How is audit built?

One installation belongs to one application. Concepts defines the vocabulary. Direct mode and stream mode show each shape as run.

The bucket contract and the capabilities matrix state what exists per platform.

The parts

flowchart TB
  E["emit library<br/>validates against<br/>the catalogue"]
  R["receiver<br/>audit-writer, front door"]
  NATS[("JetStream<br/>stream mode only")]
  W["writer<br/>audit-writer"]
  S3[("the environment's bucket<br/>THE RECORD")]
  PG[("database<br/>dedupe, registry, index")]
  O["indexer<br/>audit-observe"]
  Q["query service<br/>audit-query"]
  UI["Audit page"]
  E -- "Connect,<br/>ack = durable" --> R
  R -- "direct: put, then ack" --> W
  R -- "stream: publish" --> NATS
  NATS --> W
  W -- "locked objects" --> S3
  W --> PG
  O -- "lists from a cursor" --> S3
  O -- "index rows, cursors" --> PG
  UI -- "console's own token" --> Q
  Q --> PG
  Q --> S3

part	runs as	holds	never holds
emit	library in the application	compiled-in catalogue, bounded in-memory queue	bucket, index or stream credentials
receiver	audit-writer, one or two pods	stream credentials in stream mode	none
writer	consumer mode, N pods; the receiver in direct mode	write on its prefix, a dedupe and registry role	the index, any way to hand a record back
indexer	audit-observe, one pod	read on the archive, the index role	write on the archive, the dedupe table
notary	audit-notary, hourly CronJob	read on the prefix, put under seals/, the seal key	write on records/
verify	nightly CronJob	read on the prefix	write on the archive
query service	audit-query, one or two pods	read-only index role, read on the prefix	write on the archive or index
Audit page	React component in the console	nothing	credentials
usage consumer	designed, not built: a small Deployment for quotas	the counter cache	none

The application holds no credentials for the trail. A compromised application pod can reach the archive only by emitting records. Nothing in the write path can read a record back. The query service is the only way in, and it records every read, a refused one included. Every listener refuses a request body over 4 MiB.

Each install preset has a bucket of its own, and Object Lock in compliance mode belongs to the attested preset's bucket (0068). The index is one database in the application's Postgres: rows, facet counts, cursors, and the writer's dedupe table under another role. The index rebuilds from the archive. The dedupe table does not, and neither do wrapped pseudonym keys kept there with keys.state.backend: database.

The catalogue

One YAML file lives with the application code, with a JSON Schema per action that carries data. It names each action's operation, framework categories, destination category, subject, actors, data schema, delivery and sentence template.

• A catalogue is one per application. It names framework profiles and never defines them.
• Every record names its catalogue version. The archive keeps every version.
• The catalogue reaches the receiver at start-up over RegisterCatalogue. A malformed catalogue is refused and the application does not start. An unreachable receiver is retried.
• Use one constructor per action, so check-emitters can hold the code to the file.

CI runs audit validate and check-emitters. The emit library refuses a mismatching record, and the receiver and writer validate again. See the catalogue reference.

What an acknowledgement means

Delivery is set per action in the catalogue.

delivery	the call returns	receiver down	for
block	when the receiver acknowledges durability	the action fails	privileged sign-in, key destruction, billable operations
async (default)	at once	the record waits in a bounded queue and retries with backoff	everything else

The acknowledgement states durability. Archived is the object in the bucket, which the receiver puts before answering in direct mode. Queued is a replicated queue's acknowledgement. Logged is a log line. The start-up guard sink.Require refuses a chain that can never give what the deployment needs.

What each mode can lose

A block record is never lost. For async:

event	direct mode	stream mode
application pod dies with records queued	unacknowledged records: one flush interval (default one second) plus the batch in flight	the same, shorter
writer dies holding stream records	not applicable	nothing, because the stream redelivers
receiver or writer crashes	nothing	nothing
long receiver outage overflows the queue	oldest dropped and counted	the same
container restarts, pod intact	the queue is gone, as in row one	the same

The emitter logs each record the queue gives up. Alert on audit.emit.records.dropped and watch audit.emit.queue.pending.

The life of one record

1. The emitter fills id, time, source, sequence, client address and ids. It validates against the catalogue and delivers as declared.
2. The receiver stamps recorded_at, the observer from the caller's verified token, and origin_hash. In stream mode it stamps here, and the writers keep a stamp whose hash matches. From an SQS queue the writer stamps from SenderId and SentTimestamp: see authentication.
3. The writer splits the record into one copy per profile, keeps only the allowed fields, and treats each identity by category. See the writer.
4. The writer rolls copies into objects per ingest batch, profile and tenant, keyed by the hour of ingest. It puts each under an Object Lock retention computed from the profile. A record it cannot take goes to the dead-letter prefix.
5. The writer marks the id as written. The indexer lists the bucket once an object passes the settle window and writes the rows and facet counts a couple of minutes later.
6. The verify job checks the previous day's objects nightly. The notary makes seals hourly, checked with --root.
7. A reader asks the query service. Their grant becomes one term of the query, and the read is recorded.

The two shapes

	direct	stream
for	internal service, low volume, no stream	product with many pods, metering, quotas
receiver	is the writer and puts to the bucket	publishes to JetStream
writers	the receiver's pods	N consumers, scaled apart
needs	bucket, database	bucket, database, NATS account

Switching from direct to stream changes only the receiver's configuration. The archive layout is the same, so installations on different versions can share a bucket.

Extensions

Billing and usage quotas are designed, not built. The chart accepts extensions.billing.enabled and extensions.quotas.enabled and renders nothing for either. The metering copy of a billable record is built: see billing and usage quotas.

Projections (index rows, facet counts) rebuild from the archive with audit reindex, so none is backed up. The writer rolls many records into one object per profile, tenant and ingest hour.

What is built

	state
record, catalogue, framework profiles, emitter, writer, query service, v1 bucket layout, verify, legal holds, retention addenda	built
searchers: Postgres, archive scan, memory	built
signers: key file, AWS KMS, OpenBAO transit	built
key providers local, OpenBAO transit and AWS KMS	built, off by default
one validated configuration file per binary	built
chart per application, golden per shape	built
@truvity/audit, @truvity/audit-react	built; published to GitHub Packages at each release tag
TypeScript emitter	designed, not built
hourly seals, the notary on Kubernetes and on Lambda, audit verify --root	built; not yet run live
export from the query service, as NDJSON or CSV	built
metering copy of a billable record	built
billing rollups and statement, usage consumer, reconciler	designed, not built
exporters (OCSF, ECS, Parquet), adapters	designed, not built

Decided in

• 0045 S3 Object Lock as the record.
• 0053 One installation per service or product.
• 0058 Three parts installed independently.
• 0059 Sink durability and transports.
• 0060 The v1 bucket layout.
• 0061 Seals.
• 0062 Observe follows the bucket.
• 0066 Indexer and query are separate processes.