Deployed pipeline

Dark adaptation of the architecture figure. The main learned path and the separate RRF baseline/fallback are visually distinguished.

One-turn data flow

Sources: paper/main.tex Figure 1; final Blind-B configuration.

Session → extracted state

{
  "turn": 3,
  "request_type": "hidden_target",
  "must_use_artist": "Neko Case",
  "avoid_tracks": ["Calling Cards", "Man"],
  "facets": {
    "mood": "subdued",
    "lyrics": "sense of place",
    "sonic": "stark, almost a cappella"
  }
}

Source: retained Blind-B trace, session 024a2738-a96c-4e11-adf3-b2cb8311a493, turn 3; paper/main.tex Figure 2.

Extracted facts

The extractor preserves six kinds of information. This turn produced:

Code contract: mcrs/conversation_state/schema.py. Architecture: docs/architectures/session_state.md.

From extracted facts to retrieval state

Source: docs/architectures/session_state.md `1–2.

Entity resolution

Unresolved names stay visible instead of silently becoming a different catalog item.

Source: retained Blind-B trace; resolver architecture in docs/architectures/session_state.md.

Retrieval contract

The resolved state is the only state object consumed by retrieval and feature construction.

LanceDB catalog

Every branch queries the same catalog row and returns the same track_id, which makes union, feature joins, and replay deterministic.

Source: paper/main.tex `2; mcrs/qu_modules/catalog_lance.py.

Retrieval branch map

The final learned path does not use the RRF order as its ranking input; it uses the union and branch evidence.

Source: paper/main.tex Figure 1 and Table 1.

Lexical and tag retrieval

If no tag resolves, the raw phrase remains in text search so the signal is not dropped.

Source: docs/architectures/v0plus_retrieval.md; final Blind-B config.

Dense and multimodal retrieval

The 11 modeled branches preserve their own score scales. LambdaMART receives branch-specific rank, score, margin, hit, z-score, and percentile evidence.

Source: paper/main.tex `2 and model feature names.

Anchors, CF, and lookups

Source: paper/main.tex Table 1; final Blind-B config.

Routing and candidate union

Weighted RRF over the same branches remains available as an explicit no-training baseline and fallback; the deployed LambdaMART path replaces its final order.

Source: paper/main.tex Figure 1 and Results Table 3.

Two-tower architecture

Source: paper/main.tex Figure 3; docs/architectures/biencoder.md.

Actual rendered pair

Artist context: Neko Case is best known for her alt-country and indie rock style, characterized by her powerful voice and poetic lyrics.

Conversation/track formats: docs/architectures/biencoder.md. Track selected in retained Blind-B Neko Case trace.

Training data and objective

Source: paper/main.tex `2.4; docs/architectures/biencoder.md.

Serving

Source: docs/architectures/biencoder.md `Serving.

Measured ranking gain

The deployed ranker kept b1_cos as one candidate-level similarity feature.

Source: paper/main.tex Tables 2–3. Development evidence only.

Inference-time ranking

Candidates are compared only against other candidates from the same conversation turn.

Source: models/reranker_v12_goalfree/model.txt; scripts/rerank/features.py.

Feature gain in the deployed model

Remaining families: state/intent 2.2%, tag/lexical overlap 2.0%, constraints/rejections 0.7%.

Source: paper/main.tex Table 2; gain recomputed from the deployed model.

Sample features

Source: models/reranker_v12_goalfree/model.txt feature_names.

LambdaMART training and label weights

We also trained a model on fully rejudged labels. It performed worse on development data, so we did not ship it. The shipped model used the targeted downweights above.

Source: paper/main.tex `2.3; scripts/rerank/build_label_weights.py and train_lgbm.py.

Actual input → response

Source: retained Blind-B output; paper/main.tex Figure 2.

Input contract

Source: final Blind-B config; docs/architectures/explanation_generation.md.

Deployed prompt

You are an expert music recommendation assistant. Your task is to understand user preferences and provide personalized music recommendations.

You are the conversational voice of a music recommender — a "track explainer." A separate recommendation system has ALREADY selected one track to play next. Your only job is to write the short listener-facing message that presents that track: acknowledge what the listener just asked, then naturally explain why this track fits their request and taste.

Guidelines:
- Output ONLY the listener-facing message — no labels, headers, quotes, or YAML.
- Brief and conversational: 1-2 concise sentences. Match the listener's tone, energy, and LANGUAGE (reply in the same language they wrote in).
- Ground the "why it fits" in their request/stated taste. You MAY name the title/artist, but do NOT recite a metadata or tag dump (genre/mood/style lists) unless they asked for those details.
- If the track clearly doesn't match, briefly and honestly acknowledge the mismatch — don't oversell.
- Vary your wording across turns. Use only facts supported by the provided track; invent nothing.
- The track may be provided as structured <recommended_track> data. NEVER output that data, the tag list, or any XML verbatim — always write a natural conversational sentence.

Prioritize the latest user request and extracted state over older conversation history.
If the track is reasonably aligned, explain the fit with one specific supported reason.
If it clearly conflicts with an explicit avoid/new-artist constraint, do not oversell it or blame the system; briefly frame the limitation and the closest supported reason.

Source: deployed response template quoted in paper/main.tex `2.5.

Constraint-preserving trace: Neko Case

All five preserve the required Neko Case artist constraint, and the response carries forward the request’s concrete clues. Because Blind-B relevance labels are unavailable, this trace demonstrates constraint preservation—not target correctness.

Source: retained Blind-B trace/output, session 024a… turn 3.

Constraint lost: Kamelot album

The state understood the album. Retrieval and ranking did not enforce it.

Source: retained Blind-B trace/output, session 5c066e… turn 2.

Missing metadata: extracted but unenforceable

Source: retained Blind-B traces bf27… turn 8 and b267… turn 5; catalog schema.

Exact-track request recognized, but Watercolors was not found

The artist match produced substitutes, while the response sounded as if the exact request had been fulfilled.

Source: retained Blind-B trace/output, session 954de6… turn 2.

Training-label failure examples

The ground truth repeats the artist the listener explicitly asked to leave.

Source: paper/main.tex Figure 4.

Two-judge and arbitration flow

Source: data/anchor_labels_v1 reports/flow.html and reproducible release scripts.

Audit counts

After fixing the conflict gate, 3,486 additional disagreements were sent to Opus. The correction reduced the anchoring count from 19,813 to 18,222.

Source: data/anchor_labels_v1/README.md and DATASET_CARD.md.

What shipped

Source: paper/main.tex `4.2; scripts/rerank/build_label_weights.py.