Classification boundaries — evidence and remaining work

Bounded review, 5 October 2026. Two residual meaning choices are prepared. No current assignment, migration or complete pipeline has been certified. Named feature handling applies the existing Products model rather than requesting the relationship again.

History and scope

Read the current Products and Moments chapters, the classification proposal and its promotion definitions. The historical proposal calls the three OC6 pairs “same-idea pairs.” Its own descriptions differ: Chat Interaction describes an exchange, while AI Chatbot Interaction specifies a participant; Data Analysis specifies a task, while AI Analysis attributes interpretation to AI. Those differences need preserving or explicit rejection, not an assumed synonym mapping.

Focused searches of the current ruling registry, Decided, the July 21 replies, August 11 records and the creative-taxonomy master did not locate pair-specific approvals for Chat Interaction or Data Analysis. The July source mentions the overall AI-analysis pipeline; that is not an approval of these moment definitions. This is bounded history recovery, not a claim that every conversation was available or searched.

Products already specifies feature identities, product-to-feature connections and content demonstrating features. The exact Heel Max Air Unit assignment remains an evidence and implementation question. OC2's other candidates are not closed by applying this principle to the named feature. OC6/active-use remains unprepared; OC7 depends on framing. The earlier Format approval and scoped shared-name approval remain intact; neither approves the individual mappings discussed here.

Independently checked sources

• Intercom's chat guide, particularly escalation guidance and handover troubleshooting, distinguishes Fin from human teammates within a conversation. It supports a participant distinction, not a claim that our collected clip implements it.
• OpenAI's data-analysis guide describes analysis of uploaded data. It supports a task example with AI involvement; it does not establish any specific stored moment's correctness.
• Nike's LeBron 17 page names heel Max Air and forefoot Zoom Air. This is evidence of the publisher's product-feature description, not independent substantiation of its performance claims.
• W3C SKOS Reference, semantic relations and mapping properties: broader, related and equivalent concepts are distinct relationships. This is a useful conceptual check on a proposed merge, not a recommendation to adopt RDF or SKOS storage.

Retaining related descriptions can preserve meaning but requires explicit search relationships. Simplifying into a task or visible interaction plus evidenced AI involvement avoids multiplying every task by every participant type. A selective change is preferable to blind synonym replacement. Exact storage and all consumers remain for the implementation contract.

Selected live checks

Literal SQL and returned results are saved in docs/reference/2026-10/2026-10-04/classification-boundaries-live-inspection.json. Column probes preceded the selected joins. The receipt includes one unused OR branch with an unverified dummy identifier; it supplies no evidence and is excluded from findings.

The selected rows join content_segment to content_segment_classifications and their parent content_item. The parent playback locations returned in gcs_path; an empty gcs_video_path is not evidence that media is missing. The selected classification rows have analyzer provenance. Their labels, analyzed status and timestamps do not prove semantic accuracy or that the media revision matches the one used for classification.

Media inspection log

These observations were made from browser-rendered sampled frames and the video element's current time. They are not continuous playback, audio inspection or full interval-boundary verification. Exact rows and paths are in the SQL receipt; observations are also saved in classification-boundaries-media-observations.json.

Collected video	Stored description being checked	Observed sample
Introducing the LeBron 17	Heel Max Air Unit Force Absorption, 9.033–11.967 seconds	Samples around 10–11 seconds show a component and the named feature caption. Supports the name appearing; performance and full boundaries unverified.
Nike Playlist, Emma Raducanu and Soyeon	Tennis Swing with VFX, 123.8–125.8 seconds	At 124.216 seconds, a dance studio and smiley overlay with “RELAXATION” appear. The sampled frame does not support the stored tennis description.
HiBob turned AI into a flywheel	Chat Interaction, 5.54–11.09 seconds	At 10.207 seconds, a chat interface discusses AI use cases. Participant identity and the whole interval are not established by the frame alone.
Replying to @internal_glowup	AI Coverage Feedback, 127.68–133.23 seconds	At 128.257 seconds, the source presents English-to-Luo translation. This does not support the makeup-coverage description at that timestamp.
Sales productivity: working smarter	Typing on Laptop, 81–86.833 seconds	At 83.205 seconds, a speaker gestures behind a laptop. This is not the described hand close-up; typing elsewhere in the interval remains unchecked.

Two additional samples were non-supporting: the Nike poncho interval sampled at 109.2 seconds showed a graphic transition; the R&D video at 392.223 seconds showed a race pit stop. Neither sample proves an activity absent throughout its interval. The R&D classification lead comes from an earlier receipt rather than this turn's selected segment query and is not promoted into a fresh row-level finding.

Confirmed limit: these examples cannot establish synonym equivalence or lossless conversion. Root cause unresolved: possible classification error, temporal mismatch, media revision or selection mismatch must be tested rather than guessed. Retain original evidence and human corrections while investigating; no production record was changed.

Local dependencies inspected

Read the relevant blocks in backend/routers/pipeline_scene_classify.py. _merge_adjacent compares family and moment type plus embedding similarity on the llm_detection fallback. _dedup_adjacent_beats, in the locally documented preview path, groups by moment category and moment type. These are selected local consumers, not a full deployed trace. Changing labels may affect interval grouping as well as text shown in a filter.

Before implementation, find every caller and inspect the deployed path. Trace source media identity and revision → cut list and timestamps → model input and composed prompt → returned descriptions → validation and normalization → segment/classification writes → thumbnail/playback selection → queries and saved references. Compare actual media duration and timebase at each step. Test reruns, correction precedence, changed media, failed writes and unsupported feature links.

Human and technical relationships to preserve

A task, the participant performing it, the product offering it, and a publisher's claim about it are different assertions. More than one participant or product can be relevant. Unknown AI involvement cannot become a negative or positive fact by default. A captured app is not automatically the depicted product, and a feature mention is not automatically a verified demonstration or campaign membership.

For schema review, inspect stable identities, relationship direction, multiplicity, required values, uniqueness and foreign keys before choosing physical fields. Inspect junctions and indexes against actual query patterns. A narrower AI-chat browse path must include only supported matches and preserve the evidence scope; it must not turn a whole video into a chat or AI-analysis piece. Equivalent collection and tracker rules must yield equivalent authorized results.

For workflow review, preserve the taxonomy version, media identity, evidence and human changes across reanalysis, merges, corrected boundaries, withdrawals and missing media. Scheduling can refresh evidence but cannot silently upgrade an inference. Costs, activation gates and delivery remain part of the comprehensive audit. This packet changes no schema, data, prompt, API or automation setting.

Queue mapping and completion

• OC6/chat: prepared recommendation; no answer or conversion approved.
• OC6/analysis: prepared recommendation; no answer or conversion approved.
• OC2/heel_max_air_unit: existing feature model applies; exact identity, assignments and replacement descriptions remain to verify. No new capability question.
• OC6/active-use, the other OC2 candidates, and all unrelated promotions remain open in their existing obligations.

The inventory records this bounded evidence and the remaining media investigation. Readiness requires independent challenge, rendering checks and hosted verification. The complete decision pass, full Docs audit, production repair and product acceptance remain separate milestones.