Research Loop Audit

Original goal

# North Star: Investment-Grade GW Forecast vs Actual

## Mission

Build **production-grade research** we can trust for **investment decisions** on whether **more gigawatts (GW) of data-center / AI infrastructure capacity are coming online** — past, present, and future.

For each of **17 target companies**, answer:

1. **What did they say?** — GW forecasts, pipeline, and guidance each quarter  
2. **How did forecasts change?** — vintage panel: Oracle said X GW in Q1, revised to Y in Q3  
3. **What actually happened?** — delivered / operating / under-construction GW vs the original forecast  
4. **Can we see it on a chart?** — one **line graph per company** + one **portfolio line graph**

Example: *"Oracle said they'll have X data centers / Y GW by 2027. Each quarter we capture that statement. We track how the forecast moved quarter-to-quarter and compare the first forecast to what ultimately landed."*

## Target Companies (17)

ORCL, MSFT, AMZN, GOOGL, META, CoreWeave (CRWV), DLR, EQIX, IRM, APLD, CORZ, IREN, WULF, CIFR, HUT, NBIS/Nebius, NVDA

*(Vertiv/Eaton dropped from v1 — infrastructure suppliers, not capacity owners.)*

## Core Metrics (GW-first)

- operating / live GW (actual)
- GW under construction (actual snapshot)
- GW pipeline / future capacity (forecast)
- GW added this quarter (actual)
- forecast vintage: `as_of_quarter` → `target_period` → `forecast_gw`
- capex actuals & guidance (supporting evidence)
- revenue conversion per MW/GW where disclosed

## Canonical Output Layout

All production artifacts live under `research_loop/production/`:

```
production/
  {TICKER}/
    source_manifest.csv
    forecast_vintage.csv      # how forecasts changed each quarter
    capacity_timeseries.csv   # actual vs latest vs original forecast by period
    company_gw_chart.html     # line chart for this company
    extractor.py
    validator.py
  portfolio_gw_timeseries.csv
  portfolio_gw_chart.html
  goal_progress.json          # machine-readable completion status
```

## Done When (`north_star.json`)

- Every company has ≥8 quarters in `forecast_vintage.csv` and `capacity_timeseries.csv`
- Every company has provenance-complete rows (https source URLs, snippets, validation PASS)
- Every company has `company_gw_chart.html`
- Portfolio chart aggregates all companies
- `goal_completion.py` reports ≥85% investment-confidence score

## Source Preference

Primary: SEC 10-Q/10-K/8-K, earnings supplements, IR pages, investor decks, transcripts  
Secondary: industry press (discovery only — not row-level evidence)

## Supervisor Mandate

The auto-continue supervisor runs **until the north star is achieved**, not for a fixed loop count. Each child loop picks **one company or one repair task** that moves the portfolio closest to done.

Completed tasks

Repair the failed prior child run `gw-forecast-production-001-loop04`. The previous task was: Complete AMZN production pipeline: replace cross-ticker fallbacks (MSFT manifest/vintage/chart + EQIX timeseries) with real Amazon/AWS primary-source artifacts in production/AMZN/, following the proven ORCL/MSFT hyperscaler capex-proxy pattern (SEC 10-Q/10-K cash-flow capex actuals + earnings-call guidance vintages; normalized_gw blank with explicit caveat). Fix the persisted artifacts from that prior run so the machine gate passes. Do not choose a new task. Use the copied context for `gw-forecast-production-001-loop04` and update the same expected artifact names. Prior child public URL: http://204.168.210.83:8130/research-loop/gw-forecast-production-001-loop04/index.html Machine-gate blockers to fix: - reviewer codex exec failed (exit 1): codex exited 1 but wrote no output to /home/dev/research_loop/runs/gw-forecast-production-001-loop04/agent_io/iter001_reviewer_c1/output.md OpenAI Codex v0.137.0 -------- workdir: /home/dev model: gpt-5.5 provider: openai approval: never sandbox: danger-full-access reasoning effort: medium reasoning summaries: none session id: 019ea2fb-5593-7433-b45d-bfdc8d0a5bad -------- user You are the Reviewer agent. You verify work quality but do NOT fix issues. ## Your job Review the **persisted artifact file contents** against the task acceptance criteria. Mechanical checks (file exists, tickers present, PRIMARY/SECONDARY labels) are already verified by the harness — focus on content quality. ## Output format Start with exactly one of: VERDICT: PASS or VERDICT: CHANGES_REQUIRED If CHANGES_RE Prior artifacts to inspect: - build_timeseries.py - capacity_timeseries.csv - company_gw_chart.html - extractor.py - forecast_vintage.csv - source_manifest.csv - validator.py — verdict: PASS
Explicit seed task provided on the command line.

Task history

Repair the failed prior child run `gw-forecast-production-001-loop04`. The previous task was: Complete AMZN production pipeline: replace cross-ticker fallbacks (MSFT manifest/vintage/chart + EQIX timeseries) with real Amazon/AWS primary-source artifacts in production/AMZN/, following the proven ORCL/MSFT hyperscaler capex-proxy pattern (SEC 10-Q/10-K cash-flow capex actuals + earnings-call guidance vintages; normalized_gw blank with explicit caveat). Fix the persisted artifacts from that prior run so the machine gate passes. Do not choose a new task. Use the copied context for `gw-forecast-production-001-loop04` and update the same expected artifact names. Prior child public URL: http://204.168.210.83:8130/research-loop/gw-forecast-production-001-loop04/index.html Machine-gate blockers to fix: - reviewer codex exec failed (exit 1): codex exited 1 but wrote no output to /home/dev/research_loop/runs/gw-forecast-production-001-loop04/agent_io/iter001_reviewer_c1/output.md OpenAI Codex v0.137.0 -------- workdir: /home/dev model: gpt-5.5 provider: openai approval: never sandbox: danger-full-access reasoning effort: medium reasoning summaries: none session id: 019ea2fb-5593-7433-b45d-bfdc8d0a5bad -------- user You are the Reviewer agent. You verify work quality but do NOT fix issues. ## Your job Review the persisted artifact file contents against the task acceptance criteria. Mechanical checks (file exists, tickers present, PRIMARY/SECONDARY labels) are already verified by the harness — focus on content quality. ## Output format Start with exactly one of: VERDICT: PASS or VERDICT: CHANGES_REQUIRED If CHANGES_RE Prior artifacts to inspect: - build_timeseries.py - capacity_timeseries.csv - company_gw_chart.html - extractor.py - forecast_vintage.csv - source_manifest.csv - validator.py

Explicit seed task provided on the command line.

Status: passed

Review cycles: 1

Acceptance criteria

Addresses every listed machine-gate blocker or explains an exact unavailable item in a machine-readable row.
Returns complete corrected versions of the expected artifact files, not a prose-only explanation.
Does not start a new research direction until the failed prior task passes.
Preserves source provenance, snippets, confidence, caveats, and validation summaries.
production/AMZN/source_manifest.csv exists with ≥8 Amazon fiscal quarters, ticker=AMZN, document_type sec_10q/sec_10k/earnings materials, and https primary-source URLs (SEC EDGAR + IR)
production/AMZN/extractor.py reads source_manifest.csv, fetches sources, and writes production/AMZN/forecast_vintage.csv with canonical schema rows for capex_actual and capex_guidance vintages
production/AMZN/forecast_vintage.csv has ≥16 rows spanning ≥8 distinct as_of_period values; all rows have ticker=AMZN and no MSFT/ORCL/EQIX contamination
production/AMZN/build_timeseries.py (or equivalent) derives production/AMZN/capacity_timeseries.csv with ≥8 fiscal_quarter rows including actual_gw, latest_forecast_gw, original_forecast_gw, and capex proxy columns
production/AMZN/company_gw_chart.html renders actual vs latest-forecast vs original-forecast lines from capacity_timeseries.csv
production/AMZN/validator.py enforces schema, ≥8 quarters, https provenance, snippet checks, and exits 0
python3 production/AMZN/validator.py reports status PASS
goal_progress.json shows AMZN complete=true with empty fallback_used list

Expected artifacts

source_manifest.csv
extractor.py
forecast_vintage.csv
build_timeseries.py
capacity_timeseries.csv
company_gw_chart.html
validator.py

Agent calls & decisions

Time	Iter	Role	Task	Verdict	Decisions	Blockers	Artifacts
2026-06-07T16:48:07.866958+00:00	0	system	start		[ "run started" ]	[]
2026-06-07T16:48:08.374708+00:00	1	planner	Repair the failed prior child run `gw-forecast-production-001-loop04`. The previous task was: Complete AMZN production pipeline: replace cross-ticker fallbacks (MSFT manifest/vintage/chart + EQIX timeseries) with real Amazon/AWS primary-source artifacts in production/AMZN/, following the proven ORCL/MSFT hyperscaler capex-proxy pattern (SEC 10-Q/10-K cash-flow capex actuals + earnings-call guidance vintages; normalized_gw blank with explicit caveat). Fix the persisted artifacts from that prior run so the machine gate passes. Do not choose a new task. Use the copied context for `gw-forecast-production-001-loop04` and update the same expected artifact names. Prior child public URL: http://204.168.210.83:8130/research-loop/gw-forecast-production-001-loop04/index.html Machine-gate blockers to fix: - reviewer codex exec failed (exit 1): codex exited 1 but wrote no output to /home/dev/research_loop/runs/gw-forecast-production-001-loop04/agent_io/iter001_reviewer_c1/output.md OpenAI Codex v0.137.0 -------- workdir: /home/dev model: gpt-5.5 provider: openai approval: never sandbox: danger-full-access reasoning effort: medium reasoning summaries: none session id: 019ea2fb-5593-7433-b45d-bfdc8d0a5bad -------- user You are the Reviewer agent. You verify work quality but do NOT fix issues. ## Your job Review the persisted artifact file contents against the task acceptance criteria. Mechanical checks (file exists, tickers present, PRIMARY/SECONDARY labels) are already verified by the harness — focus on content quality. ## Output format Start with exactly one of: VERDICT: PASS or VERDICT: CHANGES_REQUIRED If CHANGES_RE Prior artifacts to inspect: - build_timeseries.py - capacity_timeseries.csv - company_gw_chart.html - extractor.py - forecast_vintage.csv - source_manifest.csv - validator.py		[ "Explicit seed task provided on the command line." ]	[]
2026-06-07T16:51:10.403505+00:00	1	worker	Repair the failed prior child run `gw-forecast-production-001-loop04`. The previous task was: Complete AMZN production pipeline: replace cross-ticker fallbacks (MSFT manifest/vintage/chart + EQIX timeseries) with real Amazon/AWS primary-source artifacts in production/AMZN/, following the proven ORCL/MSFT hyperscaler capex-proxy pattern (SEC 10-Q/10-K cash-flow capex actuals + earnings-call guidance vintages; normalized_gw blank with explicit caveat). Fix the persisted artifacts from that prior run so the machine gate passes. Do not choose a new task. Use the copied context for `gw-forecast-production-001-loop04` and update the same expected artifact names. Prior child public URL: http://204.168.210.83:8130/research-loop/gw-forecast-production-001-loop04/index.html Machine-gate blockers to fix: - reviewer codex exec failed (exit 1): codex exited 1 but wrote no output to /home/dev/research_loop/runs/gw-forecast-production-001-loop04/agent_io/iter001_reviewer_c1/output.md OpenAI Codex v0.137.0 -------- workdir: /home/dev model: gpt-5.5 provider: openai approval: never sandbox: danger-full-access reasoning effort: medium reasoning summaries: none session id: 019ea2fb-5593-7433-b45d-bfdc8d0a5bad -------- user You are the Reviewer agent. You verify work quality but do NOT fix issues. ## Your job Review the persisted artifact file contents against the task acceptance criteria. Mechanical checks (file exists, tickers present, PRIMARY/SECONDARY labels) are already verified by the harness — focus on content quality. ## Output format Start with exactly one of: VERDICT: PASS or VERDICT: CHANGES_REQUIRED If CHANGES_RE Prior artifacts to inspect: - build_timeseries.py - capacity_timeseries.csv - company_gw_chart.html - extractor.py - forecast_vintage.csv - source_manifest.csv - validator.py		[ "## Decisions", "- Confirmed loop04 content repair is complete. `production/AMZN/` and `loop04/artifacts/production/AMZN/` both pass `python3 validator.py` (exit 0, status PASS). Cross-ticker MSFT/ORCL/EQIX fallbacks are gone; all rows are `ticker=AMZN` with SEC EDGAR + Amazon IR primary URLs.", "- Loop04 machine-gate blocker was infrastructure, not data. Reviewer Codex failed with `OSError: [Errno 7] Argument list too long` \u2014 no content verdict was produced. Re-emitting the verified canonical artifacts for loop05 harness extraction.", "- Hyperscaler capex-proxy pattern (ORCL/MSFT). Amazon does not disclose MW/GW; `normalized_gw` and GW chart columns stay blank. Capex actuals from `us-gaap:PaymentsToAcquireProductiveAssets` in 10-Q/10-K; guidance vintages from MD&A + SEC 8-K ex99 ($200B FY2026 from FY2025-FQ4). FY2025 guidance ($105B\u2192$118B\u2192$125B) is derived with explicit `is_derived=1` caveats.", "- FY2026 guidance rollup anchor: `GUIDANCE_ORIGIN_QUARTER=FY2025-FQ4`, `GUIDANCE_TARGET=FY2026`, original/latest both $200B from first FY2026 guide vintage.", "## Evidence", "PRIMARY", "- SEC EDGAR CIK 1018724: FY2024\u2013FY2026 10-Q/10-K cash-flow capex (e.g. `https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm`)" ]	[]	build_timeseries.py, capacity_timeseries.csv, company_gw_chart.html, extractor.py, forecast_vintage.csv, source_manifest.csv, validator.py
2026-06-07T16:53:23.776088+00:00	1	reviewer	Repair the failed prior child run `gw-forecast-production-001-loop04`. The previous task was: Complete AMZN production pipeline: replace cross-ticker fallbacks (MSFT manifest/vintage/chart + EQIX timeseries) with real Amazon/AWS primary-source artifacts in production/AMZN/, following the proven ORCL/MSFT hyperscaler capex-proxy pattern (SEC 10-Q/10-K cash-flow capex actuals + earnings-call guidance vintages; normalized_gw blank with explicit caveat). Fix the persisted artifacts from that prior run so the machine gate passes. Do not choose a new task. Use the copied context for `gw-forecast-production-001-loop04` and update the same expected artifact names. Prior child public URL: http://204.168.210.83:8130/research-loop/gw-forecast-production-001-loop04/index.html Machine-gate blockers to fix: - reviewer codex exec failed (exit 1): codex exited 1 but wrote no output to /home/dev/research_loop/runs/gw-forecast-production-001-loop04/agent_io/iter001_reviewer_c1/output.md OpenAI Codex v0.137.0 -------- workdir: /home/dev model: gpt-5.5 provider: openai approval: never sandbox: danger-full-access reasoning effort: medium reasoning summaries: none session id: 019ea2fb-5593-7433-b45d-bfdc8d0a5bad -------- user You are the Reviewer agent. You verify work quality but do NOT fix issues. ## Your job Review the persisted artifact file contents against the task acceptance criteria. Mechanical checks (file exists, tickers present, PRIMARY/SECONDARY labels) are already verified by the harness — focus on content quality. ## Output format Start with exactly one of: VERDICT: PASS or VERDICT: CHANGES_REQUIRED If CHANGES_RE Prior artifacts to inspect: - build_timeseries.py - capacity_timeseries.csv - company_gw_chart.html - extractor.py - forecast_vintage.csv - source_manifest.csv - validator.py	PASS	[ "review cycle 1" ]	[]	build_timeseries.py, capacity_timeseries.csv, company_gw_chart.html, extractor.py, forecast_vintage.csv, source_manifest.csv, validator.py
2026-06-07T16:53:23.880078+00:00	1	machine_gate	Repair the failed prior child run `gw-forecast-production-001-loop04`. The previous task was: Complete AMZN production pipeline: replace cross-ticker fallbacks (MSFT manifest/vintage/chart + EQIX timeseries) with real Amazon/AWS primary-source artifacts in production/AMZN/, following the proven ORCL/MSFT hyperscaler capex-proxy pattern (SEC 10-Q/10-K cash-flow capex actuals + earnings-call guidance vintages; normalized_gw blank with explicit caveat). Fix the persisted artifacts from that prior run so the machine gate passes. Do not choose a new task. Use the copied context for `gw-forecast-production-001-loop04` and update the same expected artifact names. Prior child public URL: http://204.168.210.83:8130/research-loop/gw-forecast-production-001-loop04/index.html Machine-gate blockers to fix: - reviewer codex exec failed (exit 1): codex exited 1 but wrote no output to /home/dev/research_loop/runs/gw-forecast-production-001-loop04/agent_io/iter001_reviewer_c1/output.md OpenAI Codex v0.137.0 -------- workdir: /home/dev model: gpt-5.5 provider: openai approval: never sandbox: danger-full-access reasoning effort: medium reasoning summaries: none session id: 019ea2fb-5593-7433-b45d-bfdc8d0a5bad -------- user You are the Reviewer agent. You verify work quality but do NOT fix issues. ## Your job Review the persisted artifact file contents against the task acceptance criteria. Mechanical checks (file exists, tickers present, PRIMARY/SECONDARY labels) are already verified by the harness — focus on content quality. ## Output format Start with exactly one of: VERDICT: PASS or VERDICT: CHANGES_REQUIRED If CHANGES_RE Prior artifacts to inspect: - build_timeseries.py - capacity_timeseries.csv - company_gw_chart.html - extractor.py - forecast_vintage.csv - source_manifest.csv - validator.py	PASS	[ "expected artifact validators and production gates passed" ]	[]	build_timeseries.py, capacity_timeseries.csv, company_gw_chart.html, extractor.py, forecast_vintage.csv, source_manifest.csv, validator.py

Time

Iter

Role

Task

Verdict

Decisions

Blockers

Artifacts

2026-06-07T16:48:07.866958+00:00

system

start

[
  "run started"
]

[]

2026-06-07T16:48:08.374708+00:00

planner

[
  "Explicit seed task provided on the command line."
]

[]

2026-06-07T16:51:10.403505+00:00

worker

[
  "## Decisions",
  "- **Confirmed loop04 content repair is complete.** `production/AMZN/` and `loop04/artifacts/production/AMZN/` both pass `python3 validator.py` (exit 0, status PASS). Cross-ticker MSFT/ORCL/EQIX fallbacks are gone; all rows are `ticker=AMZN` with SEC EDGAR + Amazon IR primary URLs.",
  "- **Loop04 machine-gate blocker was infrastructure, not data.** Reviewer Codex failed with `OSError: [Errno 7] Argument list too long` \u2014 no content verdict was produced. Re-emitting the verified canonical artifacts for loop05 harness extraction.",
  "- **Hyperscaler capex-proxy pattern (ORCL/MSFT).** Amazon does not disclose MW/GW; `normalized_gw` and GW chart columns stay blank. Capex actuals from `us-gaap:PaymentsToAcquireProductiveAssets` in 10-Q/10-K; guidance vintages from MD&A + SEC 8-K ex99 ($200B FY2026 from FY2025-FQ4). FY2025 guidance ($105B\u2192$118B\u2192$125B) is derived with explicit `is_derived=1` caveats.",
  "- **FY2026 guidance rollup anchor:** `GUIDANCE_ORIGIN_QUARTER=FY2025-FQ4`, `GUIDANCE_TARGET=FY2026`, original/latest both $200B from first FY2026 guide vintage.",
  "## Evidence",
  "**PRIMARY**",
  "- SEC EDGAR CIK 1018724: FY2024\u2013FY2026 10-Q/10-K cash-flow capex (e.g. `https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm`)"
]

[]

build_timeseries.py, capacity_timeseries.csv, company_gw_chart.html, extractor.py, forecast_vintage.csv, source_manifest.csv, validator.py

2026-06-07T16:53:23.776088+00:00

reviewer

PASS

[
  "review cycle 1"
]

[]

build_timeseries.py, capacity_timeseries.csv, company_gw_chart.html, extractor.py, forecast_vintage.csv, source_manifest.csv, validator.py

2026-06-07T16:53:23.880078+00:00

machine_gate

PASS

[
  "expected artifact validators and production gates passed"
]

[]

build_timeseries.py, capacity_timeseries.csv, company_gw_chart.html, extractor.py, forecast_vintage.csv, source_manifest.csv, validator.py

Generated 2026-06-07T16:53:24.307985+00:00

Original goal

Current task

Completed tasks

Task history

Acceptance criteria

Expected artifacts

Artifacts

Agent calls & decisions