Fastmod

Fastmod records independent 0–1 rubric scores from one LLM call per evaluation, without running full post analysis. Optional root meta tagging can apply the meta result to existing tag discovery; other scores remain advisory. Fastmod is disabled by default and intended for approved operators or internal workflows.

Configuration

Set FAST_SCORING_ENABLED and FAST_SCORING_AUTOMATIC_ENQUEUE_ENABLED to true. The scorer always uses the normal LLM_* provider settings. It may use FAST_SCORING_LLM_MODEL to select a lower-cost model from that same provider. An absent or blank override inherits LLM_MODEL; generated configuration defaults to inheritance. Explicit overrides are preserved, so use a model identifier valid for the configured endpoint (for example, gpt-5-nano for direct OpenAI).

FAST_SCORING_PROMPT_PATH selects the shared task introduction (default: prompts/fastmod_introduction.txt). The former generated default, prompts/fast_post_scoring_system.txt, is automatically treated as the new introduction path. Custom introductions must request CSV, not JSON. The model receives target text plus bounded parent/root context for replies.

Rubric selection

Open Tools → Fastmod as a root-approved operator. Each rubric displays its repository prompt beside its checkbox. Enable one or more rubrics and save using the browser identity signature. No individual rubric is mandatory, but the panel and harness reject an empty selection. Saving settings does not enqueue work or call the model. Disabled rubric text is omitted from the prompt and response contract entirely.

Each rubric is a compact .txt template under prompts/fastmod/, ideally a single bullet. The filename without .txt is its stable identity (lowercase letters, digits, underscores; first character a letter). Initial rubrics are moderation and meta. Meta covers catch-all application support: feature requests, bug reports, and help concerning the platform, instance, or interaction with the application and interfaces. Ordinary subject-matter questions are not meta. Support and hostility may overlap.

Adding a template automatically adds a panel option; changing/removing templates requires no scoring-pipeline changes. Preserve an identity's meaning; use a new identity for a different signal. Before removing an enabled rubric, disable it and leave at least one other selected. Missing, empty, or invalid selected rubrics fail configuration rather than silently disappearing. A renamed rubric is a removed rubric plus a new one.

Selection is stored privately and atomically in state/private/fastmod_rubrics.json, shared by web and worker processes; ensure both run against the same application root and can access this file. Without a saved selection, both initial rubrics are enabled; a PHP private configuration may set FAST_SCORING_RUBRICS to a nonempty array of IDs as the initial selection. The saved panel selection takes precedence. IDs are sorted lexically for the prompt and CSV order.

The response is only comma-separated probabilities, for example 0.8,0.2 for meta,moderation, or 0.8 for meta alone. The harness trims surrounding/field whitespace and validates exact count, numeric syntax, finite values, and range. It rejects prose, markdown, missing/extra fields, and invalid values as a whole; it never fills missing scores, clamps values, or accepts partial results. Recorded results use rubric IDs, not CSV positions. The legacy probability field is moderation only and is null for meta-only results; it is never a blend.

FAST_SCORING_DATABASE_PATH optionally selects private SQLite score state. It defaults to state/private/fast_scores.sqlite3, outside the published read-model database. Stored results are current only for the matching post content hash and effective introduction/ordered-rubric revision. Changing selection or templates does not automatically score existing posts. Pending work from an older composition is excluded with rubric_changed_before_score before a provider call. Use the existing explicit audit/backfill workflow if re-evaluation is wanted; historical results keep their original rubric identity.

For an audit or historical backfill using a model other than gpt-5-nano or openai/gpt-5-nano, set FAST_SCORING_INPUT_USD_PER_MILLION and FAST_SCORING_OUTPUT_USD_PER_MILLION to that model's current provider prices. They can also be supplied as matching command options for one invocation.

Batch scoring

Only newly published posts create ordinary private Fastmod work. The worker never scans the existing corpus. It processes that work through the existing task queue without making the publishing request wait for the provider:

./v3 task-queue enqueue-fast-score
./v3 task-queue run --limit=1 --score-limit=25 --work-limit=250

The enqueue command is coalesced: it leaves one queued or running sweep rather than creating duplicates. A sweep makes up to --score-limit provider calls (25 by default); a provider failure or invalid response consumes one call because the request was made. --work-limit separately bounds all examined private work rows (250 by default), including local heuristic exclusions that need no provider call. --limit is the number of queue tasks a worker claims, not a post or provider-call limit. Provider and invalid-response failures receive at most two delayed retries after the initial attempt; further retries require an explicit operator command.

Install the existing task-queue cron reference with ./v3 task-queue cron. It runs one locked worker task per minute by default; --score-limit defaults to 25 provider calls and --work-limit defaults to 250 examined work rows per sweep. The execution lock keeps concurrent cron invocations from running overlapping sweeps.

Add --verbose to task-queue run to print each Fastmod result and provider request as it happens. --quiet suppresses that progress output.

API

POST /api/score_post is retired and returns 410; it cannot bypass the private worker pipeline.

• scored results contain named rubric scores in the inclusive 0–1 range.

• excluded is a deterministic, heuristic-only outcome with no probability.

• disabled, config_missing, provider_error, and invalid_response are explicit non-score outcomes.

Operator commands

./v3 fast-score status is the operator overview. It separates regular new-content work from historical backfill work, shows retained outcomes and batch progress, reports reserved estimate against the batch cap, and gives the next queue action. Regular work intentionally runs before historical backfill. Use ./v3 fast-score status --verbose for recent individual work rows; hashes are shortened there to keep the output scannable. retry and invalidate require an explicit post ID, content hash, and rubric revision; they do not support historical backfill. smoke --post-id=... makes one deliberate provider request without creating a score row or corpus sweep. prune removes Fastmod rows and fast_post_score exchange records older than one year by default.

Historical audit and backfill

Historical content is never included unless an operator explicitly requests it:

./v3 fast-score audit --include-existing
./v3 fast-score backfill --include-existing --confirm --max-posts=100 --max-cost-usd=0.10
./v3 task-queue run --limit=1 --score-limit=25 --work-limit=250

audit is read-only. It reports current-content/current-rubric state counts, the selected model, and a cost estimate using recorded usage for that model or a conservative fallback. It writes no score, work, task, or exchange rows.

backfill snapshots only currently unrated candidates into a private batch. Both --max-posts and --max-cost-usd are required; the lower resulting limit wins. The worker reserves the batch's estimated cost before each provider attempt, including a retry, and stops a batch with budget_exhausted when the remaining reservation cannot cover another attempt. fast-score status shows recent batch IDs, state, and reserved estimate. Inspect exact provider payloads and token usage through the existing private LLM-exchange page.

Historical backfill work is isolated from ordinary new-content sweeps. It is private, advisory, and retained/pruned with the same one-year Fastmod policy.

Model exchanges use the existing private LLM-exchanges audit surface. Each recorded score is labeled independently on its public /posts/{id} detail page; the page reads it from the private score database while rendering, with legacy/earlier-revision markers where applicable. Scores are not added to thread cards, APIs, canonical records, or the read model, and do not make an automated moderation decision. Optional meta tagging publishes only the resulting root label using existing presentation; probabilities and evidence remain in private storage and are excluded from public backups.

Deployment, recovery, and rollback

Deploy code and prompt files together, including all selected templates. The private Fastmod store migrates automatically and preserves old single-score rows as legacy moderation-only results. The core read-model schema is unchanged. Custom prompt overrides need a CSV introduction; generated legacy default paths are translated automatically. Preserve private score and settings files across application releases. No automatic backfill or new scheduling trigger is added.

Inspect configuration errors in the panel and ./v3 fast-score status; correct missing/empty templates or permissions before retrying. For a malformed manually edited selection, restore a valid nonempty JSON array in the private settings file (or remove it to restore configured defaults). Provider and invalid CSV responses use existing bounded delayed retries and explicit operator retry. Do not interpret failures as a score of zero.

Before reverting application code, stop the worker and preserve a backup of the private score/settings/task databases. Older binaries understand only the former single-score contract and must not execute new-format queued work. Restore a matching prior private-state backup or keep workers stopped until the compatible release is restored; never relabel new score maps as one legacy probability.

Verification: php tests/run.php and node tests/browser/fastmod.mjs. The browser check uses temporary private state and a deterministic local HTTP provider to exercise signed selection, normal publication, worker evaluation, omission, multiple scores, malformed output, and explicit retry. It makes no external provider requests. Real-model rubric accuracy remains an operator evaluation.

Automatic root meta tagging

Enable Automatic Fastmod tagging in Tools → Feature Flags → Fastmod, with scoring/ordinary enqueue enabled and the meta rubric selected. One shared FASTMOD_AUTOMATIC_TAGGING_ENABLED flag controls all automatic tag mappings; this release includes only the root meta mapping. FASTMOD_META_THRESHOLD defaults to 0.90, inclusive. Only newly published local roots qualify; replies, imports, old work and backfill do not. Existing tags are stable across configuration/model/rubric changes.

Use ./v3 fast-score meta-status, meta-remove --post-id=ID, meta-retry and meta-rollback for shell-based operation. Removal suppresses reapplication; rollback preserves human tags. Disabling stops additions without removing tags. The enable switch reuses existing Feature Flags; no tag styling, mapping editor, restore workflow or automatic reclassification is added. Shared environment override takes priority over a saved site value, then legacy FASTMOD_META_TAGGING_ENABLED is a compatibility fallback. A saved shared false overrides legacy true. See the runbook for backup, artifact freshness and interrupted recovery.