Merge pull request #4 from tecunningham/codex/conduct-literature-search-for-data-sources

Add local tech-dashboard ingestion pipeline and mark LLM-generated files

Author: tecunningham

posts/2026-02-14-tech-progress-dashboard.data-sources.llm.md

@@ -0,0 +1,222 @@
+# Tech Progress Dashboard — Data Source Landscape (Research Log)
+
+_Last updated: 2026-02-14_
+
+## 1) Scope definition (v1)
+
+### Goal
+Build a maintainable, high-trust dataset of **technology progress time series** for use in `posts/2026-02-14-tech-progress-dashboard.llm.qmd`.
+
+### Time scope
+- Include data points from **1950 onward**.
+- Prefer series with annual observations; allow coarser frequencies only when they represent key frontier milestones.
+
+### Metric scope
+Include series that track one of the following:
+1. **Cost decline** (e.g., inflation-adjusted $/unit performance)
+2. **Performance improvement** (e.g., speed, efficiency, accuracy, yield)
+3. **Scale and diffusion** (e.g., deployment, adoption, capacity)
+4. **Input intensity** (e.g., compute used in training, R&D effort)
+
+### Domain scope
+- Compute hardware and semiconductors
+- AI/ML model capability and compute
+- Energy generation and storage
+- Communications and networking
+- Biotech/health technology
+- Manufacturing and industrial productivity
+- Agriculture (selected benchmark technologies)
+- Transportation systems
+
+### Out-of-scope (for now)
+- Pre-1950 historical reconstructions
+- Single-paper one-off benchmark results with no replicable update path
+- Data locked behind expensive subscriptions unless a consistent legal extraction path exists
+
+---
+
+## 2) Quality bar and scoring rubric
+
+Each source is scored on 5 dimensions (1–5):
+
+- **Coverage**: length and completeness (1950+ preferred)
+- **Methodological quality**: transparent definitions, data provenance, revision policy
+- **Accessibility**: open download/API, machine-readability, legal reuse clarity
+- **Updateability**: likely to keep updating for future dashboard refreshes
+- **Operational fit**: ease of integrating into normalized long-format pipeline
+
+### Tiers
+- **Tier A (include now)**: total score >= 22/25, no red flags
+- **Tier B (include selectively)**: 17–21, with caveats documented
+- **Tier C (monitor / not yet include)**: <= 16 or major access/provenance issues
+
+---
+
+## 3) Candidate source longlist (deep scan)
+
+## Compute hardware, semiconductors, AI compute
+
+| Source | Example metrics | Coverage | Access | Quality notes | Score | Tier |
+|---|---|---:|---|---|---:|---|
+| Our World in Data (OWID) grapher | Transistors/chip, FLOPS/$, storage price | Varies; many start 1950–1980s | CSV + API-like URL params; open | Strong docs, reproducible, easy extraction; often sourced from underlying institutions | 24 | A |
+| OpenAlex | AI paper counts, citations by field | 1950s+ publication metadata | Open API + snapshots | Massive coverage, good for diffusion proxies; requires careful field classification | 22 | A |
+| Crossref | DOI publication trends, venue metadata | Broad historical coverage | Open API | Good scale; noisier than curated bibliometric sources | 19 | B |
+| arXiv bulk metadata | CS/AI preprint trends | 1991+ | Open bulk + API | Clean for modern AI era, not 1950-complete | 17 | B |
+| TOP500 | Frontier supercomputer LINPACK performance | 1993+ biannual | Public lists/download | Trusted frontier proxy; no 1950s coverage | 18 | B |
+| MLPerf results | Training/inference performance curves | ~2018+ | Public tables | High benchmark quality but short horizon | 14 | C |
+| Epoch AI dataset pages | Frontier model training compute | Mostly 2010+ | Public with mixed machine-readability | Substantive but methodology can evolve; verify each datapoint | 17 | B |
+| SemiAnalysis articles/data snippets | AI hardware economics | Recent | Mostly paywalled | Valuable context but weak reproducible access | 10 | C |
+| Semiconductor Industry Association (SIA) | Industry shipments/sales | Late 20th c.+ | Reports (mixed openness) | Useful aggregate indicators; sometimes paywalled details | 15 | C |
+| FRED (selected semiconductor indexes) | Price indexes, producer prices | Various; often decades | API | Reliable for macro proxies, less direct technical frontier | 20 | B |
+
+## Energy generation, storage, and power systems
+
+| Source | Example metrics | Coverage | Access | Quality notes | Score | Tier |
+|---|---|---:|---|---|---:|---|
+| IEA Data Browser | Electricity generation tech mix, efficiency, prices | Often 1960+ | Mixed: many series free, some restricted | High quality official source; access terms vary | 21 | B |
+| EIA (US Energy Information Administration) | Generation costs, fuel efficiency, capacity | Many series 1949+ | Open API + downloads | Excellent long-run US series, consistent metadata | 24 | A |
+| Lazard LCOE reports | Utility-scale cost benchmarks | ~2009+ | Public PDF | Good for modern cost snapshots; short horizon, methodology changes | 14 | C |
+| NREL ATB | Forward-looking cost/performance assumptions | Mostly 2010+ | Public CSV | Great structure but not long-run historical primary source | 15 | C |
+| BP Statistical Review / Energy Institute | Global energy production/consumption | ~1965+ | Public spreadsheets | Long history, broadly used, transparent revisions | 22 | A |
+| Ember electricity datasets | Power-sector trends | mostly 2000+ | Open downloads | Strong recent coverage, not enough historical depth alone | 16 | C |
+| BNEF battery price announcements | Li-ion pack $/kWh | 2010s+ | Mostly press releases | Important indicator but source data often not fully open | 14 | C |
+| IRENA renewables cost database | Utility-scale technology costs | ~2010+ | Public reports/data | Good methodology, insufficient pre-2000 depth | 16 | C |
+
+## Communications and digital infrastructure
+
+| Source | Example metrics | Coverage | Access | Quality notes | Score | Tier |
+|---|---|---:|---|---|---:|---|
+| ITU (International Telecommunication Union) | Broadband subscriptions, telecom diffusion | Several series from 1960s/70s | Public tables + API-like access | Canonical global telecom source; some usability friction | 22 | A |
+| OECD broadband portal | Prices, speeds, subscriptions | Mostly 1990s+ | Open tables/API | High trust for OECD countries; narrower geography | 20 | B |
+| World Bank WDI | Internet users, secure servers, telecom variables | Many series 1960+ depending variable | Open API + bulk CSV | Excellent operational access and metadata | 23 | A |
+| FCC data (US) | Broadband deployment/performance | Modern era | Public releases | Good US detail but short horizon | 15 | C |
+| Cisco annual internet reports (archival) | Traffic growth | 2000s+ | Public PDFs | Useful context but report methodology evolved | 13 | C |
+
+## Manufacturing, macro productivity, technology diffusion proxies
+
+| Source | Example metrics | Coverage | Access | Quality notes | Score | Tier |
+|---|---|---:|---|---|---:|---|
+| Penn World Table (PWT) | TFP, capital deepening | 1950+ | Open downloads | Standard in growth accounting; model-based estimates need caveats | 22 | A |
+| EU KLEMS / World KLEMS | Sector productivity and ICT capital | Often 1970+ | Open downloads | Great granularity, weaker pre-1970 coverage | 20 | B |
+| Conference Board Total Economy Database | Productivity levels and growth | Mid-20th c.+ | Public downloads | High-quality macro benchmark | 21 | B |
+| UNIDO INDSTAT (where accessible) | Industrial output and intensity | Varies | Mixed access | Valuable, but access constraints in some components | 15 | C |
+| FRED macro series | Price deflators, productivity, industrial production | Many start pre-1950 | Open API | Strong operational backbone for normalization/deflation | 23 | A |
+
+## Agriculture and food technology
+
+| Source | Example metrics | Coverage | Access | Quality notes | Score | Tier |
+|---|---|---:|---|---|---:|---|
+| FAOSTAT | Crop yields, inputs, production | 1961+ | Open API + bulk | Global, consistent, machine-readable | 23 | A |
+| USDA Quick Stats | US yields, adoption, productivity | Many series 1950+ | Open API | Excellent US depth and metadata | 24 | A |
+| Our World in Data agriculture grapher | Harmonized yield/diet/land series | Varies | Open CSV | Great dashboard-ready integration; verify underlying source transformations | 22 | A |
+| ASTI (agricultural R&D indicators) | R&D spending and personnel | Mostly 1980+ | Public tables | Important input proxy, but limited long-run depth | 17 | B |
+
+## Biotech and health technology
+
+| Source | Example metrics | Coverage | Access | Quality notes | Score | Tier |
+|---|---|---:|---|---|---:|---|
+| NIH RePORTER / historical NIH funding | Biomedical R&D inputs | Mid-20th c.+ aggregates | API + files | Good funding signals; variable granularity over time | 20 | B |
+| FDA (Drugs@FDA, device approvals) | Approval counts/timelines | Modern electronic era primarily | Public DB | Robust for modern period; weak earlier machine-readable data | 16 | C |
+| CDC/NCHS (selected tech-related outcomes) | Mortality/outcome proxies | Long historical | Open datasets | Useful as outcomes, less direct as technical frontier measure | 18 | B |
+| GenBank growth statistics (NCBI) | Sequence database scale | 1980s+ | Public | Useful diffusion proxy; not direct performance/cost | 17 | B |
+
+## Patents, innovation intensity, cross-domain
+
+| Source | Example metrics | Coverage | Access | Quality notes | Score | Tier |
+|---|---|---:|---|---|---:|---|
+| USPTO PatentsView | Patent counts, citations, classifications | 1976+ structured modern data; historical text back further | API + bulk | High-quality US patent backbone | 21 | B |
+| EPO PATSTAT (via access agreements) | Global patent panel data | Long historical | Restricted/licensed | Very strong data quality but access barriers | 15 | C |
+| WIPO IP statistics | Patent applications by country/field | Multi-decade | Public downloads | Strong international comparability at aggregate level | 21 | B |
+| OECD STAN / ANBERD / MSTI | R&D, industry innovation stats | Often 1970+ | Mixed (some open, some subscription) | High quality but access friction | 18 | B |
+
+---
+
+## 4) Shortlist for immediate integration (recommended subset)
+
+These are highest-priority because they combine 1950+ relevance, quality, and accessible extraction:
+
+1. **OWID Grapher series** (cross-domain normalized quick wins)
+2. **EIA API** (US energy long-run, 1949+ in many series)
+3. **World Bank WDI API** (cross-country diffusion and infrastructure)
+4. **FAOSTAT + USDA Quick Stats** (agriculture and biological productivity anchors)
+5. **FRED API** (deflators + macro-normalization backbones)
+6. **PWT** (macro productivity context)
+7. **BP/Energy Institute Statistical Review** (global energy long-run)
+8. **ITU datasets** (communications diffusion)
+
+Secondary additions after MVP:
+- WIPO/PatentsView for innovation intensity overlays
+- TOP500 for compute frontier scaling (1993+)
+- Epoch AI model-compute series (careful provenance tagging)
+
+---
+
+## 5) Proposed ongoing pipeline (for future source additions)
+
+## A. Repository structure
+
+- `posts/2026-02-14-tech-progress-dashboard.data-sources.llm.md` (this human-readable research log)
+- `posts/2026-02-14-tech-progress-dashboard.source-registry.llm.csv` (machine-readable source registry)
+- Future (optional): `posts/2026-02-14-tech-progress-dashboard.raw/` for fetched raw extracts
+- Future (optional): `posts/2026-02-14-tech-progress-dashboard.processed.csv` for normalized long-format panel
+
+## B. Source intake workflow
+
+1. **Discover** candidate source and add one row to registry with status `candidate`
+2. **Score** using rubric (coverage/method/access/updateability/fit)
+3. **License check**: verify legal reuse and citation requirements
+4. **Extraction test**: confirm scriptable fetch and parse
+5. **Quality audit**:
+   - Missingness profile
+   - Unit consistency
+   - Breaks/redefinitions over time
+   - Update cadence and revision handling
+6. **Decision gate**:
+   - Promote to `approved` if Tier A or strong Tier B with caveats
+   - Otherwise set `watchlist` or `rejected`
+7. **Integrate** into dashboard canonical schema
+8. **Document** provenance note and assumptions in the post
+
+## C. Canonical schema for series integration
+
+For each final series keep:
+- `id`, `name`, `domain`, `metric_type`, `frontier_class`, `direction`, `unit`
+- `year`, `value`
+- `source_name`, `source_url`
+- `provenance_note`
+- `quality_tier`, `last_verified_at`
+
+## D. Quality controls before publish
+
+- Check each selected series has at least 8 observations (or explicit frontier exception)
+- Verify at least one citation URL resolves
+- Validate no mixed nominal/real units without explicit deflator
+- Confirm directionality (`higher_better` / `lower_better`) and transformations
+- Keep changelog of source revisions affecting historical points
+
+---
+
+## 6) Known risks and mitigation
+
+- **Risk: methodology changes over time** (e.g., benchmark definitions)
+  - Mitigation: version-tag each source and keep per-series caveats.
+- **Risk: paywalled sources become inaccessible**
+  - Mitigation: prefer open replicas or use source only for contextual triangulation.
+- **Risk: frontier series are sparse and noisy**
+  - Mitigation: require minimum-point rule + robustness checks on trend fits.
+- **Risk: mixed geographic scope (US-only vs global)**
+  - Mitigation: encode geography explicitly and avoid combining incompatible units.
+
+---
+
+## 7) Next execution step
+
+For the next pass, prioritize building fetch scripts and adding first approved series from:
+- OWID
+- EIA
+- World Bank WDI
+- FAOSTAT
+- USDA Quick Stats
+- FRED
+
+Then rerun dashboard with provenance table showing quality tier and last-verified date.

…s/2026-02-14-tech-progress-dashboard.qmd …26-02-14-tech-progress-dashboard.llm.qmdposts/2026-02-14-tech-progress-dashboard.qmd renamed to posts/2026-02-14-tech-progress-dashboard.llm.qmd posts/2026-02-14-tech-progress-dashboard.qmd renamed to posts/2026-02-14-tech-progress-dashboard.llm.qmd

@@ -8,6 +8,16 @@ format: html
 
 <script src="https://cdn.plot.ly/plotly-2.35.2.min.js"></script>
 
+## Data-source scope and research log
+
+- Scope is currently restricted to **1950 onward** data.
+- Source search, quality scoring, and pipeline notes live in:
+  - `posts/2026-02-14-tech-progress-dashboard.data-sources.llm.md`
+  - `posts/2026-02-14-tech-progress-dashboard.source-registry.llm.csv`
+- Local normalized ingestion output lives in:
+  - `posts/data/tech-progress-dashboard.llm/normalized-series.llm.csv`
+  - Refresh command: `python tools/ingest_tech_progress_dashboard.py`
+
 <style>
 #tpd-controls {
   display: grid;
@@ -514,4 +524,4 @@ This MVP is built to support **forecasting and measuring AI's impact on technolo
 5. Data inclusion threshold: at least 4 time points, direction-of-improvement defined, source URL recorded.
 6. Fitting method: per-series log-linear fit on historical data (`log(value) = a + b * year`).
 7. v1 excludes causal attribution and forecasting beyond observed history.
-8. v1 data are seed-series and should be upgraded to automated ingestion in a follow-up.
+8. v1 data are seed-series and should be upgraded to automated ingestion in a follow-up.

posts/2026-02-14-tech-progress-dashboard.source-registry.llm.csv

@@ -0,0 +1,16 @@
+source_id,source_name,domain_group,example_metrics,coverage_start,coverage_end,access_mode,coverage_score,method_score,access_score,updateability_score,operational_fit_score,total_score,tier,status,notes
+owid_grapher,Our World in Data Grapher,multi-domain,"transistors_per_chip; gpu_perf_per_dollar; crop_yields",1950,2025,open_csv,5,5,5,4,5,24,A,approved,"Strong first-line source; verify underlying original source for each series"
+eia_api,US EIA API,energy,"generation; prices; efficiency",1949,2025,open_api,5,5,5,5,4,24,A,approved,"Excellent long-run US energy coverage"
+world_bank_wdi,World Bank WDI,communications_macro,"internet_users; broadband; infrastructure",1960,2025,open_api,5,4,5,4,5,23,A,approved,"Cross-country comparability with strong metadata"
+faostat,FAOSTAT,agriculture,"yields; production; inputs",1961,2025,open_api,5,5,5,4,4,23,A,approved,"Global ag backbone"
+usda_quickstats,USDA Quick Stats,agriculture,"us crop yields; adoption",1950,2025,open_api,5,5,5,4,5,24,A,approved,"Deep US historical coverage"
+fred_api,FRED,macro,"deflators; productivity; industrial indexes",1950,2025,open_api,5,5,5,4,4,23,A,approved,"Normalization backbone and supporting indicators"
+penn_world_table,Penn World Table,macro,"tfp; capital; gdp per worker",1950,2023,open_download,5,5,4,4,4,22,A,approved,"Model-based estimates require caveats"
+energy_institute_review,Energy Institute Statistical Review,energy,"global energy production and use",1965,2024,open_spreadsheet,4,5,4,4,5,22,A,approved,"Long-run global energy comparisons"
+itu_data,ITU datasets,communications,"subscriptions; access diffusion",1960,2025,open_table,5,5,4,4,4,22,A,approved,"Canonical telecom source"
+top500,TOP500,compute,"frontier_supercomputer_flops",1993,2025,public_download,3,5,4,5,4,21,B,candidate,"Frontier compute proxy; post-1993 only"
+patentsview,USPTO PatentsView,innovation,"patent_counts; citations",1976,2025,open_api,4,4,5,4,4,21,B,candidate,"Great for innovation intensity overlays"
+wipo_stats,WIPO IP Statistics,innovation,"country patent applications",1970,2024,public_download,4,5,4,4,4,21,B,candidate,"Global aggregate patent trends"
+epoch_ai,Epoch AI model datasets,ai_models,"training_compute; parameter_counts",2010,2025,public_web,2,4,4,4,3,17,B,watchlist,"Useful but needs per-point provenance auditing"
+bnef_battery,BloombergNEF battery prices,energy_storage,"li_ion_pack_price",2013,2025,press_release,2,4,2,4,2,14,C,watchlist,"Important metric but restricted underlying dataset"
+mlperf,MLPerf,ai_benchmarks,"training_time; inference_perf",2018,2025,public_table,1,5,4,5,3,18,B,watchlist,"High quality but short horizon for 1950+ dashboard"

posts/data/tech-progress-dashboard.llm/README.llm.md

@@ -0,0 +1,18 @@
+# Tech Progress Dashboard local data cache
+
+This folder contains locally cached and normalized data used by:
+- `posts/2026-02-14-tech-progress-dashboard.llm.qmd`
+
+## Files
+- `normalized-series.llm.csv`: canonical long-format panel used for chart ingestion.
+- `ingest-log.llm.md`: last ingestion summary (counts and replacement coverage).
+- `raw/`: raw downloaded source extracts (currently OWID grapher CSVs).
+
+## Refresh
+Run:
+
+```bash
+python tools/ingest_tech_progress_dashboard.py
+```
+
+The script parses seed rows from the post, refreshes downloadable sources, and writes normalized output.

posts/data/tech-progress-dashboard.llm/ingest-log.llm.md

@@ -0,0 +1,10 @@
+# Tech Progress Dashboard ingestion log
+
+- Source post: `posts/2026-02-14-tech-progress-dashboard.llm.qmd`
+- Seed rows parsed from embedded JS: **84**
+- Series replaced with fresh OWID downloads: **transistors_per_chip, us_corn_yield, solar_pv_module_price, hdd_cost_per_gb**
+- Final normalized rows: **260**
+
+## Notes
+- This pipeline keeps seed data for sources that are not yet script-downloaded (e.g., BNEF, Green500, NHGRI page-based data).
+- All rows are normalized to a single long-format CSV for local dashboard ingestion.