Jev decisions: operator reference
Start with mise run classification -- status --database /absolute/path/hippo.db. This reads classification availability and queue state without starting inference or migrating the database.
The implementation provides optional Jev reranking, deterministic rules, controlled-topic classification, topic retrieval, and isolated evaluation. Every production feature starts disabled. Implementation tests do not establish the quality, latency, or human-review acceptance criteria in the approved specification ↗.
Configuration and compatibility
Use the normal Hippo configuration file, ~/.config/hippo/config.toml. These values show the defaults:
[retrieval]
rerank = false
rerank_backend = "local"
rerank_recipe = "note-v1"
rerank_adaptive = false
rerank_pool = 30
decision_deadline_ms = 2000
topic_retrieval = false
decision_log = false
[classification]
enabled = false
# recipe_path = "/absolute/path/calibrated-topics.json"
Enable a qualified Jev ranking recipe with rerank = true and rerank_backend = "jev". Backend "rules" runs the existing declarative lexical decision table. Backend "local" retains local LLM ranking. Classification and topic retrieval have independent switches. Configuration changes require restarting the affected brain and MCP processes through the normal service workflow.
Set [brain] typesafe_api_key_op_ref in ~/.config/hippo/config.toml to a 1Password secret reference such as op://<vault>/<item>/<field>. The installed brain-wrapper.sh resolves it with op read at service start and exports the result as TYPESAFE_API_KEY, so no plaintext copy is stored. The op executable must be available on the LaunchAgent PATH or at /opt/homebrew/bin/op or /usr/local/bin/op, with 1Password authentication available to that process; shell-only authentication is not inherited. If op is missing or its lookup fails, the wrapper emits a credential-free warning and still starts the brain. New hippo-mcp processes read the same reference with a ten-second timeout and pass the key directly to Jev. When a reference is configured, the process removes any inherited TYPESAFE_API_KEY from its own environment, so a failed lookup can never send a stale key. Each rerank then falls back with cause no_credentials and makes no request. Each hippo-mcp process retries the lookup in the background, at most once every five minutes and only when a rerank needs it. The lookup fails when 1Password is locked or its authorization prompt is not approved in time, and each attempt can raise that prompt. Without a configured reference, Jev clients use TYPESAFE_API_KEY from their environment. Setting the variable in a terminal does not update an already running LaunchAgent. Keep the reference in machine-local installation configuration, selecting the HIPPO_TYPESAFE_API_KEY item and its credential field in your vault. After adding or changing the reference, run hippo daemon install --force to regenerate the plist, then restart the brain service; restarting alone does not apply the new reference. An empty reference with no existing key emits a startup warning and the brain continues without Jev credentials. The client sends bounded, redacted input to https://api.typesafe.ai/v1/systemone, pins version 1.13.0 using API model name jev-1.13.0, and makes one HTTP attempt per assessment. It permits four concurrent requests per client and enters a 30-second cooldown after three endpoint failures. Missing credentials or endpoint failures preserve the last complete ranking; the fast path does not invoke the local LLM as a fallback.
Install matching Rust and Python components through the normal deployment workflow before enabling classification. See schema compatibility for migration ownership and accepted database versions. Missing classification schema disables classification with a diagnostic. Do not apply experimental SQL directly to the production database. Classification CLI commands do not migrate it.
note-v1 is the default recipe. It asks one relevance Noul per candidate and places that candidate inside its own question as a readable note: the full summary (up to 1,000 characters), intent, outcome, up to four decisions, up to three problems, the project name, the branch, a calendar date, and 300 characters of identifiers. Each question also states that the note is untrusted text and that Jev must not follow an instruction inside it. The state holds only the query and the requested project scope. Each note is redacted before it is sent and is held to 3,000 encoded bytes, so that 30 notes and a query stay inside the 128 KB request limit; a note over that size loses its identifiers first, then its problems and decisions. The Noul probability is the ranking score, and note-v1 never refines. The rules backend ignores the recipe and always reads the shared candidate list. single-v1 retains the original relevance rubric. multi-v1 combines relevance, direct evidence, and scope fit with weights 0.50, 0.35, and 0.15. Both older recipes send one shared list of candidates, cut to 300 characters per field, and address each candidate by index.
On 82 real queries with blind relevance labels, counting only notes written before each query, note-v1 scored nDCG 0.873 at each query’s own limit, multi-v1 0.796, and retrieval order 0.729. The gain of note-v1 over retrieval order held in both halves of the query set. Its gain over multi-v1 (+0.077, 95% interval +0.022 to +0.132) is less certain: the note-v1 variant was chosen after the second half had been scored. A note-v1 request uses a median of about 14,400 input tokens against 23,324 for multi-v1. The labels came from model judges, not from people. See the evaluation record. Adaptive mode is a development recipe: it fetches eligible source evidence for at most eight uncertain candidates and rescoring stops on stable order, sufficient separation, missing new evidence, or the budget. Online decisions admit at most 30 candidates, three passes, and four reserved attempts within a maximum 2,000 ms deadline, including client queue wait. Repeating identical input is not refinement.
Classification and topic discovery
The built-in classification recipe has 14 topics, a development threshold of 0.8, and a maximum of five accepted topics. Those defaults are not a claim that publication precision has passed its gate.
An evaluation on 2026-11-13 found no retrieval use that gains from these topics. On 160 model-judged nodes, the accepted labels of the built-in recipe were at least a minor subject of the node 86% of the time and a substantive subject 56% of the time. Enable classification only for a reader that needs the labels. A candidate recipe with one plain question per topic scored higher and was not adopted. See the evaluation record.
A calibrated recipe can override thresholds and limit publication to supported topics:
{
"taxonomy_version": "technical-topics-v1-calibrated-01",
"thresholds": {"database-storage": 0.94},
"publish_topics": ["database-storage"],
"max_topics": 5
}
The number above demonstrates the file format; it is not a recommended calibrated threshold. Set classification.recipe_path to the reviewed file. Omitted topic thresholds retain 0.8. An empty publish_topics list publishes no memberships. Recipe model identity must match the pinned client. Recipes also support explicit topic definitions and input version; unsupported fields and invalid values are rejected.
HTTP, MCP, enrichment scripts, and classification commands load the same configured recipe. CLI --recipe PATH overrides it for an individual command. A different recipe changes the hash and makes prior results ineligible for that reader. Keep the recipe fixed throughout a run and use a new backfill cursor after changing it.
Source writers enqueue the latest node input in their transaction. The worker checks node UUID, input hash, revision, recipe, and lease before publication. It uses two concurrent assessments, three attempts per revision, and expiring leases. Invalid answers never publish a partial topic set. An empty accepted set is a valid result. Original tags, content, FTS, vectors, and source links remain unchanged.
Owned concept identities use hippo/topic/ plus an explicit ownership marker. Source concepts using reserved names retain their display text and receive an escaped canonical identity. A pre-existing unowned identity collision causes abstention; classification does not take ownership of that entity. Changed nodes lose stale owned memberships even when classification is disabled.
After an approved migration and rollout, queue one bounded backfill page:
mise run classification -- backfill \
--database /absolute/path/approved-working.sqlite \
--cursor "$HOME/.local/share/hippo-bench/classification/backfill.json" \
--limit 100
This command writes queue state to the selected database. Use an approved deployment or a writable experiment clone, never the immutable benchmark baseline. It does not call Jev. Repeat the same command until complete is true; the cursor binds the database, recipe, and initial node-ID high watermark. Replayed pages are idempotent. New nodes use the writer hooks, or a new backfill job. The enabled brain worker drains the queue. The worker claims two nodes per brain poll, about 22 nodes per minute at the default 5-second interval, so a 30,000-node corpus drains in about 22 hours. A 1,000-row page holds the SQLite write lock long enough to make concurrent worker claims fail with database is locked; a node can exhaust its three attempts during that window. Use the default 100-row limit while the worker runs.
After resolving an HTTP failure or lease exhaustion, requeue eligible failures:
mise run classification -- retry-failed \
--database /absolute/path/approved-working.sqlite \
--error HTTPStatusError --error 'lease attempts exhausted' --limit 100
This command writes queue state without calling Jev. Only the two shown error reasons are supported; repeat --error to select both. It resets attempts and leases for failed rows matching the configured recipe (or --recipe override), with no applied revision and unchanged node UUID and input hash. The limit is 1–100 requeues, default 100, not a scan limit: invalid or changed inputs are skipped without hiding later eligible rows. Output reports selected, requeued, and skipped; recipe mismatches and applied revisions are excluded from selected. Repeat as needed after fixing the cause; the enabled worker performs the retries.
Export classification state to Parquet for offline analysis:
mise run classification -- export \
--database /absolute/path/hippo.db \
--out "$HOME/.local/share/hippo-investigations/classification-$(date -u +%Y%m%dT%H%MZ)"
The command reads the database once, so the export is consistent even while the worker runs. It writes classifications.parquet with one row per classification record (nodes never enqueued are absent; status, attempts, timestamps, latency_ms, recipe and input identity, and raw probability and accepted-topic JSON) and topic-probabilities.parquet with one row per ready node and topic (prob, accepted). The output directory must be outside the repository and the Hippo data directory, and existing export files are never overwritten. Both files are staged before publication. Publication links them sequentially and removes already-published outputs on a caught failure; it is not an atomic two-file operation for concurrent readers or process crashes. The Parquet writer is DuckDB from the brain’s development dependency group, so run the command from a checkout synced with uv sync --project brain. Reclassification overwrites rows in place, so an export is the only record of a run’s results after a recipe or threshold change.
Inspect current connections without inference:
mise run classification -- connections \
--database /absolute/path/hippo.db \
--node KNOWLEDGE_NODE_UUID --limit 20 --project /absolute/path/project
Optional filters are --source, --branch, --entity, --memory-category, and --since-ms. Connections show the exact shared topic and provenance for both current memberships. They express shared subject matter, not causal or evidential support.
topic_retrieval = true adds an independent retrieval channel using reviewed query aliases. It admits at most three query topics, 20 candidates per topic, and 40 additional candidates before bounded fusion. Existing source/project/time/exclusion filters still apply. This path uses current metadata and leaves semantic vectors unchanged.
Leave this channel off unless a new measurement supports it. On 90 real queries, the aliases matched a topic for 5 queries and changed the candidate pool for 2, with no gain in ranking quality. When the query topics came from Jev instead, the channel admitted 103 candidates, of which 8 scored as relevant, and it displaced 10 candidates that judges had graded relevant. Topic agreement between a query and a note also added nothing as a ranking signal. The evaluation record has the numbers.
Capture and external artifacts
All generated corpora, labels, logs, clones, responses, manifests, and archives belong outside the checkout and outside Hippo’s production data directory. The default root is $XDG_DATA_HOME/hippo-bench/decisions, or ~/.local/share/hippo-bench/decisions. Output guards resolve symlinks and reject repository or production paths. Maintained code, handwritten fixtures, recipes, and documentation can remain in Git.
Set these environment variables on the serving process only when capture is intended:
| Variable | Output under the decision root | Behavior |
|---|---|---|
HIPPO_QUERY_CAPTURE=1 | query-captures/ | Captures ordinary queries and candidate identities even when reranking is disabled. |
HIPPO_DECISION_CAPTURE=1 | captures/, decision-traces/ | Captures reranker inputs and decision diagnostics when their paths execute. |
Capture invokes no additional inference. Files are redacted, created with private permissions, limited to 128 KB each, and capped at 1,000 entries per directory. Query records identify their origin; benchmark and probe traffic is excluded from ordinary capture. Duplicate query inputs are skipped. Capture failure does not fail a query. Move completed captures to an external archive before starting another bounded collection.
Corpus review and replay
The corpus target is 1,000 retrieval questions over at least 10,000 nodes, plus 3,000 independently sampled classification nodes. Preparation creates source-grounded authoring and review queues. Empty authoring slots are not completed questions, and model annotations are not human ground truth.
-
Freeze source data and inspect actual coverage:
mise run bench:knowledge:corpus -- prepare \ --source /absolute/path/hippo.db \ --questions /external/path/questions.json \ --captures /external/path/query-captures \ --query-target 1000 --node-target 3000 \ --out /external/path/corpus-draft mise run bench:knowledge:corpus -- validate /external/path/corpus-draft mise run bench:knowledge:corpus -- review-package /external/path/corpus-draft \ --out /external/path/review-package--questionsand--capturesare optional. Preparation uses a consistent SQLite backup and records source/input hashes, code identity, embedding metadata, and actual population counts. It does not call inference. Inspectmanifest.json,questions.json,query-authoring.json, andclassification-review.json. The separate review package supplies source evidence, taxonomy definitions, blind judgment queues, and the review rubric; it does not create human judgments. -
Have reviewers author missing questions, resolve answerability, and audit task families before tuning. Keep related sessions, duplicates, and paraphrases in one split. Family hashing produces approximate split counts; it does not force quotas by splitting a family. Record evidence for unanswerable cases and retain missing-source or missing-node cases as coverage failures. Publish reviewed questions to a new directory:
mise run bench:knowledge:corpus -- finalize /external/path/corpus-draft \ --questions /external/path/reviewed-questions.json \ --family-auditor REVIEWER_NAME --out /external/path/corpus-reviewedFinalization requires actual
human_reviewedquestion records and a named family audit. It does not perform that review. To correct family grouping, supply the reviewed commontask_family_idon the affected input questions and runprepareinto a new directory before finalization. The preparer unions their source families before assigning splits. Source-identity corrections also require a new freeze; do not edit a frozen manifest. -
Freeze candidate pools and run ranking through the production boundary:
mise run bench:knowledge:replay -- \ --cases /external/path/cases.json \ --questions /external/path/corpus-reviewed/questions.json \ --db /external/path/corpus-reviewed/baseline.sqlite \ --arms retrieval,rules,jev-single,jev-multi,jev-adaptive \ --repeats 1 --max-requests 300 --max-tokens 50000000 \ --out /external/path/ranking-runThese explicit budgets demonstrate a bounded run and can stop before a large corpus finishes. Choose the budget before dispatch.
cases.jsonis a JSON array of{id, state: {query, candidates}}; each candidate needs a stableuuidand frozen result fields. IDs/text must match--questions, with at most 30 candidates per case. Archivedknowledge_probeinputs can use--targets targets.jsonto recover their original UUID mapping. Raw query captures are inputs for authoring; select and freeze their candidate pools before acceptance replay.Supported arms also include
local-original,local-optimized, andlocal-cached. Fresh local arms require--local-urland--local-model. Cached local responses require--cached-localand are quality replay, not fresh timing. Adaptive replay requires the frozen database. Every arm requires--dbto measure source-policy violations; without it, the measurement remains unknown and cannot pass the source-policy gate. The default--concurrency 1rotates arms sequentially.--concurrency 4runs one arm at a time in batches of four, preserving local-comparator isolation. Every terminal result is checkpointed, and requests/tokens are reserved before dispatch. Resume an unchanged run with the same arguments plus--resume; inspectcompletion.jsonfor exhaustion, interruptions, or incomplete rows. Unknown usage remains unknown in the evidence and consumes its conservative budget reservation. Run the specified repeated latency slice separately with--repeats 3and compare single-query and concurrency-four measurements. -
Pool all tested candidates for blind relevance review, then import completed labels:
mise run bench:knowledge:corpus -- pool /external/path/corpus-reviewed \ --rankings /external/path/ranking-run/results.jsonl --cutoff 30 \ --out /external/path/relevance-review.json mise run bench:knowledge:corpus -- import-labels /external/path/corpus-reviewed \ --labels /external/path/relevance-reviewed.json --out /external/path/relevance-imported.json mise run bench:knowledge:corpus -- import-node-labels /external/path/corpus-reviewed \ --labels /external/path/topics-reviewed.json --out /external/path/topics-imported.jsonRelevance judgments require grade 0 to 3, evidence or a negative rationale, named annotator and type, rubric
relevance-0-3-v1, and reviewed/adjudicated status. Preserve their semantic-query and node-input fingerprints. Changing question text, filters, as-of time, source IDs, or node state requires pooling and grading again; changing only annotation status preserves those fingerprints. Legacy labels without fingerprints cannot qualify release acceptance. Topic judgments must label every taxonomy topic aspresent,absent, orinsufficient_evidence, with evidence. Disagreements require human adjudication. Test acceptance also checks human labels, family audit, coverage, and double review. An unjudged candidate is not an irrelevant candidate. -
Report the locked recipe against retrieval and a selected local comparator:
mise run bench:knowledge:corpus -- report /external/path/corpus-reviewed \ --rankings /external/path/all-arms.jsonl --labels /external/path/relevance-imported.json \ --candidate jev-multi --baseline retrieval --local local-optimized \ --out /external/path/retrieval-report.json mise run bench:knowledge:corpus -- classification-report /external/path/corpus-reviewed \ --predictions /external/path/topic-predictions.json --labels /external/path/topics-imported.json \ --backend jev --recipe /external/path/calibrated-topics.json \ --out /external/path/classification-report.jsonRanking reports require one record per query/arm with complete identities and terminal status; keep repeated latency trials separate from that quality input. Use
--adaptivewhen reporting the adaptive latency gate.sweep --signals PATH --labels PATH --weights PATH --out PATHrecomposes recorded development signals without new inference. Select on development/calibration data, then evaluate the locked test set once. Review the complete acceptance policy ↗, including operational load, exact-topic connections, candidate discovery, and answer/citation checks. A report is evidence for its implemented checks, not blanket release approval.
Review the exact topics behind proposed connections separately from individual-node labels:
mise run bench:knowledge:corpus -- connection-review /external/path/corpus-reviewed \
--predictions /external/path/topic-run/predictions.jsonl --limit 200 \
--out /external/path/connection-review.json
mise run bench:knowledge:corpus -- connection-report /external/path/corpus-reviewed \
--predictions /external/path/topic-run/predictions.jsonl \
--pairs /external/path/connections-reviewed.json \
--plan /external/path/connection-review.plan.json --out /external/path/connection-report.json
Reviewers judge the connecting topic and usefulness for a named discovery task separately. The default sample uses test-only eligible nodes and disjoint endpoints without materializing all node pairs. Preserve the adjacent immutable plan unchanged and annotate every planned pair. The report checks that coverage and uses family-aware uncertainty bounds; generating the queue does not supply its judgments. Legacy reports without a plan can score results but cannot pass the connection gate. Candidate-discovery acceptance uses recall at 30 with equal candidate pools of at most 30; a larger diagnostic pool does not qualify.
Classify the fixed node sample independently for each comparator:
mise run bench:knowledge:tags -- classify-corpus /external/path/corpus-reviewed \
--backend jev --recipe /external/path/calibrated-topics.json \
--limit 3000 --max-requests 3000 --max-tokens 50000000 \
--out /external/path/topic-run
Use predictions.jsonl from that directory for classification-report. The other backends are rules, stored, and local; fresh local classification also requires --model MODEL --base-url URL. Run each into its own directory with the same frozen sample and recipe. The report checks backend/model/input/recipe identities; comparator results cannot qualify Jev publication. Benchmark classification commands use the built-in recipe when --recipe is absent, so pass the configured recipe explicitly for parity.
Re-running the unchanged classify-corpus command resumes its output directory after validating the manifest and checkpoints. It reserves request/token budget before dispatch. Interrupted reservations become terminal error records with unknown usage charged conservatively; they cannot qualify a complete run. Inspect completion.json instead of inferring completion from the existence of predictions.
The metadata-only clone command reuses production classification validation and publication:
mise run bench:knowledge:tags -- metadata /external/path/baseline.sqlite \
/external/path/topic-run/predictions.jsonl --recipe /external/path/calibrated-topics.json \
--out /external/path/metadata.sqlite
Each prediction requires node_uuid, current input_hash, and recipe_hash. Jev predictions require the complete typed response; comparator predictions retain their backend, model, and complete probabilities. Historical summary-only classifications cannot be relabeled as current source-aware predictions. The command verifies preservation and writes only a new external clone.
Run content/FTS and semantic-vector projections as separate ablations from the same untouched baseline:
mise run bench:knowledge:tags -- projection /external/path/baseline.sqlite \
/external/path/topic-run/predictions.jsonl --recipe /external/path/calibrated-topics.json \
--arm fts --out /external/path/fts.sqlite
mise run bench:knowledge:tags -- projection /external/path/baseline.sqlite \
/external/path/topic-run/predictions.jsonl --recipe /external/path/calibrated-topics.json \
--arm vectors --model SNAPSHOT_EMBEDDING_MODEL --base-url EMBEDDING_API_URL \
--out /external/path/vectors.sqlite
The vector arm requires the embedding model recorded in the baseline. Each arm checks original tags, source identity, command vectors, and isolation from the other projection. A second disposable clone proves restoration of content, FTS postings, and vectors; inspect the adjacent completion record. These are explicit clone mutations with additional embedding requests in the vector arm. Runtime flags do not roll them back. Older experiments remain described in the decision-sidecar reference; their results do not authorize production FTS or vector rewriting.
Observability and rollback
mise run brain:api:health exposes classification availability, queue counts, oldest pending age, recent batch outcomes, claim-gate failures, and Jev cooldown diagnostics. Decision traces record recipe/input identity, raw typed judgments, rule decisions, usage, timing, and stop/fallback reasons. Existing telemetry exports hippo.brain.decision.* counters for decisions, actual HTTP dispatches, errors, tokens, and unknown measurements, plus a stage-duration histogram. Labels are bounded to backend/task/outcome/stage/direction, never node/query IDs. On startup, each process adds zero to those counters for every outcome its configured backends can produce and immediately flushes them to the collector. Prometheus rate() and increase() then see the first decision of a per-session hippo-mcp process, unless that decision happens before Prometheus first scrapes the collector (up to one 15 s scrape interval after process start). The duration histogram has no such placeholder, so a process with a single decision contributes no latency sample. The brain also reports classification rows by status through hippo_brain_classification_queue_depth, separate from the enrichment queue gauge. Stage durations overlap and must not be summed. HTTP time is client wall time, not server inference time. Missing token usage is not zero, and latency ratios are not dollar-cost estimates.
Rerank decisions carry four more measurements:
| Signal | Meaning |
|---|---|
Outcome skipped | The decision needed no backend call (one candidate, invalid input, or no budget). It is not counted as a success. |
hippo.brain.decision.fallbacks, label cause | Why a rerank kept retrieval order: deadline, connect_timeout, transport, http_auth, http_rate_limit, http_server, http_other, no_credentials, cooldown, invalid_request, invalid_response, or other. |
hippo.brain.decision.effects, label effect | What a completed ranking did to the delivered results against retrieval order: unchanged, reordered, top_changed, set_changed, or set_and_top_changed. |
hippo.brain.decision.best_relevance | The highest composed Jev relevance (0 to 1) among the candidates of a decision. A low value means that no candidate matched the query. |
The duration histogram has bucket boundaries at 1,500, 2,000, and 2,500 ms, and a stage overrun that records the time a decision ran past its deadline. Each rerank decision also logs one line, rerank decision backend=… outcome=… cause=… effect=… elapsed_ms=… input_tokens=…, with no query or candidate text. The Rerank reliability and effect row of the hippo-enrichment dashboard charts these signals.
hippo-mcp runs each Jev rerank on a private event loop. Tool handlers run synchronous SQLite on the server loop, and a decision awaiting there could not observe its response or its deadline while another call held the loop. Measured before this change, 96% of deadline fallbacks happened while another rerank ran in the same process.
Results that Jev ranked carry a relevance field (the composed score, 0 to 1) in MCP search hits. A result without that field came from retrieval order: the rerank was disabled, skipped, or fell back. agent_query also returns fallback_cause and effect in its retrieval object.
Set decision_log = true in [retrieval] to append one line per rerank decision to decision-log/YYYY-MM.jsonl under the decision root. A line holds the outcome, cause, effect, timings, token counts, candidate UUIDs, the final order, and the scores. It holds no query or candidate text. This log is the only decision record that outlives the 30-day telemetry and log retention. A monthly file stops growing at 32 MB.
hippo doctor reports classification and Jev endpoint health from the brain /health response. It fails when enabled classification has an unavailable, stopped, or erroring worker, or when the Jev client is in cooldown. It warns on failed rows, recent endpoint failures, or a pending row older than 24 hours. The Decisions row of the hippo-enrichment Grafana dashboard plots decision outcomes, stage latency, classification backlog, Jev tokens, error ratio, and HTTP dispatches.
The brain worker yields new claims while paused or while queries are in flight. HTTP and MCP queries share a file-lock gate with classification when classification is configured. The lock lives beside the canonical database as DATABASE_NAME.classification-claims.lock, with mode 0600. Queries hold shared locks; classification takes a nonblocking exclusive lock only around its short claim transaction and releases it before inference. Do not delete the lock file while processes use it. A gate failure allows the query to proceed, prevents a background claim, and increments process-local diagnostics. Disabled classification creates no gate file. Standalone MCP processes consume classification state but never start workers.
Classifier claim, publication, and failure writes use owned background-thread connections so SQLite busy waits do not block the brain event loop. Cancellation checks prevent pending transactions from publishing; a claim committed before cancellation can remain until its bounded lease expires.
Classification does not depend on local chat-model preflight. Already dispatched work remains bounded by its timeout; query arrival does not retroactively cancel it. Verify query latency under the declared deployment load before claiming the operational acceptance gate.
Disable a feature by changing its configuration and restarting affected processes:
- Set
retrieval.rerank = falseto restore retrieval order, or selectrerank_backend = "local"to retain local reranking. - Set
retrieval.topic_retrieval = falseto remove classification-driven candidate expansion. - Set
classification.enabled = falseto stop new background assessments. Valid state remains stored; source updates still invalidate stale memberships. - Remove the capture environment variables to stop capture. Preserve collected evidence in an external archive.
Metadata-only rollback does not require deleting original content or vectors. Do not delete all concepts or entity links to remove classification. FTS/vector experiment clones are separate artifacts; disabling a flag cannot remove words or embeddings already written by an experimental transformation.