary entries;
- proposal generation through `internal/framework/proposal_generation` and `contracts.StructuredLLMClient`;
- scheduler-aware proposal calls through existing `contracts.LLMScheduler` hooks;
- replacement policy `replace_all` (matching Python glossary behavior);
- validator chain integration using existing deterministic + LLM-backed validators;
- glossary confidence threshold enforcement through existing validator/config infrastructure;
- module-level reporting and diagnostics capture through existing runner/reporting paths;
- explicit support for repeated glossary stages with deterministic instance names (`glossary_1`, `glossary_2`, ...), where later stages see prior-stage working transcript changes.
## Implemented protected-term behavior Module packages provide:
`internal/framework/validators/protected_terms.go` provides deterministic glossary-derived protected vocabulary: - module identity and replacement policy;
- extracts protected terms from glossary names and aliases; - module-specific prompt message building;
- includes explicit plural fields and synthetic plural forms where safe; - built-in validator chain selection.
- deduplicates and returns stable ordering for repeatable behavior/tests.
This vocabulary is used by deterministic validators for both glossary-stage and non-glossary-stage protection checks, keeping protected-term guardrails active across modules. Shared prompt payload helpers are in `internal/framework/promptcontext`.
## Implemented homophones production module ## Validator architecture
`internal/modules/homophones` now provides the third production module: Built-in validator construction and chain composition live in `internal/validators`.
- prompt builder aligned to Python homophones-module intent, constrained to conservative homophone/near-homophone/mistranscription corrections;
- prompt context includes protected glossary names/aliases/plurals to avoid damaging known terms;
- explicit guardrails against punctuation cleanup, grammar cleanup, style rewriting, summarization, and content invention;
- proposal generation through `internal/framework/proposal_generation` and `contracts.StructuredLLMClient`;
- scheduler-aware proposal calls through existing `contracts.LLMScheduler` hooks;
- replacement policy `require_unique` (matching Python homophones behavior);
- validator chain integration using existing deterministic + LLM-backed validators;
- homophones confidence threshold enforcement through existing validator/config infrastructure;
- protected-term guardrails for non-glossary modules remain active and are exercised through the homophones path;
- module-level reporting and diagnostics capture through existing runner/reporting paths.
## Implemented spoken_word production module Shared validator runtime mechanics live in `internal/framework/validators`.
`internal/modules/spoken_word` now provides the fourth production module:
- prompt builder aligned to Python spoken_word-module intent, constrained to conservative dysfluency cleanup;
- strong prompt guardrails preserving meaning/intent/voice/named entities/domain terms and substantive content;
- explicit guardrails against summarization, style rewriting, grammar-only cleanup, punctuation-only cleanup, invention, and meaning-changing rewrites;
- proposal generation through `internal/framework/proposal_generation` and `contracts.StructuredLLMClient`;
- scheduler-aware proposal calls through existing `contracts.LLMScheduler` hooks;
- replacement policy `require_unique` (matching Python spoken_word behavior);
- validator chain integration using existing deterministic + LLM-backed validators, including strong semantic guardrails (`spoken_word_review`, `meaning_reversal_review`);
- spoken_word confidence threshold enforcement through existing validator/config infrastructure;
- protected-term guardrails for non-glossary modules remain active and are exercised through the spoken_word path;
- module-level reporting and diagnostics capture through existing runner/reporting paths.
## Reports and diagnostics (implemented) Execution class metadata (deterministic vs LLM-backed) is centralized in `internal/validators/metadata` and used for ordering and reporting classification.
Current per-run artifacts include:
- `source-transcript.json`
- `source-transcript-parsed.json`
- `normalized-transcript.json`
- `normalization-summary.json`
- `chunking-summary.json`
- `utilization-diagnostics.json`
- `correction-ledger.json`
- `invocation.json`
- `effective-config.json` (redacted credentials)
- `report.json`
- `error.log` on failure
`--report-json` writes a separate report file when requested. ## Structured LLM boundary
All production LLM calls go through the internal contract:
- `contracts.StructuredLLMClient`
- `CompleteStructured(ctx, req, out)`
Current process reports include diagnostics metadata references for: The OpenAI-compatible HTTP adapter is implemented in `internal/framework/llm`.
- diagnostics directory path;
- source transcript artifact path;
- parsed source transcript artifact path;
- normalized transcript artifact path;
- normalization summary artifact path;
- chunking summary artifact path;
- utilization diagnostics artifact path;
- correction ledger artifact path;
- invocation metadata artifact path;
- redacted effective-config artifact path;
- error-log artifact path on failure.
Current process reports also include: Structured response schemas are registered in `internal/framework/responseschema` and attached to requests via `response_format` metadata.
- module-level results (when runner modules execute), including applied/skipped proposal changes;
- run-level module summary totals and failed module instance metadata.
- module-level warning records for malformed proposal-generation payloads and malformed validator batches.
- module-level validator decisions and validator rejections.
- optional decision-level diagnostic artifact paths for validator LLM interactions when available.
- stable validator keys in `validator_name` fields for validator decisions/rejections.
- explicit report metadata:
- report schema name;
- report schema version;
- selected output schema;
- config file version when config file input is used.
- review/observability artifacts:
- run-level and module-level utilization/timing summaries;
- flattened correction ledger entries for applied/rejected/skipped/failed correction dispositions.
Utilization diagnostics collection: Malformed structured-output detection is centralized in `internal/framework/structuredoutput` and reused by proposal generation and validator execution so downgrade behavior stays consistent.
- collection is performed in the runner path via lightweight instrumentation around LLM scheduler and structured-client execution (`internal/framework/runner`);
- instrumentation is observational only and does not change scheduler acquisition/release semantics or module execution order;
- serialized artifact: `utilization-diagnostics.json`.
Utilization diagnostics high-level shape: ## Stage naming and diagnostics metadata
- `effective_concurrency`: Diagnostics stage naming is centralized in `internal/framework/stagename`:
- `total_llm`, `proposal_llm`, `validation_llm`; - module proposal stage names;
- `run_timing`: - proposal-generation stage names;
- `run_wall_time_ms`; - validator batch stage names.
- `scheduler_queue_wait_ms`;
- `llm_execution_time_ms`;
- `deterministic_validation_time_ms`;
- `max_in_flight_llm_calls`;
- `average_in_flight_llm_calls`;
- `llm_calls`:
- `total_proposal_calls`;
- `total_validation_calls`;
- `modules`:
- per-module key/instance timing summaries including module wall time and per-module call counts;
- `validators`:
- per-validator summaries keyed by stable validator key with elapsed time and LLM-backed marker.
Correction ledger construction: Prompt metadata and response-schema metadata each expose canonical diagnostics maps via:
- ledger entries are built from runner module results in the CLI report/diagnostics path (`internal/cli/review_artifacts.go`); - `prompts.Metadata.DiagnosticsMap()`
- serialized artifact: `correction-ledger.json`; - `responseschema.Schema.DiagnosticsMap()`
- one flattened record per applied/validator-rejected/application-skipped outcome where data is available, plus module-failed records for failed module instances.
Correction ledger high-level shape: ## Diagnostics and reporting
- run/module/proposal identity: Run-directory artifacts are owned by `internal/core/diagnostics`.
- `run_id`, `module_key`, `module_instance`, `proposal_index`, `segment_id`;
- correction payload:
- `original_text`, `proposed_corrected_text`, `applied_corrected_text` (when applied), `replacement_policy`;
- disposition:
- `disposition` in `{applied,rejected,skipped,failed}`;
- `disposition_reason_code`, `disposition_message`;
- validator decision snapshots:
- `deterministic_validator_decisions[]`;
- `llm_validator_decisions[]`;
- each decision uses stable validator keys and reason codes.
Identity and metadata boundaries: Stable artifact names are centralized constants (for example transcript artifacts, `invocation.json`, `effective-config.json`, `utilization-diagnostics.json`, `correction-ledger.json`, `report.json`, `error.log`).
- stable module keys/instance names and stable validator keys are included directly in ledger records;
- prompt metadata and structured response schema metadata remain in LLM interaction diagnostics payloads and are not duplicated into every ledger row;
- reports reference artifact paths for utilization and ledger files through diagnostics metadata.
Redaction and retention: Report diagnostics path metadata is constructed through `BuildDiagnosticsMetadata`, which keeps run-directory artifact references consistent between success and failure reports.
- secret redaction guarantees continue to apply to diagnostics/report artifacts;
- utilization and ledger artifacts are emitted within the existing run-directory retention model (`auto|always|never`) and are retained/removed with the run directory.
Current report schema metadata values: ## Secret redaction
- `report_metadata.report_schema_name = "audita-process-report"` Redaction responsibilities are split by concern:
- `report_metadata.report_schema_version = "v1"` - structural config redaction: `config.Config.Redacted()`
- byte/string payload redaction for diagnostics and surfaced errors: framework redaction utilities.
Retention modes implemented in `ApplyRetention`: Configured LLM secret extraction is centralized in `llm.ConfiguredSecrets(cfg)` and reused across proposal and validator diagnostics paths.
- `always`: keep all run directories.
- `never`: keep successful run directories.
- `auto`: keep failed runs and successful runs with skipped corrections.
- failed runs are always retained.
Current runtime note: ## Output contracts
- default non-explicit runs usually have no module-level skipped corrections, so `auto` commonly removes clean successful run directories. Transcript output schema selection is owned by `internal/core/outputschema`.
- explicit grammar/glossary/homophones/spoken_word runs can produce validator rejections and application skips, which are reflected in reports and retention input.
## Current tests and quality posture Supported schemas:
Implemented tests currently cover: - `bare-segments`
- CLI argument handling and behavior (`internal/cli/run_test.go`) - `audita-v1`
- subprocess stdout/stderr and exit-code behavior (`cmd/audita/main_integration_test.go`)
- config/env/override validation (`internal/core/config/*_test.go`)
- transcript and glossary schema validation (`internal/core/schema/*_test.go`)
- deterministic normalization (`internal/core/normalization/*_test.go`)
- deterministic chunking and summaries (`internal/core/chunking/*_test.go`)
- proposal preview/apply semantics (`internal/framework/proposals/*_test.go`)
- contracts/foundation composition tests (`internal/framework/contracts/*_test.go`)
- runner sequencing and failure behavior with deterministic fake modules (`internal/framework/runner/*_test.go`)
- CLI runner integration through injected fake module factories (`internal/cli/run_test.go`)
- validator models, cardinality enforcement, and deterministic validators (`internal/framework/validators/*_test.go`)
- LLM-backed validator batching, prompt builders, structured-response safety, scheduler hooks, and diagnostics redaction (`internal/framework/validators/*_test.go`, `internal/framework/runner/*_test.go`)
- shared proposal-generation request/response parsing, deterministic indexing, scheduler hooks, and diagnostics redaction (`internal/framework/proposal_generation/*_test.go`, `internal/framework/runner/*_test.go`)
- production module-registry known-key recognition and unsupported/internal-registry error behavior (`internal/framework/modules/*_test.go`, `internal/cli/run_test.go`)
- production grammar module prompt constraints, proposal mapping, validator-chain behavior, confidence-threshold enforcement, diagnostics redaction, and explicit CLI/runtime integration (`internal/modules/grammar/*_test.go`, `internal/cli/run_test.go`, `internal/framework/runner/*_test.go`)
- production glossary module prompt constraints, proposal mapping, validator-chain behavior, confidence-threshold enforcement, diagnostics redaction, repeated-stage behavior, and explicit CLI/runtime integration (`internal/modules/glossary/*_test.go`, `internal/cli/run_test.go`, `internal/framework/runner/*_test.go`)
- production homophones module prompt constraints, proposal mapping, validator-chain behavior, confidence-threshold enforcement, diagnostics redaction, protected-term behavior, and explicit CLI/runtime integration (`internal/modules/homophones/*_test.go`, `internal/cli/run_test.go`, `internal/framework/runner/*_test.go`)
- production spoken_word module prompt constraints, proposal mapping, validator-chain behavior, semantic guardrail behavior, confidence-threshold enforcement, diagnostics redaction, protected-term behavior, and explicit CLI/runtime integration (`internal/modules/spoken_word/*_test.go`, `internal/cli/run_test.go`, `internal/framework/runner/*_test.go`)
- glossary-derived protected-term extraction and stable behavior (`internal/framework/validators/protected_terms_test.go`)
- default full-pipeline runtime shape and ordering (`internal/cli/run_test.go`, `cmd/audita/main_integration_test.go`, `internal/cli/parity_test.go`)
- subprocess operational hardening behavior including large-input, failure-mode, timeout/cancellation, backend-failure, and partial-progress paths (`cmd/audita/main_integration_test.go`)
- report/diagnostics redaction and artifact-shape behavior across success and failure paths (`internal/cli/run_test.go`, `cmd/audita/main_integration_test.go`)
- curated release-fixture and idempotence-oriented readiness checks using fake structured LLM responses (`internal/cli/release_fixtures_test.go`, `internal/cli/testdata/release`)
## Operational hardening status Unknown schema keys fail validation and runtime resolution.
The runtime now includes hardened subprocess behavior for parent-process callers:
- deterministic success/failure exit codes;
- strict stdout/stderr separation suitable for machine orchestration;
- failure stderr summaries that include diagnostics location when available;
- retained failure diagnostics (`report.json`, `error.log`, and artifacts written before failure);
- deterministic timeout/cancellation behavior in tests;
- redaction coverage for API keys/secrets across reports, diagnostics artifacts, and surfaced errors.
- stable output routing behavior:
- with `--output`, stdout remains empty on success;
- without `--output`, stdout contains only transcript JSON in the selected output schema;
- `--report-json` writes report data to file only (never stdout).
Operational caller guidance is documented in [`docs/subprocess-operations.md`](docs/subprocess-operations.md). ## Key package map
Core packages:
- `internal/core/config`
- `internal/core/schema`
- `internal/core/normalization`
- `internal/core/chunking`
- `internal/core/diagnostics`
- `internal/core/reporting`
- `internal/core/modulecatalog`
- `internal/core/outputschema`
## Final status Framework packages:
- Audita's default full module-sequence runtime is implemented and tested. - `internal/framework/contracts`
- Parity fixtures and operational hardening coverage are in place. - `internal/framework/proposals`
- Historical migration context is documented in [`docs/migration-from-python.md`](docs/migration-from-python.md). - `internal/framework/proposal_generation`
- `internal/framework/promptcontext`
- `internal/framework/runner`
- `internal/framework/validators`
- `internal/framework/llm`
- `internal/framework/responseschema`
- `internal/framework/stagename`
- `internal/framework/structuredoutput`
Domain packages:
- `internal/modules/*`
- `internal/validators/*`
- `internal/prompts`

View File

@@ -1,31 +1,25 @@
# Audita Public Contract # Audita Public Contract
This document defines stability expectations for Audita's external process and data interfaces.
## Scope ## Scope
This document defines stability expectations for Audita's external runtime interfaces.
This contract covers: Covered interfaces:
- CLI invocation and behavior - CLI commands and major flags;
- versioned config file behavior - versioned config behavior and precedence;
- transcript/glossary input forms - transcript/glossary input forms;
- transcript output schema selection - output schema selection;
- process report schema metadata - report schema metadata;
- stable validator key identifiers in report/diagnostics records - diagnostics artifact path metadata;
- prompt metadata identifiers in diagnostics - stdout/stderr and exit-code behavior;
- diagnostics directory behavior - redaction guarantees.
- utilization diagnostics and correction-ledger artifact presence/pathing in diagnostics metadata
- stdout/stderr and exit-code behavior
- secret redaction guarantees
- compatibility and deprecation policy
## CLI stability expectations
## CLI contract
Stable commands: Stable commands:
- `audita process` - `audita process`
- `audita config validate` - `audita config validate`
- `audita config print-effective` - `audita config print-effective`
For `audita process`, stable high-value flags include: Stable high-value `process` flags:
- `--config` - `--config`
- `--glossary` - `--glossary`
- `--output` - `--output`
@@ -33,136 +27,95 @@ For `audita process`, stable high-value flags include:
- `--modules` - `--modules`
- `--output-schema` - `--output-schema`
Compatibility flags and lower-level tuning flags remain available; they may be narrowed over time with explicit compatibility notes. ## Config contract
Supported config format:
- YAML;
- `version: 1`;
- strict unknown-field rejection.
## Config file stability expectations Path resolution for `process` and `config print-effective`:
1. `--config`
2. `AUDITA_CONFIG`
3. `/usr/local/etc/audita/config.yml`
4. `/etc/audita/config.yml`
Supported file format: Missing explicit path is an error. Missing default paths is non-fatal.
- YAML
- strict unknown-field rejection
- explicit `version`
Supported version: Precedence for `process`:
- `version: 1` 1. defaults
Precedence for `audita process`:
1. built-in defaults
2. file config 2. file config
3. environment overrides 3. environment overrides
4. CLI overrides 4. CLI overrides
Config source behavior: `config validate` remains file-only validation (defaults + file config; no env overrides).
- `--config <path>`: missing path is a clear failure
- `AUDITA_CONFIG`: missing path is a clear failure
- defaults `/usr/local/etc/audita/config.yml`, then `/etc/audita/config.yml`: both missing is non-fatal
## Supported transcript input forms Module and output-schema keys are validated against built-in catalogs. Unknown keys fail validation.
Audita accepts transcript JSON as either: ## Input contract
- a top-level array of segments Supported transcript JSON top-level forms:
- an object with a `segments` array - array of segments
- object with `segments` array
Segments must satisfy the schema and validation rules enforced by `internal/core/schema`. Supported glossary YAML form:
- top-level `glossary` list with required entry fields validated by schema parsing.
## Supported glossary input form ## Output schema contract
Supported transcript output schemas:
Audita accepts glossary YAML with a top-level `glossary` entry list and validates required fields per entry.
## Supported output schema names
Built-in output schema registry supports:
- `bare-segments` (default) - `bare-segments` (default)
- `audita-v1` - `audita-v1`
`seriatim-intermediate` is planned but not implemented. Unknown schema keys fail before output write.
Unknown output schema names fail clearly. ## Report metadata contract
Process reports include stable report metadata fields:
## Report schema/versioning expectations
Process report payloads include `report_metadata` with:
- `report_schema_name` - `report_schema_name`
- `report_schema_version` - `report_schema_version`
- `output_schema` - `output_schema`
- `config_version` when file config is used - `config_version` (when file config is loaded)
Current values: Current values:
- `report_schema_name`: `audita-process-report` - `report_schema_name = audita-process-report`
- `report_schema_version`: `v1` - `report_schema_version = v1`
`--report-json` output and diagnostics run-dir `report.json` use the same report schema metadata. `--report-json` output and run-directory `report.json` use the same report schema metadata.
Validator decision/rejection records in reports use stable validator keys in `validator_name`. Validator decision/rejection records use stable validator keys via `validator_name`.
Module results may also include warning records for malformed module-stage LLM payloads.
Report diagnostics metadata includes artifact-path fields for utilization diagnostics and correction ledger when diagnostics initialization succeeds.
## Diagnostics directory behavior ## Diagnostics metadata contract
When run-directory initialization succeeds, diagnostics metadata paths reference stable artifacts, including:
- transcript and normalization artifacts;
- chunking summary;
- invocation metadata;
- redacted effective config;
- utilization diagnostics;
- correction ledger;
- `error.log` on failures.
When diagnostics directory creation succeeds, Audita writes run artifacts including: LLM interaction diagnostics include stable prompt and structured-schema identifiers where applicable.
- invocation metadata
- redacted effective config
- transcript/normalization/chunking artifacts
- utilization diagnostics (`utilization-diagnostics.json`)
- correction ledger (`correction-ledger.json`)
- report and failure error log (when applicable)
- module/LLM diagnostics artifacts as available
Retention behavior is controlled by configured retention mode; failed runs are retained. ## Stdout/stderr and exit codes
Success:
- with `--output`, stdout is empty;
- without `--output`, stdout contains transcript JSON only;
- report JSON is not written to stdout.
Diagnostics metadata for LLM interactions may include semi-public prompt identifiers: Failures:
- `prompt_id` - nonzero exit;
- `prompt_version` - human-readable stderr summary;
- `prompt_source` - diagnostics directory path on stderr when available.
- `embedded_path`
- `sha256`
These are diagnostic identifiers, not user-facing prompt override controls. Exit codes:
- `0` success
- nonzero failure
## Stdout/stderr behavior ## Redaction contract
Configured secrets are redacted from:
- effective config outputs;
- diagnostics artifacts;
- report artifacts;
- surfaced adapter/runtime errors.
Success behavior: ## Compatibility policy
- with `--output`, stdout is empty Stable command behavior, schema names, report metadata keys, diagnostics-path field semantics, and validator key identities are treated as public contract.
- without `--output`, stdout contains only transcript JSON in selected output schema
- report JSON is not written to stdout
- success stderr remains empty even when reports/diagnostics contain module warnings
Failure behavior: Additive fields are acceptable when existing fields and behavior remain compatible.
- stderr contains human-readable error summary
- nonzero exit
- diagnostics path is printed when available
## Exit-code behavior
- `0`: success
- nonzero: failure
Treat any nonzero exit as a failed invocation.
## Secret redaction guarantees
Audita redacts API keys and authorization secrets from:
- effective config outputs (`audita config print-effective`, diagnostics effective-config artifact)
- report artifacts
- LLM diagnostics artifacts
- surfaced request/response error messages
Config files should reference secrets via environment variable names (`api_key_env`) rather than embedding secret values.
## Compatibility and deprecation policy
- Existing stable schema names, report metadata keys, and top-level command behavior are treated as public contract.
- Existing stable validator keys remain public contract values even when validator semantics are refined.
- Compatibility inputs (legacy flags/env aliases) may remain during transition windows.
- Any planned removal or behavior change should include clear compatibility notes and migration guidance.
## Breaking changes after 1.0
After 1.0, breaking changes include, for example:
- changing default success/failure exit-code semantics
- changing stdout/stderr routing semantics
- silently changing default output schema shape
- removing supported output schema names without compatibility strategy
- changing report schema fields or meanings incompatibly
- changing config version semantics incompatibly without version bump
Additive fields, additive diagnostics, and new optional schema names are generally non-breaking when existing behavior remains intact.

View File

@@ -1,90 +1,72 @@
# Structured LLM Architecture # Structured LLM Architecture
## Purpose ## Scope
This document describes Audita's structured LLM runtime boundary and adapter behavior. This document describes Audita's structured LLM runtime boundary and adapter behavior.
## Why Audita owns the adapter ## Runtime boundary
Production LLM integration depends on the internal contract only:
Audita owns a small structured LLM adapter so that core runtime behavior is controlled inside the repository: - `contracts.StructuredLLMClient`
- request construction and schema handling are explicit and testable;
- retries, timeouts, cancellation, and error redaction are consistent across modules and validators;
- provider SDK types are not exposed outside the adapter boundary;
- dependency weight and transitive provider-specific behavior are reduced.
At runtime, the rest of Audita depends only on the internal contract:
- `StructuredLLMClient`
- `CompleteStructured(ctx, req, out)` - `CompleteStructured(ctx, req, out)`
## OpenAI-compatible request shape Provider SDK types do not leak past this boundary.
At a conceptual level, Audita sends chat completion requests with: ## Adapter ownership
- `model` `internal/framework/llm` owns the OpenAI-compatible HTTP adapter and shared LLM runtime utilities.
- `messages` (role/content pairs)
- `response_format`:
- `type = "json_schema"`
- `json_schema.name` (stable schema name)
- `json_schema.strict = true`
- `json_schema.schema` (registered JSON Schema payload)
The adapter uses OpenAI-compatible `POST {base_url}/chat/completions` over `net/http`. Key responsibilities:
- request assembly;
- timeout/cancellation propagation;
- bounded retry behavior;
- scheduler integration;
- provider response decoding;
- error redaction.
## Structured response schema registry ## Structured schema registry
Structured response schemas are registered in `internal/framework/responseschema` and include stable metadata:
- `id`
- `version`
- `name`
- `json_schema`
- `sha256`
Structured response schemas are registered in `internal/framework/responseschema` with stable metadata: Current schema keys:
- schema key - `correction_set`
- schema ID - `validator_decision_set`
- schema version
- schema name (OpenAI-compatible `response_format` name)
- raw JSON Schema payload
- SHA-256 hash
Current schemas: Schema metadata is attached to diagnostics through `Schema.DiagnosticsMap()`.
- `correction_set`:
- id `audita.correction_set`
- version `v1`
- name `audita_correction_set_v1`
- `validator_decision_set`:
- id `audita.validator_decision_set`
- version `v1`
- name `audita_validator_decision_set_v1`
## Provider compatibility assumptions ## Request shape assumptions
Audita targets OpenAI-compatible chat-completions endpoints and sends structured requests with:
- model;
- chat messages;
- `response_format.type = json_schema`;
- schema name and JSON schema payload.
Audita assumes an OpenAI-compatible chat-completions endpoint that: ## Local validation remains mandatory
- accepts message arrays with model selection; Provider schema enforcement is treated as transport-level guardrails.
- accepts `response_format.type = json_schema`;
- returns a completion with assistant message content and optional usage metadata.
Provider-specific differences are expected in strictness and error payload shapes, so the adapter treats provider output as untrusted until locally decoded. Audita still validates output locally before applying behavior changes:
- proposal decoding and proposal invariants;
- validator decision decoding and cardinality checks;
- deterministic validation and apply-time rules.
## Local decode and validation remain mandatory ## Shared malformed-output policy
Malformed structured-output classification is centralized in `internal/framework/structuredoutput`.
Provider-level structured output is a transport guardrail, not final validation. Proposal generation and validator execution both use this shared classifier so downgrade behavior cannot drift between the two paths.
After receiving a response, Audita still: ## Secrets and redaction
- decodes assistant content into typed request-specific structs; Secret extraction for LLM redaction is centralized in `llm.ConfiguredSecrets(cfg)` and reused by proposal and validator diagnostics writers.
- validates proposal and validator payload invariants locally;
- enforces deterministic validator/cardinality rules before any transcript application.
This protects runtime correctness even when provider responses are malformed, partial, or semantically inconsistent. Secrets are redacted from:
- diagnostics artifacts;
- report artifacts;
- surfaced adapter/runtime errors.
## Diagnostics and redaction ## Concurrency and scheduling
LLM execution is constrained by composed scheduler limits:
- total LLM concurrency;
- proposal LLM concurrency;
- validation LLM concurrency.
When structured schemas are used, diagnostics metadata records: The scheduler is FIFO and context-aware so permits are released on success, failure, and cancellation.
- schema ID
- schema version
- schema name
- schema hash
Diagnostics and surfaced errors preserve secret redaction:
- API keys and bearer tokens are redacted from request/response/error artifacts;
- redaction is applied before diagnostic files are written.
## Runtime behavior guarantees
The structured LLM path preserves existing runtime guarantees:
- bounded LLM call execution through schedulers;
- context-aware cancellation and timeout propagation;
- retry behavior for transient failures and retryable malformed structured responses;
- deterministic module/chunk/proposal/validator behavior outside provider nondeterminism.

View File

@@ -1,79 +1,36 @@
# Audita Validators # Audita Validators
This document describes Audita's built-in validator registry and module validator chains.
For LLM-backed validator prompt asset details, see [`docs/prompts.md`](prompts.md).
## Package ownership
Built-in validator construction is package-owned under `internal/validators/<validator_key>`:
- `internal/validators/confidence_threshold`
- `internal/validators/proposal_shape`
- `internal/validators/original_text_presence`
- `internal/validators/non_empty_corrected_text`
- `internal/validators/no_effect`
- `internal/validators/protected_terms`
- `internal/validators/spoken_form_plausibility`
- `internal/validators/meaning_reversal_review`
- `internal/validators/editorial_review`
Registry and chain wiring stay in:
- `internal/validators/registry.go`
- `internal/validators/chains.go`
Shared validator runtime mechanics stay in `internal/framework/validators`:
- request/result/decision models
- decision cardinality helpers
- protected vocabulary helpers
- shared LLM validator runtime, batching, and diagnostics helpers
Execution classification metadata is defined in `internal/validators/metadata`:
- `deterministic`
- `llm_backed`
Runner ordering uses this metadata so deterministic validators run before LLM-backed validators without concrete framework type assertions.
## Scope ## Scope
This document defines the built-in validator system used by production module runs.
Validator chains are built-in runtime behavior. ## Ownership boundaries
Built-in validator keys, constructors, and module chains are owned by `internal/validators`.
Current 1.0 boundary: Shared runtime execution mechanics are owned by `internal/framework/validators`, including:
- built-in validator keys and built-in module chains are stable runtime identifiers; - validator request/result models;
- thresholds and batching knobs remain configurable where already supported; - deterministic proposal checks;
- arbitrary user-defined validator chains are deferred. - LLM validator batching and execution;
- decision-cardinality enforcement;
- diagnostics integration.
## Built-in validator keys Execution class metadata is owned by `internal/validators/metadata`.
### Deterministic validators
## Stable validator keys
Deterministic:
- `proposal_shape` - `proposal_shape`
- rejects malformed proposal fields before other validators run.
- `confidence_threshold` - `confidence_threshold`
- checks proposal confidence against module-specific configured threshold.
- `original_text_presence` - `original_text_presence`
- ensures target segment exists and `original_text` exists in current working segment text.
- `non_empty_corrected_text` - `non_empty_corrected_text`
- rejects proposals whose previewed resulting segment text would be empty or whitespace-only.
- `no_effect` - `no_effect`
- rejects proposals where `original_text == corrected_text`.
- `protected_terms` - `protected_terms`
- protects glossary-derived terms from unsafe mutations in non-glossary modules.
- glossary stages use glossary-specific protection logic but still report this same stable key.
### LLM-backed validators
LLM-backed:
- `spoken_form_plausibility` - `spoken_form_plausibility`
- checks whether proposed spoken-form change remains plausible in transcript context.
- `meaning_reversal_review` - `meaning_reversal_review`
- checks for likely meaning reversal or semantic contradiction.
- `editorial_review` - `editorial_review`
- performs conservative editorial safety review.
## Built-in module chains ## Built-in module chains
`glossary`:
Current built-in chains resolved from `internal/validators/chains.go`:
- `glossary`
- `proposal_shape` - `proposal_shape`
- `no_effect` - `no_effect`
- `original_text_presence` - `original_text_presence`
@@ -83,7 +40,7 @@ Current built-in chains resolved from `internal/validators/chains.go`:
- `spoken_form_plausibility` - `spoken_form_plausibility`
- `meaning_reversal_review` - `meaning_reversal_review`
- `homophones` `homophones`:
- `proposal_shape` - `proposal_shape`
- `no_effect` - `no_effect`
- `original_text_presence` - `original_text_presence`
@@ -93,7 +50,7 @@ Current built-in chains resolved from `internal/validators/chains.go`:
- `spoken_form_plausibility` - `spoken_form_plausibility`
- `meaning_reversal_review` - `meaning_reversal_review`
- `spoken_word` `spoken_word`:
- `proposal_shape` - `proposal_shape`
- `no_effect` - `no_effect`
- `original_text_presence` - `original_text_presence`
@@ -103,7 +60,7 @@ Current built-in chains resolved from `internal/validators/chains.go`:
- `editorial_review` - `editorial_review`
- `meaning_reversal_review` - `meaning_reversal_review`
- `grammar` `grammar`:
- `proposal_shape` - `proposal_shape`
- `no_effect` - `no_effect`
- `original_text_presence` - `original_text_presence`
@@ -113,48 +70,27 @@ Current built-in chains resolved from `internal/validators/chains.go`:
- `editorial_review` - `editorial_review`
- `meaning_reversal_review` - `meaning_reversal_review`
## Protected terms construction ## Ordering and execution semantics
Validator ordering is based on canonical metadata:
- deterministic validators run before LLM-backed validators.
`protected_terms` has explicit constructors: Within each module stage:
- general constructor used by non-glossary modules through the built-in registry - proposals are generated per section;
- glossary-stage constructor used by glossary chain resolution - validator chains execute on those proposals;
- approved proposals are applied once after section work settles.
Both variants preserve existing behavior and report the stable key `protected_terms`. ## Malformed payload behavior
Malformed structured-output from proposal generation and LLM validator calls is downgraded, not treated as a process-fatal transport error.
## Execution semantics Current outcomes:
- malformed proposal-generation payloads produce section/module warnings and zero proposals for the affected section;
- malformed validator decision payloads reject the affected validator batch with warnings;
- deterministic validator behavior and runner order remain unchanged.
- modules execute serially; ## Reporting identity
- section proposal work can run concurrently within a module; Reports and diagnostics use stable validator keys as identifiers.
- deterministic validators run before LLM-backed validators;
- malformed module proposal payloads are downgraded to section-scoped module warnings with zero proposals for the affected section rather than module failure;
- malformed/missing/duplicate/unknown LLM validator decisions reject the affected validator batch with warnings instead of failing the module;
- oversized single-proposal validator inputs reject only the affected proposal under that validator;
- approved proposals are applied once per module after section work settles.
## Validator rejections vs proposal-application skips Correction-ledger deterministic-vs-LLM classification is derived from canonical validator metadata, not package-local hardcoded maps.
- validator rejection: ## Prompt assets
- proposal is denied by validator-chain review and appears in validator rejection reporting with validator key and reason code. LLM validator prompt assets and prompt metadata are documented in [Prompts](./prompts.md).
- proposal-application skip:
- proposal passed validators but could not be applied under replacement-policy semantics (for example no matching span at apply time).
- module warning:
- malformed proposal-generation payloads and malformed validator batches are recorded in module warning records and diagnostics without writing success stderr.
These are separate outcomes and are reported separately.
## Reporting and diagnostics identity
- report validator decision/rejection entries use stable validator keys in `validator_name`.
- report module results include warning records for malformed module-stage LLM payloads.
- validator LLM diagnostics include validator identity in interaction metadata and structured response schema metadata.
- correction ledger entries include deterministic and LLM validator decision snapshots keyed by the same stable validator keys, and keep validator rejection distinct from application-level skip.
Prompt assets are unchanged by the validator package-ownership refactor and remain built-in under `internal/prompts`.
## Configurable knobs that remain supported
- per-module confidence thresholds (`thresholds.*` / equivalent env+CLI overrides)
- validation batching limits (`validation_max_prompt_tokens` / equivalent env+CLI overrides)
- validation LLM model/base URL/timeout/retries/concurrency settings
These tune validator behavior without exposing arbitrary user-defined chains.

View File

@@ -1,51 +1,48 @@
# Audita Configuration # Audita Configuration
This document describes Audita's versioned YAML config support and related commands. ## Scope
This document defines the supported versioned YAML configuration model and runtime precedence behavior.
## Purpose
Audita's config file provides a stable place for pipeline defaults and runtime tuning that would otherwise require many environment variables or CLI flags.
Use config files for baseline settings, then use environment variables and CLI flags for deployment and per-run overrides.
## Supported version
Current supported config version:
## Supported file version
Current supported config file version:
- `version: 1` - `version: 1`
Rules: Validation rules:
- missing `version` fails;
- missing `version` fails validation; - unsupported version fails;
- unknown versions fail validation; - unknown YAML fields fail (strict decoding).
- unknown fields fail validation (strict decoding).
## Config path resolution ## Config path resolution
For `audita process` and `audita config print-effective`, path resolution order is:
1. `--config <path>`
2. `AUDITA_CONFIG`
3. `/usr/local/etc/audita/config.yml` (if present)
4. `/etc/audita/config.yml` (if present)
For `audita process`, config path resolution is: Missing-path behavior:
- missing `--config` path is an error;
- missing `AUDITA_CONFIG` path is an error;
- missing both default paths is non-fatal.
1. `--config <path>` if provided ## Effective precedence
2. `AUDITA_CONFIG` if set and `--config` is not provided `audita process` effective precedence:
3. default `/usr/local/etc/audita/config.yml` if present 1. defaults
4. fallback default `/etc/audita/config.yml` if present
Missing-file behavior:
- missing `--config` path: hard failure;
- missing `AUDITA_CONFIG` path: hard failure;
- missing both default-path files: non-fatal, run continues.
## Precedence model
Effective config precedence is:
1. built-in defaults
2. file config 2. file config
3. environment overrides 3. environment overrides
4. CLI overrides 4. CLI overrides
## Supported YAML fields `audita config print-effective` uses:
1. defaults
2. file config
3. environment overrides
`audita config validate` intentionally uses file-only validation:
1. defaults
2. file config
Environment overrides are not applied in `config validate`.
## Supported top-level YAML fields
```yaml ```yaml
version: 1 version: 1
@@ -62,7 +59,6 @@ llm:
api_key_env: AUDITA_LLM_API_KEY api_key_env: AUDITA_LLM_API_KEY
timeout: 120s timeout: 120s
max_retries: 3 max_retries: 3
validation: validation:
base_url: https://openrouter.ai/api/v1 base_url: https://openrouter.ai/api/v1
model: openrouter/google/gemma-4-31b-it model: openrouter/google/gemma-4-31b-it
@@ -100,91 +96,54 @@ diagnostics:
retention: auto retention: auto
``` ```
`context.description` provides background-only transcript context for prompts. ## Module and output-schema validation
If both config and CLI provide a description, `--transcript-description` takes precedence. `pipeline.modules` keys are validated against the built-in supported module catalog.
`output.schema` supports the built-in output schema registry values: Supported module keys:
- `bare-segments` (default) - `glossary`
- `homophones`
- `spoken_word`
- `grammar`
Repeated supported module keys are allowed.
`output.schema` is validated against the built-in output schema catalog.
Supported output schema keys:
- `bare-segments`
- `audita-v1` - `audita-v1`
Unknown schema names fail clearly before transcript output is written. Unknown module keys and unknown output schema keys fail validation.
Duration-like fields accept either: ## Duration field parsing
Duration-like fields support:
- numeric seconds (for example `120`, `3.5`)
- duration strings (for example `120s`, `2m`)
- numeric seconds (for example `120`, `3.5`), or LLM timeout duration strings must resolve to whole seconds.
- duration strings (for example `120s`, `2m`).
For LLM timeouts, duration strings must resolve to whole seconds.
## Secret handling ## Secret handling
Use `api_key_env` fields for secrets:
Use `api_key_env` for secrets:
- `llm.proposal.api_key_env` - `llm.proposal.api_key_env`
- `llm.validation.api_key_env` - `llm.validation.api_key_env`
These fields must contain environment variable names, not secret values. These fields store environment variable names, not secret values.
At runtime, Audita resolves those names from the process environment. Resolved secret values are redacted from:
- `audita config print-effective` output;
Redaction behavior: - diagnostics `effective-config.json`;
- report and diagnostics payloads.
- run diagnostics `effective-config.json` is redacted;
- `audita config print-effective` output is redacted;
- API keys are never emitted in plaintext by those outputs.
## Config commands
Validate a config file:
## Commands
Validate a file config:
```sh ```sh
audita config validate --config ./audita.yml audita config validate --config ./audita.yml
``` ```
Print redacted effective config: Print redacted effective config:
```sh ```sh
audita config print-effective --config ./audita.yml audita config print-effective --config ./audita.yml
``` ```
`print-effective` loads defaults, then file config, then environment overrides.
## Example: local OpenAI-compatible endpoint
```yaml
version: 1
llm:
proposal:
base_url: http://localhost:8000/v1
model: local/proposal-model
api_key_env: AUDITA_LLM_API_KEY
timeout: 90s
max_retries: 2
validation:
base_url: http://localhost:8000/v1
model: local/validation-model
api_key_env: AUDITA_VALIDATION_LLM_API_KEY
timeout: 90s
max_retries: 2
pipeline:
modules: [glossary, homophones, glossary, spoken_word, grammar]
diagnostics:
work_dir: /tmp/audita
retention: auto
```
## Compatibility notes ## Compatibility notes
Legacy compatibility flags and environment aliases remain available where implemented, but the stable configuration surface is the versioned YAML model described above.
Existing environment variables and lower-level CLI flags remain available for compatibility.
Current guidance:
- prefer file config for baseline behavior;
- keep environment variables for secrets/deployment-specific overrides;
- use CLI flags for per-run overrides.
- validator chains are built-in and are not user-configurable in config.
- prompt source selection and filesystem prompt overrides are not config options.

33
docs/development.md Normal file
View File

@@ -0,0 +1,33 @@
# Audita Development Workflow
## Scope
This document defines the canonical contributor workflow and engineering conventions for this repository.
## Workflow
1. Start from a clean understanding of scope and constraints.
2. Make focused changes that preserve existing public behavior unless behavior change is explicitly intended.
3. Run targeted tests for touched packages.
4. Run `go test ./...` before finalizing substantial changes.
5. Update affected documentation so it describes current behavior only.
## Engineering conventions
- Keep module packages separate: `glossary`, `homophones`, `spoken_word`, `grammar`.
- Prefer narrow shared helpers and catalogs over broad abstractions.
- Preserve diagnostics artifact naming and report field contracts unless intentionally changed.
- Preserve CLI/config precedence semantics unless intentionally changed.
- Treat stable validator keys, prompt identifiers, and output-schema keys as contract surfaces.
## Configuration and runtime expectations
- `audita process` precedence is defaults -> file -> env -> CLI.
- `audita config validate` validates file config merged onto defaults only.
- `audita config print-effective` includes environment overrides and prints redacted JSON.
## Testing expectations
- Add tests for new behavior and for bug fixes.
- Keep deterministic fixtures stable.
- Do not reduce existing parity, release-fixture, subprocess, or module-specific coverage without equivalent replacement.
## Commit discipline
- Keep commits scoped and reviewable.
- Avoid mixing unrelated refactors with behavior changes.
- Use clear plain-English commit messages.

View File

@@ -0,0 +1,27 @@
# Documentation Policy
## Scope
This policy defines how project documentation should be authored and maintained.
## Core rules
- Document the current behavior of the codebase.
- Remove stale behavior descriptions promptly when code changes.
- Do not describe development history in architecture or behavior docs unless a document is explicitly historical.
- Do not use architecture or behavior docs as changelogs.
- Prefer rewriting stale sections from scratch when substantial behavior or ownership changes occur.
## Consistency requirements
- Keep command examples aligned with current CLI surfaces.
- Keep configuration examples aligned with supported fields and precedence.
- Keep architecture package ownership descriptions aligned with current code layout.
- Keep stable contract identifiers accurate (module keys, validator keys, output-schema keys, report metadata fields).
## Cross-document expectations
- `docs/architecture/*` documents runtime behavior and package ownership.
- `docs/configuration.md` documents config schema and precedence.
- `docs/development.md` documents contributor workflow and engineering conventions.
## Review expectations for documentation changes
- Verify referenced files and links exist.
- Verify examples match current behavior.
- Prefer concise, direct language and avoid speculative future claims.