Finance/Quant Source Ingestion Queue

See also: finance-quant-ai-objective-coverage-audit-2026.md · finance-quant-ai-coverage-audit-2026.md · buy-side-quant-ai-practitioner-signals-2026.md · …/13-multimodal-sources/finance-quant-audio-video-deep-dive-2026.md

Source ledger: sources/06-industry-verticals/finance-quant-source-ingestion-queue-2026-raw.md

This queue converts the remaining finance/quant source gaps into ingestion work. It is deliberately stricter than a watchlist: an item belongs here only when it can be captured through an existing repo pipeline or converted into a StateBench source-discovery fixture.

Promotion Rules

Priority Queue

Recently Captured Benchmark Sources

Pipeline Commands

PDF/report ingestion:

scripts/.venv-tts/bin/python3 scripts/ingest_pdfs.py --pillar 06
scripts/.venv-tts/bin/python3 scripts/ingest_pdfs.py --pillar 21

Raw PDF files captured under research/**/raw are local source material and are ignored for new ingests. Commit the OCR output (*.chandra.txt or chapter chunks), *.pdf.meta.md sidecars, source-card markdown, and any StateBench fixture entries that prove the source was captured. Some older PDFs are already tracked in git; do not use that precedent for new finance/quant captures.

Podcast/audio ingestion:

scripts/.venv-tts/bin/python3 scripts/podcast_mine.py \
  --show risk-net-quantcast \
  --limit 3 \
  --archive-audio

Archived podcast and conference audio/video stays local under sources/13-multimodal-sources/. Commit only the .meta.md sidecar, ASR transcript in research/13-multimodal-sources/**/raw, promoted episode note, and source-discovery or benchmark metadata.

Finance replay model scoring:

python3 -m statebench.runners.suite_runner \
  --suite-file statebench/suites/finance-replay-v0.json \
  --suite-id finance-replay-v0-<artifact>-<backend>-YYYYMMDD \
  --backend <frontier-api-or-openai-compatible-url> \
  --model <model-id> \
  --policy resume \
  --artifact-id <exact-artifact-id> \
  --artifact-format <api|hf|gguf|mlx|neuron_compiled|awq|gptq|fp8> \
  --artifact-precision <api|bf16|fp16|q4_k_m|fp8|...> \
  --weights-revision <revision-or-snapshot> \
  --backend-hardware <api|apple_m3_max|cuda_h100|aws_inf2|...> \
  --backend-constraint "record backend/operator/context/license limits" \
  --context-length <effective-context-tokens> \
  --preflight-record statebench/preflights/<artifact>-preflight.json

Scorecard summary:

python3 -m statebench.runners.scorecard \
  --results-root statebench/results \
  --suite-prefix finance-replay-v0 \
  --format markdown

What To Capture In Each Source Card

Every promoted source should preserve:

• URL and retrieval date;
• publication date;
• publisher/owner;
• named speaker/author where applicable;
• source type: paper, PDF, podcast, official page, job listing, social post, conference session, competition, repository, or model card;
• claim type: workflow, model, benchmark, governance, data connector, execution, alpha/performance, hiring/talent, or failure mode;
• credibility tier and conflict caveat;
• ingestion path: PDF OCR, podcast ASR, source card, StateBench fixture, or source-discovery-only;
• promotion decision: finding, source-discovery, watchlist, or reject.

Stable Source-Discovery Fixture Ideas

These should become frozen source-acquisition tasks rather than live-web leaderboard runs:

1. Find an arXiv finance benchmark paper and record its PDF, title, authors, date, and benchmark purpose.
2. Find a named quant podcast episode, archive the audio/transcript, and extract only workflow claims tied to a guest.
3. Find an official hedge-fund AI/careers page and classify whether it proves workflow, tooling, hiring demand, governance, or only marketing.
4. Find a Kaggle finance competition and extract metric, public/private split, leakage controls, and official data schema.
5. Find a small stable Kaggle-related PDF or paper, download it to research/21-benchmarks/raw, OCR it with scripts/ingest_pdfs.py --pillar 21, and record why the source is eval-design evidence rather than alpha evidence.
6. Find a vendor finance-agent product page and separate product claims from named-customer architecture evidence.
7. Find a Snowflake/Postgres/NL2SQL source and extract dialect, schema, execution, cost, and permission constraints.
8. Find an NL2SQL benchmark-quality paper and record whether it changes the StateBench SQL harness, the public-leaderboard caveat, or the internal gold query audit process.

The vendor/product promotion boundary is now executable through finance-vendor-product-signal-v0. Use it when a source-discovery model tries to promote QuantConnect-style assistant teams, finance-agent connector pages, incumbent terminal AI, AI-native workflow vendors, named-customer quotes, podcast mentions, or terminal replacement claims.

The first live-source eval queue is now tracked in statebench/suites/finance-source-discovery-eval-v0.md. It records a Better System Trader RSS dry run and marks the discovered episodes as source-discovery until transcript review proves they contain useful workflow or failure-mode evidence.

Validate that queue before promoting any live discovery batch:

python3 -m statebench.runners.source_discovery_eval \
  --queue statebench/suites/finance-source-discovery-eval-v0.json \
  --format markdown

Generate frozen source-acquisition task templates from the queue with:

python3 -m statebench.runners.source_discovery_eval \
  --queue statebench/suites/finance-source-discovery-eval-v0.json \
  --section fixture-templates \
  --format markdown

Check which template families already have source-acquisition task examples:

python3 -m statebench.runners.source_discovery_eval \
  --queue statebench/suites/finance-source-discovery-eval-v0.json \
  --section fixture-coverage \
  --format markdown

The eval now tracks twelve fixture families, matching the current backlog: podcast RSS episode discovery, official firm/careers page discovery, official social-post discovery, negative-evidence discovery, paper/PDF discovery, Kaggle competition discovery, SQL/retrieval source discovery, domain model/runtime discovery, vendor/customer architecture discovery, vendor-product signal classification, finance workflow vendor platform discovery, and SQL benchmark-quality / dialect-portability discovery. This closes the earlier gap where LinkedIn/social evidence, negative-evidence, model/runtime, vendor/product, and SQL benchmark-quality batches were listed in the backlog but were not first-class frozen-fixture templates.

Newly Captured Gap Items

• BizFinBench business-driven financial benchmark - arXiv 2505.19457, submitted 2025-05-26 and retrieved 2026-06-05. Captured through the pillar 21 PDF/OCR pipeline as research/21-benchmarks/raw/bizfinbench-business-driven-financial-benchmark-2505.19457.chandra.txt and promoted through sources/21-benchmarks/bizfinbench-business-driven-financial-benchmark-raw.md. Promotion: multilingual/business-finance benchmark-design evidence, not current model-ranking proof. It strengthens the regional-language, numerical-calculation, financial-time-reasoning, event-attribution, tool-usage, and LLM-as-judge lanes by requiring task-level decomposition, Chinese/local-market scope labels, and reruns before trusting paper-snapshot model rankings.

• XFinBench complex financial problem solving - ACL Findings 2025, published July 2025 and retrieved 2026-06-05. Captured through the pillar 21 PDF/OCR pipeline as research/21-benchmarks/raw/xfinbench-complex-financial-problem-solving-2025.findings-acl.457.chandra.txt and promoted through sources/21-benchmarks/xfinbench-complex-financial-problem-solving-raw.md.

• AIMA The Long-Short / CFM quant episode - official AIMA episode page published 2026-03-25 and retrieved 2026-06-05. Captured as sources/06-industry-verticals/aima-long-short-cfm-quant-2026-raw.md with Apple lookup verification of the Acast RSS feed and a new scripts/podcast_sources.yaml show entry aima-long-short. Audio was archived locally with a .meta.md sidecar, raw transcript was written to research/13-multimodal-sources/aima-long-short/raw/rss-8026b27a9d70.txt, and a promoted extract was added at research/13-multimodal-sources/aima-long-short/2026-03-25-philip-seager-cfm-systematic-multistrategy.md. Promotion: transcript-backed systematic-investing practitioner vocabulary. Do not use as current AI-agent deployment, alpha, allocator-suitability, or model/runtime evidence. Promotion: complex multimodal financial problem-solving benchmark-design evidence, not current model-ranking proof. It strengthens temporal reasoning, future forecasting, scenario planning, numerical modelling, visual chart/curve interpretation, knowledge-augmentation, and human-expert-baseline gates.

• LLM stock forecasting from a hedge-fund perspective - arXiv 2605.05211, submitted 2026-04-10 and retrieved 2026-06-02. Captured through the pillar 21 PDF/OCR pipeline as research/21-benchmarks/raw/llm-stock-forecasting-hedge-fund-perspective-2605.05211.chandra.txt. Promotion: benchmark-design evidence for text-alpha and trading-agent guardrails. It strengthens the negative-evidence lane by requiring full market-cycle evaluation, non-LLM baselines, trading-aligned metrics, leakage controls, and illiquidity/capacity checks before any LLM stock forecasting claim is treated as useful.

• Balyasny / OpenAI AI research engine - OpenAI customer case study, published 2026-03-06 and retrieved 2026-06-05. Captured through sources/06-industry-verticals/balyasny-openai-ai-research-engine-raw.md and materialized as source-acquisition fixture sa-098. Promotion: production-bar finance research-agent architecture evidence, not alpha proof or audited ROI. It strengthens the vendor-product and research-agent lanes by requiring pre-deployment model evaluation, scoped tools/data access, traceable reasoning paths, compliance guardrails, workflow-specific agents, and human decision-support boundaries before vendor speedup claims are trusted.

• Man Group / Anthropic / AlphaGPT partnership - official Man Group announcement and PDF mirror, published 2026-02-11 and retrieved 2026-06-05. Captured through sources/06-industry-verticals/man-anthropic-alphagpt-partnership-raw.md and materialized as source-acquisition fixture sa-099. Promotion: systematic-manager frontier-lab partnership and workflow-integration evidence, not investment-performance evidence. It strengthens the systematic quant practitioner lane by requiring agents to distinguish Claude / Claude Skills / Claude Code / AlphaGPT proprietary workflow signals from runnable local, MLX, llama.cpp, SageMaker, Bedrock, or Neuron artifacts.

• FinSheet-Bench financial spreadsheets - arXiv 2603.07316, published 2026-03-07 and retrieved 2026-05-31. Captured through the pillar 21 PDF/OCR pipeline and promoted through sources/21-benchmarks/finsheet-bench-financial-spreadsheets-raw.md. Promotion: benchmark-design evidence for private-markets diligence, spreadsheet extraction, and deterministic-computation guardrails. It strengthens the negative-evidence lane by showing that standalone frontier models remain too error-prone for unsupervised professional spreadsheet use, especially on irregular layouts, fund dividers, multi-line headers, aggregation, sorting, and complex calculations.

• Agentic AI in Finance survey - arXiv 2604.21672, published 2026-04-23 and retrieved 2026-05-28. Captured through the pillar 06 PDF/OCR pipeline and promoted through sources/06-industry-verticals/agentic-ai-finance-survey-2026-raw.md. Promotion: benchmark-design and governance evidence for multi-agent trading, portfolio, risk, compliance, market-structure, and systemic-risk tasks. It strengthens the negative-evidence lane by requiring system-level stability, liquidity, market-impact, shock-recovery, accountability, audit-trail, and continuous-validation checks before autonomous finance-agent claims are trusted.

• U.S. Senate HSGAC hedge-fund AI/ML report - June 2024 staff report, retrieved 2026-06-02. Captured through the pillar 06 PDF/OCR pipeline as research/06-industry-verticals/raw/us-senate-hedge-funds-ai-ml-2024.chandra.txt and promoted through sources/06-industry-verticals/us-senate-hedge-funds-ai-ml-2024-raw.md. Promotion: government/regulatory evidence for hedge-fund AI/ML governance, not alpha evidence. It is now frozen as source-acquisition fixture sa-085. It strengthens the negative-evidence lane by requiring consistent system definitions, human-review boundaries, testing/review cadence, client-disclosure adequacy, version control, audit trails, overfitting/backtesting skepticism, and systemic-risk checks for herding, manipulation, and market stability.

• SQL benchmark quality and dialect portability - NL2SQLBench / arXiv 2604.16493 plus PARROT / arXiv 2509.23338, promoted on 2026-06-02 through sources/21-benchmarks/sql-benchmark-quality-dialect-portability-raw.md and source-acquisition fixture sa-084. Promotion: retrieval/NL2SQL harness design evidence, not finance-alpha evidence. It strengthens the Snowflake/Postgres/custom-schema lane by requiring module-level schema-selection, SQL-generation, query-revision, dialect-portability, execution-result, cost/latency, permission, semantic-layer, provenance, abstention, and internal-gold-query checks instead of trusting generic BIRD/Spider scores.

• BankerToolBench investment-banking workflows - arXiv 2604.11304, retrieved 2026-05-29 and promoted through sources/21-benchmarks/finance-investment-banking-deal-work-benchmarks-2026-raw.md. Promotion: benchmark-design evidence for deal-work and investment-banking agents, not buy-side alpha evidence. It is now frozen as source-acquisition fixture sa-086. It strengthens the multi-artifact workflow lane by requiring senior-banker request parsing, data-room provenance, market/SEC research, Excel/PPT/PDF/Word/CSV consistency, spreadsheet formula fidelity, confidentiality/compliance controls, and human-review handoff before any client-readiness claim.

• The Fund AI Pod funds-industry AI interviews - verified on 2026-06-05 through the official site, Apple Podcasts, and executable RSS feed https://anchor.fm/s/106e86050/podcast/rss, promoted through sources/13-multimodal-sources/the-fund-ai-pod-raw.md and added to scripts/podcast_sources.yaml. Highest-value leads are Shu Bai on hedge-fund research and thesis challenge, Pat Starling on FactSet AI Foundry / APIs versus MCPs / auditability, Alex Benke on Ridgeline system-of-record agents, Alex Dunegan on FX/currency workflows, and Simmons & Simmons on governed legal AI. Archive audio/transcripts before promoting claims; sponsor, host, and LinkedIn metadata stay discovery-only. The Shu Bai episode was archived/transcribed and promoted on 2026-06-05 as practitioner workflow evidence for hedge-fund research/thesis-challenge patterns. The Pat Starling episode was archived/transcribed and promoted on 2026-06-05 as vendor/platform architecture evidence for FactSet AI Foundry, MCP/API/feed delivery, auditability, grounded answers, model routing, and buy-side analyst research workflows. Remaining episodes stay source-discovery until transcript review.

• scaledown.ai finance workflow platform signal - queued on 2026-06-05 as source-acquisition fixture sa-087, then resolved on 2026-06-06 through sources/06-industry-verticals/scaledown-ai-finance-workflow-platform-raw.md. Promotion decision: keep as finance-workflow source-discovery/watchlist, not a pillar 06 product finding. Official ScaleDown docs reviewed on 2026-06-06 prove context engineering, task-specific SLMs, compression, extraction, AST/BM25 pruning, semantic retrieval, and span offsets, but do not name a concrete finance workflow, data/source connector, investment artifact, finance governance control, auditability boundary, access-control boundary, or named finance customer architecture. Valid promotion is instead the technical sa-123 context-engineering source card. Do not promote third- party commentary, social mentions, unsourced logos, inferred deployment state, alpha, ROI, investment-performance, or autonomous capital-allocation claims.

• CurrentAI proactive investment-agent product signal - official page retrieved 2026-06-06 and promoted through sources/06-industry-verticals/currentai-proactive-investment-agents-raw.md; materialized as source-acquisition fixture sa-122. Promotion: finance workflow-vendor evidence for investment-team background agents, financial model updates, scenario analysis, value-chain read-throughs, earnings previews, and management-meeting model-impact analysis. Evidence boundary: product-surface signal only. Do not use as alpha, autonomous-trading, customer-adoption, audited-productivity, model-superiority, or backend- portability evidence.

• ScaleDown context-engineering / task-specific SLM signal - official docs retrieved 2026-06-06 and promoted through sources/21-benchmarks/scaledown-context-engineering-slm-raw.md; materialized as source-acquisition fixture sa-123. Promotion: SLM/context infrastructure evidence for task-specific context extraction, prompt compression, AST/BM25 code-context pruning, local-embedding/FAISS semantic retrieval, preservation controls for domain-specific terms, and offset/confidence-bearing extraction. Evidence boundary: not finance outcome evidence and not MLX/GGUF/CUDA/ROCm/SageMaker/Bedrock/Neuron compatibility evidence unless separate source-card proof exists.

• Hudson River Trading AI, benchmark, and data guardrail source - recorded on 2026-06-05 through sources/06-industry-verticals/hrt-ai-benchmark-data-source-raw.md and materialized as source-acquisition fixture sa-088. Promotion: first-party negative-evidence and benchmark-design source for trading-agent, quant-code, alternative-data, and benchmark-mismatch checks. HRT’s official articles preserve three durable guardrails: trading AI should be decomposed into prediction, optimization, execution, risk, and regulatory / market- participant constraints; ML papers should be judged for simplicity, reproducibility, and generality rather than leaderboard gains alone; and data sources should be tested for relevance, uniqueness, leakage/lookahead, sample size, noise, and provenance. The Bloomberg Odd Lots / Omny episode “How Hudson River Trading Actually Uses AI” was captured through the podcast pipeline on 2026-06-05 as research/13-multimodal-sources/odd-lots/raw/rss-0562c5190e8c.txt with an excluded local MP3 and committed audio sidecar. It is promoted only as named practitioner workflow and guardrail evidence for short-horizon market-data ML, infrastructure, operational risk, regulatory controls, and LLM backtest contamination checks.

• Acadian 2026 Investor Forum systematic-manager scale source - recorded on 2026-06-05 through sources/06-industry-verticals/acadian-2026-investor-forum-systematic-scale-raw.md and materialized as source-acquisition fixture sa-089. Promotion: current official public systematic-manager context, not AI-agent or alpha proof. The page records $196B of client assets as of 2026-03-31, an advanced alpha engine analyzing 65,000+ securities daily with hundreds of unique signals, more than 100 investment experts, and a five-year benchmark-outperformance claim that is explicitly limited to eligible composite AUM. Its most useful benchmark lesson is the authority boundary: Acadian separates discretionary AUM from model advisory assets where it does not have trading authority. Agents should preserve those dates, caveats, and authority boundaries rather than flattening the page into a generic “AI alpha” claim.

• Trading-R1 / Alpha-R1 / Trade-R1 trading-reasoning RL cluster - captured on 2026-06-05 through pillar 21 PDF/OCR ingestion and promoted through sources/21-benchmarks/trading-alpha-r1-reasoning-rl-raw.md; materialized as source-acquisition fixture sa-090. Promotion: systematic quant benchmark-design evidence for structured trading theses, context-aware alpha screening, stochastic market rewards, RAG-grounded process verification, and reward- hacking checks. The local PDFs remain ignored; the committed artifacts are Chandra OCR text, PDF metadata sidecars, the source card, and the frozen fixture. Do not promote reported return/drawdown results as live alpha or autonomous trading proof without code, data, transaction-cost, capacity, baseline, and cross-regime reproduction.

Current Interpretation

The next missing work is not more indiscriminate scraping. The corpus needs fewer but stronger captures:

• source cards for official firm/social/careers pages that currently exist only as brittle web signals;
• audio archives/transcripts for the highest-value quant practitioner episodes;
• OCR for newly discovered PDFs/reports;
• frozen source-discovery fixtures to test whether models can find stable papers, reports, podcasts, and official pages;
• scored StateBench runs proving which models and backends can actually do the ingestion, wiki-maintenance, and overclaim-repair work.

This keeps web discovery as an eval and prevents time-sensitive browsing from contaminating the deterministic finance replay benchmark.