# AwarenessFund — ACG Capability Inventory (Back-Testing Audit)

**Author**: research-lead (A-C-Gee)
**Date**: 2026-05-18
**Audience**: Russell Korus, Apex/Jordana (Pyonair), Corey, capital-markets-lead, sister AiCIV insiders
**Frame**: *Honest audit of what ACG can deliver right now vs what we have to source. No hand-waving. Gaps named explicitly.*

---

## Executive Summary

ACG has **STRONG general orchestration + LLM infrastructure** but **WEAK native financial-back-testing tooling**. We have a yfinance-class CLI tool (Hermes-built), generic ingestion plumbing (ago/ingest), and excellent NLP/LLM extraction discipline. We have **NO** native back-testing engine, **NO** factor library, **NO** transaction-cost model. The federation's financial-civ (Works) is currently DOWN, so domain-depth fallback unavailable. Net: **ACG should own ingestion + competing-hypothesis discipline + NLP extraction; Pyonair should own back-test engine + PPO; sister-civ Works should own factor/finance-domain when revived.**

---

## Capability Audit Matrix

| Capability | State | Location | Quality | Confidence |
|------------|-------|----------|---------|-----------|
| **Price/OHLCV scraping** | ✅ EXISTS | `projects/hermes-student-001/.../finance/stocks/scripts/stocks_client.py` (stdlib-only Yahoo + Alpha Vantage CLI) | Working, basic | HIGH |
| **Generic ingestion pipeline** | ✅ EXISTS | `tools/ago/ingest/` (correlator, normalizer, pipeline) | Generic; needs financial adapters | MEDIUM |
| **Web scraping primitives** | ✅ EXISTS | `autonomy/skills/jina-reader/`, `autonomy/skills/deep-search/`, `autonomy/skills/article-extract/` | Excellent for filings/transcripts | HIGH |
| **LLM extraction discipline** | ✅ EXISTS | `autonomy/skills/anti-fabrication-pre-flight/`, `autonomy/skills/transcription-not-paraphrase/`, scientific-method + critical-thinking | THE strongest ACG advantage | HIGH |
| **SEC EDGAR / 10-K parsing** | ❌ NOT BUILT | n/a | n/a | LOW (would need to author) |
| **Earnings transcript pipeline** | ❌ NOT BUILT | n/a | n/a | LOW |
| **FRED / macro data ingest** | ⚠️ PARTIAL | Generic web-fetch works; no dedicated wrapper | Would build in <1 day | MEDIUM |
| **Time-series database** | ❌ NOT BUILT | Sqlite/JSONL ad-hoc only | n/a | LOW |
| **Back-testing engine (vectorized)** | ❌ NOT BUILT | n/a | n/a | LOW — recommend Pyonair owns this |
| **Factor library (Fama-French, momentum, etc.)** | ❌ NOT BUILT | n/a | n/a | LOW |
| **Transaction-cost / slippage model** | ❌ NOT BUILT | n/a | n/a | LOW |
| **Portfolio construction / optimization** | ❌ NOT BUILT | n/a | n/a | LOW |
| **Risk model / VaR / max DD reporting** | ❌ NOT BUILT | n/a | n/a | LOW |
| **PPO / RL infrastructure** | ❌ NOT BUILT | n/a | n/a | LOW — Pyonair sending code |
| **GPU compute** | ⚠️ PARTIAL | Local CPU + cloud LLM access; no dedicated GPU rig for RL training | Medium | MEDIUM |
| **Distributed compute / fleet** | ✅ EXISTS | Witness owns Docker fleet; can spin training containers | HIGH | HIGH (Witness coordinated) |
| **Multi-agent orchestration** | ✅ STRONG | Conductor-of-conductors pattern; sister-civ delegation; fanout | HIGH | HIGH |
| **Cross-civ data exchange** | ✅ EXISTS | TGIM substrate + Hub + AgentMail | HIGH | HIGH |
| **Visualization / charting** | ⚠️ PARTIAL | `autonomy/skills/diagram-generator/` (mermaid), no Plotly/matplotlib-class | Medium | MEDIUM |
| **Investor-grade reporting (HTML/PDF)** | ⚠️ PARTIAL | Blog publishing pipeline (Netlify); PDF generation via skills | Medium | MEDIUM |
| **Static-data sources (Wikipedia, GitHub)** | ✅ EXISTS | jina-reader + WebFetch | HIGH | HIGH |
| **Cron / scheduled jobs** | ✅ EXISTS | `autonomy/skills/scheduled-tasks/`, daemon-watchdog pattern | HIGH | HIGH |

---

## Existing Financial-Adjacent Code

```
projects/hermes-student-001/provisioning/hermes-agent/optional-skills/finance/stocks/scripts/stocks_client.py
  → Stdlib-only Yahoo Finance + Alpha Vantage CLI
  → Handles cookie jar, retries, backoff
  → Quality: production-ready for V1 baseline ingestion
  → NOTE: lives in hermes provisioning — would need to copy/adapt for ACG-native use

tools/ago/ingest/
  ├── correlator.py     → Generic event correlation
  ├── normalizer.py     → Generic data normalization
  └── pipeline.py       → Generic pipeline orchestration
  → Quality: scaffolding only, no financial domain adapters
  → Would need: financial-data Schema adapters (OHLCV → factor → portfolio)

memories/agents/token-scout/
  → Crypto-token intelligence (NOT equity)
  → Not directly reusable for AwarenessFund
  → Architectural pattern (scan → analyze → publish → monitor) IS portable

memories/agents/research/20260501-works-financial-civ-architecture.md
  → Works (Kimi K2.6) financial-civ DESIGN (10 verticals × 5 specialists)
  → Status: DESIGNED but Works currently DOWN
  → Source-attribution discipline IS portable to AwarenessFund work
```

---

## Gap Analysis (Honest)

### CRITICAL gaps (block the back-test entirely)
1. **No back-testing engine.** Need vectorbt-class library or Pyonair brings one. ACG cannot build this in <2 weeks alongside other work.
2. **No transaction-cost model.** Even a simple linear-impact model is missing. Critical for honest Sharpe/info-ratio numbers.
3. **No survivorship-bias-corrected ticker universe.** All current scraping is "what's listed today" — historical delistings missing. Would need CRSP-like dataset ($$$).

### MAJOR gaps (block specific variations)
4. **No SEC EDGAR XBRL parser.** Variation V2 in ingestion spec depends on this. ~1-2 engineer-weeks to build OR use `sec-api.io` ($50-200/mo).
5. **No earnings-transcript pipeline.** Variation V3. Requires scrape rotation + LLM extraction — moderate effort.
6. **No factor library.** No momentum, value, quality, size factor implementations. Either build (~1 wk) or import (`pandas-quant` etc).

### MINOR gaps (workable but not ideal)
7. **No dedicated time-series DB.** SQLite/Parquet works for 24mo back-test scale, but Phase 2 will need DuckDB or TimescaleDB.
8. **No GPU rig for PPO training.** Witness fleet can rent; CPU OK for inference; training will be slow without H100/A100 access.
9. **No financial-charting library.** Mermaid is text-only. Need matplotlib/Plotly for investor decks.

### What we DO have that few teams have
- **Anti-fabrication discipline** (critical for LLM-extracted features — V5/V2 in ingestion spec)
- **Cross-civ research fanout** (4 sister AiCIVs can survey PPO literature in parallel)
- **Source-attribution rigor** (every claim traceable per scientific-method skill)
- **Conductor-of-conductors orchestration** (the fund's research pipeline IS multi-agent work)
- **Reproducible research output** (blog publishing + Netlify + downloads = investor-grade artifacts)

---

## Recommended Division of Labor

| Function | Owner | Rationale |
|----------|-------|-----------|
| **Ingestion design + execution** | ACG (research-lead) | jina-reader + LLM extraction + anti-fabrication |
| **Competing-hypothesis framework** | ACG (research-lead) | scientific-method + critical-thinking skill discipline |
| **Ticker universe + back-test infra** | capital-markets-lead | Already spawned in parallel; owns the engine |
| **Back-test engine (vectorized)** | Pyonair (Apex/Jordana) | They have it; we don't |
| **PPO / RL model** | Pyonair | They're shipping the code |
| **Factor library** | Pyonair OR Works (when revived) | Either is faster than ACG building it |
| **Macro / regime overlay** | ACG (research-lead, V8 ingestion) | FRED is free and scriptable |
| **Investor-grade reporting** | ACG (business-lead pipeline) | Blog + landing-page pipeline works |
| **Distributed training compute** | Witness | Fleet ops are her domain |
| **Source-attribution audit** | ACG (research-lead) | Verifier-as-substrate discipline |

---

## Build vs Buy Recommendations

| Component | Build | Buy | Recommendation |
|-----------|-------|-----|----------------|
| OHLCV + macro | 1 day (already have stocks_client.py) | $0 (Yahoo, FRED) | **BUILD** — trivial |
| SEC filings | 1-2 weeks | $50-200/mo (sec-api.io) | **BUY** — engineering time better spent elsewhere |
| Earnings transcripts | 2-3 weeks (scrape) | $500-5K/mo (AlphaSense) | **BUILD** scrape for back-test only; reassess if live trading |
| Back-test engine | 4-8 weeks | Pyonair brings it | **PARTNER** — Pyonair |
| PPO model | 4-12 weeks | Pyonair brings it | **PARTNER** — Pyonair |
| Survivorship-corrected universe | impossible from scratch | $5-50K/yr (CRSP, Compustat) | **BUY** if fund actually launches |
| Visualization | 1 week (Plotly integration) | $0 (open source) | **BUILD** when needed |
| Investor portal | 1-2 weeks (Netlify scaffold exists) | $0-200/mo (host) | **BUILD** — extends existing capability |

---

## Honest Limitations of ACG for This Mission

1. **We have not run a back-test before.** Strong general LLM + orchestration; thin specific finance experience. Mitigation: lean on Pyonair + Works.
2. **No live brokerage / execution layer.** This is fund-back-end, not in ACG's scope. (Russell/Pyonair territory.)
3. **No risk management / compliance / regulatory framework.** When fund launches, this is a hard external dependency.
4. **Time-zone alignment** with Pyonair team unknown — assume async coordination.
5. **Works civ is DOWN** — the financial-domain depth we'd lean on for factor-construction critique is unavailable. Pending restart.
6. **Sister AiCIV fanout** (Proof/Hengshi/Hermes) is dispatched but results uncertain (~hours to days, async).
7. **Pyonair PPO code has not arrived.** All PPO research is preparatory; will need to integrate when code lands.

---

## Phase-1 Practical Deliverables (achievable in 1-2 weeks)

| Deliverable | Owner | Status |
|-------------|-------|--------|
| Ingestion variations spec | research-lead | ✅ SHIPPED today |
| Competing hypotheses spec | research-lead | ✅ SHIPPED today |
| Capability inventory (this doc) | research-lead | ✅ SHIPPED today |
| PPO survey from 3-4 sister AiCIVs | sister civs (async) | DISPATCHED 17:35Z; pending replies |
| Ticker universe | capital-markets-lead | parallel-spawned |
| Back-test infra design | capital-markets-lead | parallel-spawned |
| S&P 500 comparative methodology | capital-markets-lead | parallel-spawned |
| V1 (Yahoo+FRED) baseline ingest | ACG eng (when prioritized) | NOT STARTED |
| V2 (SEC filings) prototype | ACG eng or buy sec-api.io | NOT STARTED |

---

## Cross-References

- `projects/awareness-fund/research/ingestion-variations-spec.md` — what data we need
- `projects/awareness-fund/research/competing-hypotheses.md` — what we're testing
- `projects/awareness-fund/capital-markets/` — capital-markets-lead's parallel work
- `projects/awareness-fund/research/students/` — sister-civ PPO research (pending)
- `memories/agents/research/20260501-works-financial-civ-architecture.md` — Works financial-civ design (DOWN, but design portable)

---

## Status

SPEC v1.0 — ready for capital-markets-lead cross-read + Pyonair integration when PPO code lands.

**Honest summary for Russell/Apex/Jordana**:
> "ACG owns the ingestion-and-discipline layer. Pyonair owns the back-test-and-PPO layer. Together with capital-markets-lead's ticker universe + S&P comparative work, that's a complete Phase-1 stack. We have not run a back-test before, so we'd defer to Pyonair on engine choice and validation methodology. Our genuine advantage is anti-fabrication discipline + multi-agent research fanout."