MOSAIC-Rec

Auditable multimodal sequential recommendation research with exact full-catalog evaluation, frozen model-selection gates, uncertainty-aware diagnostics, and reproducible evidence.

MOSAIC-Rec studies whether a support-aware router can preserve a multimodal content expert's zero-support behavior while retaining the quality of a sequence expert on warm items. Protocol mosaic-rec-2026-09-23-p2 is CLOSED with the permanent decision RETAIN_BASELINE.

What the project studies

The p2 study uses PixelRec50K under chronological FIT_VISIBLE, VALIDATION, QUALIFICATION, and LOCKED_FINAL roles. It compares:

• C_seq: a frozen three-seed BERT4Rec-style sequence ensemble;
• C_cold: a frozen three-seed multimodal Two-Tower content ensemble;
• R4: a frozen three-seed context-aware router with a warm-preservation objective.

R4 combines separately trained sequence and content experts using pre-query support, item-age, user-history, recency, preference, and content-availability covariates. Router weights are model control values, not calibrated reliability probabilities.

The protocol uses exact full-catalog ranking, identical candidate semantics, immutable top-k artifacts, one-time protected-role access, deterministic paired bootstrap inference, and a canonical Parquet/DuckDB evidence warehouse.

Locked primary results

The one-time LOCKED_FINAL evaluation contained 25,367 queries. Overall NDCG@20 is secondary and cannot override either primary gate.

Result	Estimate	Frozen inference	Decision
R4 overall NDCG@20	0.001846	Secondary metric	—
H1: R4 − C_cold, controlled zero support	+0.0008903	Paired 95% interval [0, 0.0026710]	FAIL
H2: R4 − C_seq, warm	−0.0013377	Paired 95% interval [−0.0022765, −0.0005345]; margin 0.0013209	FAIL

Both gates were required for promotion. The final scientific decision is therefore RETAIN_BASELINE. Eleven of fourteen preregistered robustness cells also violated the declared routing-direction expectation.

See the p2 final status, technical report, claim ledger, and adversarial audit.

Posthoc industry diagnostics

POST-HOC EXPLORATORY — NOT A P2 CONFIRMATORY CLAIM

Read-only diagnostics reused immutable locked outputs and did not rerun inference or alter p2:

• 25,367 locked queries and 11,214 target items were analyzed.
• In the torso-popularity slice, R4 − C_seq was +0.003296 NDCG@20 over 8,817 queries.
• In the support >50 slice, R4 − C_seq was −0.004394 over 520 queries.
• Only 1 of 295 controlled-zero queries had a positive R4 − C_cold NDCG difference.
• The experts tied on 99.68% of per-query NDCG outcomes.
• Routing-side agreement was 86.25% among the 80 informative expert non-ties.

These values support future hypotheses and diagnostics only. They do not revise the locked decision or establish model superiority. See the industry diagnostic report, dashboard, and artifact manifest.

Reproducibility

Public CI validates formatting, linting, CPU-safe contracts, synthetic data tests, and pure posthoc helpers. It does not train models or access protected roles.

python -m pip install --no-deps -e .
ruff check src tests scripts
ruff format --check src tests scripts
PYTHONPATH=src:. pytest -q --import-mode=importlib tests
PYTHONPATH=src pytest -q scripts/posthoc/test_diagnostics.py

The full scientific replay requires the authorized public datasets and the external scratch artifacts described in the p2 reproduction runbook. Large datasets, feature arrays, checkpoints, and protected full top-k outputs are intentionally excluded from Git. Checked-in compact evidence, aggregate tables, manifests, and reports retain hashes for auditability.

Repository structure

Path	Purpose
src/mosaic_rec/	Installable package, models, metrics, protocol guards, statistics, and evidence tools
configs/p2/	Preregistered and frozen p2 configuration identities
evidence/p2/	Compact canonical receipts, manifests, aggregate evidence, and warehouse
analysis/posthoc/	Read-only exploratory diagnostic tables and hash manifests
reports/p2/	Canonical closed-study reports and claim ledger
reports/posthoc/	Exploratory slicing, routing, exposure, robustness, and dashboard reports
scripts/	Crash-safe execution, validation, and analysis entry points
tests/	CPU-safe unit, schema, leakage, protocol, metric, and replay contracts

Hardware and systems disclosure

Scientific neural workloads used one NVIDIA RTX 4090 Laptop GPU under Ubuntu 22.04 in WSL. Frozen R4 exact full-catalog scoring measured 667.6 users/s. This is a local, single-GPU systems result; no multi-GPU or production-readiness claim is supported.

Claim boundaries

This repository does not claim:

• production or online impact;
• causal lift;
• statistical cold-start superiority;
• warm retention or noninferiority;
• reliability-aware routing;
• external replication;
• valid canonical ablation conclusions;
• multi-GPU scaling.

Negative and invalid results remain part of the public scientific record. The canonical p2 decision is RETAIN_BASELINE.

Data and licensing

Raw datasets and derived publisher assets are not redistributed. Dataset provenance, terms, citation requirements, derived-feature restrictions, and redistribution limits are documented in the repository's data and rights audits. TMDB content was excluded from model training and feature generation.