Processing Rules

Language Policy — English Only

All analysis output, rhetoric summary updates, tracking DB entries, and profile data MUST be written in English regardless of the talk's delivery language. For non-English talks:

• Verbatim quotes: ALWAYS write English translation FIRST, then the original in parentheses. Never the reverse. Format: "English text" (оригинальный текст). Example: "That's the whole point" (В этом весь смысл) — NOT "В этом весь смысл" (That's the whole point)
• Evidence-citation exception: evidence_citations[].quote is the machine-verification field and MUST contain only the exact source-language span. Put its English rendering in evidence_citations[].translation; renderers show the translation first and label the original. Keep the human evidence summary in English.
• Verbal signatures: store separately tagged with language code (e.g., [ru] "получается что") — do NOT merge into the main English signature list
• Slide text: translate in the analysis, note original language
• Humor/wordplay: note when a joke is language-dependent and untranslatable
• Tag the talk entry with delivery_language in the tracking DB

Pattern Taxonomy and Generation Recovery

Run "{python_path}" "{speaker_toolkit_root}/skills/vault-ingress/scripts/queue-state.py" <tracking-database.json> normalize before claiming work. The command owns both legacy source-status migration and pattern-generation recovery as one copy-on-write transaction. It uses partition_pattern_scoring_cohort from skills/vault-ingress/scripts/adherence_baseline.py; do not duplicate that selection logic or approximate it with processed_date.

Every valid processed/processed_partial result excluded from the active generation is moved to needs-reprocessing. The stored machine reason is pattern_scoring_generation:<reason-code>[+<reason-code>...], preserving the selector's ordered codes, and the same codes and observed/expected generation identity appear in the command's normalizations JSON. That gives every clean consumer exclusion a deterministic queue path instead of leaving the current cohort permanently empty.

Malformed or unknown generation identity, a current result with non-empty generation reasons, incomplete current identity, and invalid or divergent current score lanes reject the whole command with no DB write. Inflight, pending, already-queued, and skipped records remain outside generation recovery. Repeating normalization after a successful recovery is byte-stable. Existing completed claim and history evidence is preserved; the next ordinary queue claim archives the prior current claim under the normal generation transition.

Pattern Tagging Rules

Scan observations against the pattern taxonomy index at skills/presentation-creator/references/patterns/_index.md (path relative to plugin root). Skip every pattern marked observable: false. These include hidden preparation, provenance, decision, and post-event processes as well as behavior that the available artifacts cannot establish. A polished outcome is not proof that a named process produced it.

For every other entry, inspect evaluable_from, optional strong_evaluable_from and absence_evaluable_from, the required-together not_applicable_when / applicability_evaluable_from contract when present, evidence_requirements, and not_evaluable_when. The allowed evidence-source values and their limits are defined in the index's Evidence-Source Contract. A strong detection uses strong_evaluable_from (defaulting to evaluable_from); moderate and weak detections use evaluable_from. Only score an entry when an available eligible source establishes its requirements. Every detected pattern or antipattern must record concrete evidence and the qualifying evidence_source. When that source is source_comparison, also return the duplicate-free evidence_sources_used array. It must exactly equal one qualifying all-of group, while the prose evidence names what was compared.

Return v4/v5 makes "inspected" an artifact-bound statement. Alongside the exact evidence_sources set, return one closed raw source_inspection record per underlying source:

• transcript uses one or more inclusive, ascending line_ranges.
• static_slides and native_deck use inclusive, ascending page_ranges.
• delivery_video uses ascending [start, end] second time_ranges, with end > start.
• Each distinct source_comparison group uses its duplicate-free evidence_sources_used plus comparison_scope: "full"|"partial". Multiple comparison records are valid when their exact underlying groups differ.

Ranges may be adjacent but may not overlap. Coverage is complete only when the verified artifact begins at line/page 1 (or video second 0), ends at its exact verified bound, and has no gaps. A full comparison is complete only when all of its members have complete coverage; a partial comparison never authorizes an undetected outcome. The worker owns the raw ranges and scope. Persistence owns the resolved counts/duration, coverage_complete, artifact identities, and comparison identity bundle.

Each observable entry declares evidence_channels. Each detection returns a non-empty evidence_citations array through one of those channels. Use the citation shapes in schemas-db.md Pattern Evidence Citation Schema. The citation must locate proof from the qualifying source: transcript evidence uses a transcript locator, static/native slide evidence uses a slide or slide-sequence locator, and delivery-video evidence uses a video interval. A source_comparison detection supplies citations for every underlying member of evidence_sources_used. Metadata may supplement a detection but cannot replace the source/outcome gate. Timing, sequence, motion, and delivery claims use their specific timed-transcript, slide-sequence, or video locators. Put hypotheses without allowed source-located proof in clarification notes, not the score.

For an entry with no positive detection, use absence_evaluable_from (defaulting to evaluable_from) to decide whether completely inspected sources can support an undetected outcome. Return v4/v5 uses no prose waiver: every not_evaluable item contains exactly pattern_id and reason_code. Use missing_required_source_coverage when no effective absence group has complete coverage. Use absence_not_authorized_by_catalog when an explicit absence_evaluable_from: null makes the entry positive-only; this is intentional catalog policy, not unfinished owner work. Use source_gate_pending_owner_review only when the observable catalog entry has no owner-approved positive gate. That pending entry fails closed: it cannot be detected by a v4/v5 return and cannot be silently counted as absent. A valid positive detection takes precedence for a gated entry; never add the same ID to not_evaluable. Do not guess and do not interpret not_evaluable as absence. Exclude not-evaluable entries from the score. Persistence recomputes the exhaustive expected ID→reason map and rejects missing, extra, duplicate, prose-bearing, or blanket waivers.

Return v5 additionally makes applicability exhaustive. For every nondetected entry with not_applicable_when, first evaluate the complete applicability_evaluable_from gate. Without complete canonical coverage, an assessment is forbidden and the outcome is not_evaluable with missing_applicability_source_coverage. With complete coverage, exactly one applicability_assessments row is mandatory. It contains pattern_id, result, evidence_source, nonempty evidence, source-located evidence_citations, comparison-only evidence_sources_used, and a catalog-authorized condition_id only for not_applicable. An applicable assessment forbids condition_id and then proceeds through the ordinary absence gate; there is no implicit applicable default.

Persistence owns the exhaustive v5 projection. It writes exactly one sorted pattern_outcomes row per observable entry using precedence: detection; validated applicability assessment; incomplete applicability/absence gate as not_evaluable; applicable plus complete absence gate as undetected. Outcomes are exactly detected, undetected, not_evaluable, or not_applicable. The worker never returns this ledger. Persistence also hashes scoring schema, catalog fingerprint, and sorted per-pattern opportunity state into opportunity_coverage_identity; detected/undetected collapse to evaluable, while the two unavailable states remain distinct.

This is exhaustive for source gates: every undetected observable catalog entry for which no effective absence alternative is satisfied by complete, canonical inspection coverage must be represented in not_evaluable. A singleton alternative needs both complete ranges and absence-capable provenance. A full source comparison remains positive evidence but cannot authorize absence or applicability until a future canonical receipt proves aligned modality capture; mere artifact coexistence is not comparison work. Artifact scope still controls what counts as a source. In particular, an untrusted video full_frame_context may support concrete delivery_video observations but never creates static_slides or native_deck evidence.

A trusted schema-v3 video-extracted slide_region PDF is a positive-only static source. Its identity-bound pages may support citations and detections, but the sampling, transition filtering, and deduplication receipt does not prove that every delivered visual state survived. Therefore even full inspection of that PDF does not join the absence/applicability-complete source set. Bare native_deck and delivery_video are positive-only for the same reason: page ranges or full duration do not prove audience/screen/audio/session-boundary capture. Native PPTX and rendered static pages are distinct too: PPTX inspection establishes native_deck, never static_slides; a separately declared readable PDF retains its own static identity and may be absence-complete.

Canonical inspection rows expose both facts. coverage_complete reports only range coverage. Engine-owned absence_capability_complete separately gates negative/applicability inference, and absence_capability_reason explains the decision with a stable code such as authorized_transcript, authorized_rendered_static, nonexhaustive_video_extraction, bare_native_deck, bare_delivery_video, or comparison_alignment_unverified.

persist-results.py validates catalog ID/type, bucket, uniqueness, observability, source/outcome gate, channel, quote, slide range, declared inspection coverage, and available artifact context before writing. Raw transcript citations contain source, channel, quote, and optional translation; raw slide citations add slide_numbers; raw video citations add start_seconds/end_seconds; raw metadata citations add field. Workers do not return transcript lines/timestamps, artifact roots/paths/hashes, metadata value/owner_value_after_return, coverage_complete, derived counts/duration, timing/quality receipt identities, comparison artifact identities, enriched not-evaluable facts, or evidence_schema_version. Those are engine-owned canonical fields. Catalog dimensions are also engine-owned and should be omitted; a compatibility copy is accepted only when it exactly matches catalog order.

Transcript Quality and Timing Authority

Treat the readable transcript and its two receipts as three separate artifacts:

• transcripts/<id>.txt is the exact UTF-8 speech text.
• transcripts/<id>.segments.json schema v2 owns owner-bound acquisition source and optional timing.
• transcripts/<id>.quality.json owns the exact validation policy and the source of any duration that lowered the fixed short-artifact floor.

Both receipts carry SHA-256 of the exact .txt bytes. Verify against raw bytes, not newline-normalized text: replacing CRLF with LF invalidates both even when the decoded words are unchanged. Missing or rejected timing leaves ordinary transcript quotation available but cannot support timed_transcript. Quality is independent: a transcript with no timed segments can and must still carry a current quality receipt before it enters v5 scoring.

The quality policy is exactly {schema_version, min_words, duration_seconds}. A caller's --min-words may tighten the derived floor but never lower it. With no trusted duration, the floor remains 400 words. A lower short-talk floor derives only from yt-dlp provider duration for the exact YouTube ID or ffprobe over exact local media, whose digest is stored in the provenance. --duration-seconds is an expected value that must match that source-owned probe; it is not authority itself. Return fields, analysis prose, and unbound talk metadata never lower the floor.

Current v5 persistence requires a hash-current receipt with exact provenance. For youtube_duration, the receipt video ID must equal the owning talk's youtube_id. For local_media_duration, the stored media digest must equal the exact owner-bound local media. A missing legacy receipt is unverified and must be requeued through fetch-transcript.py; malformed, stale, wrong-owner, or duration-drifted receipts fail closed. Never copy a policy or duration from a worker return.

Timing schema v2 is closed and source-artifact-bound. YouTube captions/Whisper require the exact owner video ID and trusted duration; local Whisper requires the exact media digest and trusted duration; VTT requires a safe relative regular-file path, exact artifact digest, and exact final cue extent. Joined segment text must equal the transcript modulo Unicode whitespace layout, and time ranges must fit the source bound. Legacy schema v1/minimal receipts are archival: never infer missing ownership or migrate them by relabeling.

Caption timing enrichment for valid existing text is non-destructive. Pass the owner's provenance via --existing-source; only known youtube_auto text may acquire fetched caption segments, and only when caption text is identical after Unicode-whitespace collapse. The script writes only the timing sidecar and never relabels or overwrites manual, Whisper, unknown, or text-mismatched transcripts.

An existing transcript is validation-only unless --force explicitly authorizes replacement. Tightening --min-words can reject it but never licenses a provider overwrite. A caught bundle failure restores the prior transcript and receipt bytes. On a fresh/forced fetch, invalid optional segment timing degrades to unavailable and removes stale timing transactionally; valid semantic text and its quality receipt still commit.

Structured Field Extraction

The subagent's job is to return every structured field it identifies (co-presenter, delivery language, slide counts, opening/closing types, etc.) in the structured_data block per the return schema — never to leave them buried only in rhetoric_notes free text. If it's in the analysis, it must be in structured_data.

Persisting those fields is deterministic and script-owned, not a manual per-run mapping — SKILL.md Step 4 uses {speaker_toolkit_root}/skills/vault-ingress/scripts/persist-results.py for the merge. Authors do not re-derive that logic here.

Adherence Assessment

adherence_assessment measures how consistent a talk is with the speaker's established rhetorical baseline — not whether the talk was good in the abstract. Adherence is consistency with this speaker's own validated style, which is why it can only be computed once a baseline exists.

Authority: for return schema v5, the claim baseline remains immutable, but raw-score comparison also requires an exact matching canonical opportunity_coverage_identity. The worker cannot author that engine-owned identity. Therefore the exact empty adherence sentinel is always safe; any owner-side structured comparison must prove identity equality. Workers MUST NOT parse Section 15, infer a date cohort, or recompute an average from the live DB. Every member of one batch carries the same snapshot.

Worker gate: return exact adherence_assessment: "" and omit adherence_comparison. Canonical talk identity does not exist until owner-side persistence, so a worker cannot prove the comparison predicate.

Owner-side gate: inspect baseline comparison status, identity, and counts exactly.

• raw_score_comparison_status: unavailable, a null identity, an identity mismatch, or fewer than 10 scored_talk_count: do not construct a comparison.
• A structured comparison is valid only when the canonical talk identity equals the baseline identity and the baseline contains at least 10 scored talks. It carries a value-for-value copy of the immutable baseline and the validated talk score.

eligible_talk_count remains the complete fresh generation cohort for per-pattern opportunity denominators even when mixed identities suppress the raw-score lane. scored_talk_count is only the exact one-identity score cohort.

This is a global, generation-bound comparison. The baseline includes only processed/processed_partial talks stamped current with the exact catalog fingerprint and pattern-scoring schema captured by the claim. An unscored talk cannot be assessed, and a stale catalog/scoring generation requires recovery and a fresh claim rather than reinterpretation.

Three checks, in order:

1. Numeric anchor — interpret the validated pattern_score against the claim baseline's average_pattern_score and scored_talk_count. The renderer generates this anchor mechanically from adherence_comparison; the worker's prose need not restate the numbers.
2. Current-talk evidence — interpret that difference using the patterns and antipatterns detected in this return. Name a detected antipattern when one materially explains the score; do not invent population frequency that the baseline schema does not carry.
3. Departure classification — use claim/talk context and confirmed intent, when present, to distinguish a deliberate mode, co-presenter, or venue choice from likely backsliding. Context may explain the number but cannot replace or modify the claim snapshot.

Required interpretation: the assessment explains the mechanically generated anchor using current-talk evidence. Validators deliberately do not parse prose for numeric agreement. If the prose happens to repeat a number, that number is untrusted narrative; the structured comparison and renderer-generated anchor remain authoritative.

Bound: 2–4 punctuation-terminated sentences of prose, not a second score. Enforcement is deterministic: every ., ?, or ! punctuation cluster followed by whitespace or end of text is one sentence boundary, including a period in an abbreviation; the final sentence must be terminated. Spell out abbreviations that would create a false boundary.

Non-empty adherence prose from a return v1–v4 artifact remains replayable only as archival legacy-unverified text. It is never a verified numeric comparison, never enters a current baseline or Section 15 aggregate, and is never profile input.

Rhetoric Summary — Improvement & Adherence Sections (15–16)

rhetoric-style-summary.md Sections 1–14 mirror the 14 analysis dimensions. Sections 15–16 are cross-talk narratives. Rebuild Section 15 in Step 5 only after the entire batch has persisted successfully; never update it after an individual member merge.

Section 15 — Improvement & Adherence Baseline

Section 15 is a human-readable account of the verified current cohort, not the numeric authority for a worker. persist-results.py stdout supplies the post-batch current_adherence_baseline only after every merge succeeds. Treat that emitted object as the complete baseline contract. Do not reproduce its cohort selection, arithmetic, or unavailable-state rules in Section 15 prose.

The Section 15 current block is owned by skills/vault-profile/scripts/section15_pattern_history.py. Do not reproduce its cohort filter, policy resolution, classification, availability, or compatibility rules in this reference. To inspect a summary, run:

"{python_path}" "{speaker_toolkit_root}/skills/vault-profile/scripts/section15_pattern_history.py" assess \
  "{vault_root}/rhetoric-style-summary.md"

assess emits {current_contract, catalog_fields_available, classification_fields_available, available_classification_domains, block_schema_version, policy_semantic_sha256, scored_talk_count, eligible_talk_count, reason_codes, errors}. Exit 0 authorizes only the fields and domains named by that output. Exit 1 means the block cannot authorize current catalog history. Use it only as narrative context. Never derive a classification or availability decision from Section 15 prose or raw occurrence counts.

Old recurring/signature/underuse/resolved prose may remain outside the delimited block only when explicitly labeled historical or manually curated. It is non-baseline narrative and must not be regenerated from occurrence rates or consumed as current catalog classification.

Section 15 is the human-readable mirror of the profile's validated pattern_profile occurrence, policy, classification, and audit lanes (see ../../vault-profile/references/speaker-profile-schema.md). Keep the two consistent by passing the complete profile-pipeline candidate to the owner script. Section 15 prose and legacy adherence text are not machine-readable numeric inputs.

After the complete post-batch narrative and pattern_profile candidate are ready, run "{python_path}" "{speaker_toolkit_root}/skills/vault-profile/scripts/section15_pattern_history.py" replace with the summary, candidate, and live tracking-database.json:

"{python_path}" "{speaker_toolkit_root}/skills/vault-profile/scripts/section15_pattern_history.py" replace \
  "{vault_root}/rhetoric-style-summary.md" pattern-profile-candidate.json \
  "{vault_root}/tracking-database.json"

replace is the only supported current-block writer. On exit 0, it emits {path, changed, scored_talk_count, eligible_talk_count, catalog_fields_available} and writes the validated block atomically when changed is true. On exit 1, it emits a diagnostic to stderr and makes no write. Pass the complete candidate copied from the profile pipeline. Do not hand-filter, classify, or repair it. The command re-analyzes persisted data without reparsing talks or mutating tracking or raw opportunity rows. All prose outside the block remains historical/non-baseline.

Section 16 — Speaker-Confirmed Intent

Patterns the speaker confirmed as deliberate or accidental during a clarification session (see vault-clarification). Read-only during ingress — populated by clarification, consumed here as the intent-adherence input for Section 15.

Improvement Goal Verification

Each ingress run, after the final Section 15 baseline is current, verifies the speaker's active improvement_goals (set during clarification — record schema in ../../vault-clarification/references/schemas-config.md). This closes the loop: the system stops merely diagnosing and checks whether the issue the speaker chose to work on actually moved.

Before calculating any metric, run "{python_path}" "{speaker_toolkit_root}/skills/vault-clarification/scripts/goal_generation_provenance.py" with the complete active-goal array and the structured post-batch full-cohort pattern baseline. The script emits one assessment with a stable decision and reason_codes per goal; exit 1 blocks all goal writes. Require exactly one assessment for every active goal before constructing a mutation plan. It is the sole authority for generation comparability—do not reproduce its fingerprint/schema predicate and never parse Section 15 prose as its baseline.

• A comparable assessment authorizes the metric and outcome rubric below.
• A schema-v1 antipattern or underuse goal is historical, report-only unverifiable. Preserve the complete record and do not create a mutation for it.
• A schema-v1 pacing or other goal may remain comparable through its independent provenance lane. If comparable, patch only current_value, last_checked, checked_by, and status. Never add verification_state or verification_reasons to a schema-v1 record. A non-comparable assessment is report-only.
• For a schema-v2 needs_rebaseline or unverifiable assessment, copy the decision to verification_state and its codes to verification_reasons. In either case, preserve current_value and must not set status to achieved, improving, stalled, or regressed. A speaker-confirmed rebaseline is owned by vault-clarification; ingress never restamps the fixed baseline.
• Pacing and independent goals continue through their own provenance lanes and are not invalidated by a pattern-catalog generation change.

For each comparable goal with status not in (achieved, retired):

• Compute current_value for the goal's metric from the current Section 15 cohort data — and, for pacing and mode-specific goals, from the freshly regenerated speaker profile. For a policy-derived metric, consume only the projection and availability state emitted by the validated classifier payload. Never reconstruct either from raw rows. If that payload does not authorize the requested metric, the result is unverifiable/report-only, not zero. Pacing reads pacing.adherence.over_budget_rate from its independent lane. For schema v1, write only current_value, last_checked (today), checked_by: "vault-ingress", and status. For schema v2, write those fields plus verification_state: "current" and empty verification_reasons.
• Set status by comparing current_value against baseline_value and target:
  • achieved — current_value meets or beats target.
  • improving — moved toward target versus baseline_value but not there yet.
  • stalled — no meaningful movement from baseline_value.
  • regressed — moved away from target (worse than baseline_value).
• Only count talks processed after set_date toward movement — a goal can't be judged on talks that predate it.
• Never overwrite baseline_value, target, issue, or set_date — those are the fixed yardstick; the verification writer changes only the schema-authorized verification and status fields above.

Use one patch_improvement_goal_verification mutation per writable assessment. Each mutation's expect object must contain exactly the fields it sets with the values from the latest strict read. If every assessment is report-only, do not invoke the mutator. Otherwise dry-run the complete multi-goal plan, review it, apply the whole plan against the reported input SHA, and re-read the database. An assessment failure, malformed plan, stale expectation, or changed database generation installs no partial goal update.

Report each goal's status in the run summary. A regressed or stalled goal is the strongest signal to surface — it is the speaker's own priority, not a machine-chosen one.