What Team 1 Actually Built
This is not a paraphrase or a memory of the earlier "extracted and enhanced" framing — every fact on this page was pulled directly from Team 1's public repository: their README, decision log, weekly QA logs, and LLM comparison outputs.
Team 1 ran a 10-week sprint cycle: five specialist agents feed a shared pipeline context, four AI models are queried in parallel with an identical prompt, a human QA role scores the team's own final call against actual outcomes the following week, and the whole thing runs partly through GitHub Actions.
README.md, DECISION.md,
data/qa/calibration_scoring_rule_W22.md, data/qa/learning_log_W24.md,
data/llm/llm_comparison_W24.md, sprints/retrospective_w25.md,
and the full repo file tree.
Scope & Data
Team 1 tracks far more than the 3 tickers MPE currently covers.
| Asset | Ticker | Why they track it |
|---|---|---|
| S&P 500 | SPX | The benchmark index |
| Nasdaq 100 | NDX | Tech-heavy, fastest-moving |
| Russell 2000 | IWM | Small caps, most rate-sensitive |
| Gold | GC | Safe haven / inflation gauge |
| Crude Oil | CL | Inflation driver, geopolitical signal |
| 10Y Treasury Yield | ZN | The gravity on all asset valuations |
| US Bonds | TLT | Stocks + bonds both down = fear signal |
| VIX | VIX | Volatility / fear index |
| +1 more asset | — | 9 total per their README badge |
The LLM Layer — Two Generations
Team 1's model panel changed over the 10-week sprint. MPE's model choice matches their second panel, not their first.
| Phase | Models used | Notes |
|---|---|---|
| Original (early weeks) | Claude, ChatGPT, Gemini, DeepSeek | Paid/consumer-tier commercial models, queried directly through each provider's own interface |
| Later (W24–W25) | NVIDIA Nemotron 3 Super, OpenAI gpt-oss-120b, Google Gemma 4 31B, Poolside Laguna M.1 | Free-tier OpenRouter models — the exact same four MPE uses |
data/llm/llm_comparison_W24.md: Team 1 got real, usable output
from all four of these models in their own W24 run. That's independent evidence the model
IDs are valid and the approach works — MPE's current 429/empty-response failures are a
pacing and reliability problem on our side, not a sign the models don't exist or don't work.
Human QA Process — What "Manual Calibration" Actually Means
There's a real scoring rubric behind Team 1's manual process, run by a rotating "R10 QA and Learning Log Lead" role.
| Confidence | Direction correct? | Score |
|---|---|---|
| High | Correct | +3 |
| Medium | Correct | +2 |
| Low / Uncertain | Correct | +1 |
| High | Wrong | −2 |
| Medium | Wrong | 0 |
| Low / Uncertain | Wrong | +1 |
A real example from their learning_log_W24.md: the team predicted
Neutral-Bullish across SPX/NDX/IWM at Medium confidence. All three directions landed
correct; two of three ranges hit. Final score: +7 / +9 — up from
0 / +9 the week before.
Architecture Decisions (from their DECISION.md)
- Agents first: built before anything downstream, since errors there compound forward.
- Fetch + process combined per agent (no separate fetch abstraction layer) — deliberately kept simple.
- Output-writing separated from agent logic — a dedicated module handles disk writes, so storage format can change without touching agent code. MPE doesn't have this separation yet
- Shared pipeline context object collects all agent outputs before the LLM stage sees any of them, decoupling the LLM stage from agent count/order. MPE passes outputs directly, no context object
- Automated tests at agent/context/prompt-building integration points, plus type checking on every change. MPE has zero automated tests currently
Team 1 vs. MPE — Stated Plainly
| Dimension | Team 1 | MPE (current) |
|---|---|---|
| Assets tracked | 9 (equities + macro hedges) | 3 (equities only) |
| Evidence agent | built — sources, models, reporting | not started |
| Automated tests | 4 test files, CI-enforced | none |
| CI/CD | 3 workflows: pipeline, agent checks, python checks | 1 pipeline workflow, no code-quality checks |
| Output/storage separation | dedicated module | inline in run_pipeline.py |
| Calibration method | Manual weekly human scoring rubric | Automated per-model EMA trust weighting |
| Weekly effort required | A rotating QA role manually scores and logs each week | Zero — runs unattended once automation is stable |
| Chart/visual output | technical_agent generates charts (PNG) | none |