input_id,description,used_in,current_value_or_version,source_name,source_url,machine_readable,check_method,cadence,trigger,last_checked,notes mlenergy_lm_arena_chat,"GPU energy per output token for every measured open-weight model (task lm-arena-chat, 50 ms ITL operating point)","recompute_v3_final.py load(), MEMBERS (PROXY kept as the named-proxy sensitivity); recompute_v3_lifecycle.py class_members(), class_inputs(); build_detailed.py model_factors(); snapshot mlenergy_v3_lm-arena-chat_2026-09-29.json","ML.ENERGY Leaderboard v3, index last_updated 2026-02-16, 328 configurations; file Last-Modified 2026-09-23; snapshot identical to remote on 2026-10-02",ML.ENERGY Leaderboard,https://ml.energy/leaderboard/data/tasks/lm-arena-chat.json,yes,json_field,on release,"last_updated or the SHA-256 of the task file changes: save the new JSON under a dated name, point MLE in recompute_v3_final.py to it, add new models or GPUs to MEMBERS/PROXY, rerun recompute_v3_lifecycle.py and build_detailed.py; a new GPU generation also revisits the B200 scenario",2026-10-02,"last_updated is read from /leaderboard/data/index.json (same site); the task JSON has no date field of its own. Last-Modified (2026-09-23) is later than last_updated, so Last-Modified alone is not a release signal." ember_yearly,"Grid carbon intensity by country (lifecycle gCO2e/kWh), calendar year 2023","recompute_v3_final.py EMBER_G, EMBER_US_2023_G; build_detailed.py grid_intensity_all.csv; ember_yearly_full_release_long_format.csv","Ember yearly_full_release_long_format.csv, 49,079,981 bytes, file Last-Modified 2026-06-23; dataset uses calendar year 2023",Ember,https://storage.googleapis.com/emb-prod-bkt-publicdata/public-downloads/yearly_full_release_long_format.csv,yes,http_last_modified,annual,File replaced: open it and check the years present and complete per area; if a newer year is adopted change YEAR in build_detailed.py and EMBER_G in recompute_v3_final.py together with the eGRID year; Rerun recompute_v3_lifecycle.py and build_detailed.py; expect every grid-dependent factor to move,2026-10-02,"The local copy already contains 2000-2025 while the dataset deliberately uses 2023 (to match eGRID2023), so a changed Last-Modified does not by itself mean a new year is wanted. Latest year in the file is not read by the script (49 MB; HEAD only). URL from assumptions_B.md; landing page https://ember-energy.org/data/electricity-data-explorer/." ember_us_subnational,US total 2023 grid intensity (392.755 g CO2e/kWh) used to scale eGRID subregion rates to the Ember lifecycle basis,"recompute_v3_final.py EMBER_US_2023_G, US_LIFECYCLE_SCALE","Ember US subnational electricity data, US total 2023 = 392.755 g CO2e/kWh",Ember,https://ember-energy.org/data/electricity-data-explorer/,no,manual,annual,Ember publishes a newer US year: update EMBER_US_2023_G in recompute_v3_final.py in step with the eGRID year; Rerun recompute_v3_lifecycle.py and build_detailed.py; all US rows move,2026-10-02,"UNSURE: no direct file URL for the US subnational series is documented in the spec files, only the explorer page. The main yearly file gives 392.9 g for USA 2023 (not 392.755), so the two are different series; do not substitute one for the other." egrid_summary_tables,"eGRID subregion CO2e total output emission rates (lb/MWh), 27 subregions plus US total","recompute_v3_final.py EGRID_CO2E_LB, EGRID_US_CO2E_LB; egrid2023_subregions.csv; build_detailed.py grid_intensity_all.csv, cloud_region_factors.csv","eGRID2023 Summary Tables Rev 2 (2025-06-12), Table 1; PDF Last-Modified 2025-06-16, 410,720 bytes",US EPA eGRID,https://www.epa.gov/system/files/documents/2025-06/summary_tables_rev2.pdf,yes,http_last_modified,annual,"PDF replaced (new revision): recheck Table 1 for all subregions against egrid2023_subregions.csv, update the CSV and EGRID_CO2E_LB; Rerun recompute_v3_lifecycle.py and build_detailed.py; US rows and US cloud regions move",2026-10-02, egrid_release_page,EPA eGRID landing page (announces new data years and revisions),egrid2023_subregions.csv (year choice); cloud_regions.csv eGRID subregion mapping,eGRID2023 is the year used; eGRID2023 rev 2 current at last check,US EPA eGRID,https://www.epa.gov/egrid,yes,content_hash,annual,"A newer eGRID year is announced: download its summary tables, rebuild egrid2023_subregions.csv for that year, repeat the cloud-region subregion mapping if boundaries changed, pair with the Ember year; Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02,"Hash is of the visible page text; it also changes for news items and edits unrelated to a release, so read the page when flagged. The subregion shapefile used for region mapping (egrid2023_subregions.zip, named in cloud_regions.csv notes) has no URL in the spec files: UNSURE where it was downloaded from." cloud_regions_aws,List of public AWS regions,cloud_regions.csv (aws rows); cloud_region_factors.csv,"37 AWS regions in dataset, listed October 2026",AWS,https://docs.aws.amazon.com/global-infrastructure/latest/regions/aws-regions.html,yes,json_field,semi-annual,"New region: add it to cloud_regions.csv (country, eGRID subregion for US sites with confidence and note) and rebuild (python build_detailed.py); regions in the dataset but gone from the page: check for retirement; then rerun build_detailed.py",2026-10-02,"Region ids extracted by regex. cn-north-1, cn-northwest-1 and eusc-de-east-1 come from other pages (an aws.amazon.com whats-new article of 2020-04 and the AWS blog post on the European Sovereign Cloud, URLs in cloud_regions.csv) and always appear as 'in dataset, not on page'." cloud_regions_azure,List of public Azure regions,cloud_regions.csv (azure rows); cloud_region_factors.csv,"63 Azure regions in dataset, listed October 2026",Microsoft Azure,https://learn.microsoft.com/azure/reliability/regions-list,yes,json_field,semi-annual,"New region: add it to cloud_regions.csv (country, eGRID subregion for US sites with confidence and note) and rebuild (python build_detailed.py); regions in the dataset but gone from the page: check for retirement; then rerun build_detailed.py",2026-10-02,"Programmatic names are read from the last cell of each HTML table row. The 6 Azure China regions come from https://learn.microsoft.com/azure/china/overview-regions and always appear as 'in dataset, not on page'. 20 region mappings in the dataset rest on a state-only location (see notes column of cloud_regions.csv)." cloud_regions_gcp,List of Google Cloud regions,cloud_regions.csv (gcp rows); cloud_region_factors.csv,"43 Google Cloud regions in dataset, listed October 2026",Google Cloud,https://cloud.google.com/compute/docs/regions-zones,yes,json_field,semi-annual,"New region: add it to cloud_regions.csv (country, eGRID subregion for US sites with confidence and note) and rebuild (python build_detailed.py); regions in the dataset but gone from the page: check for retirement; then rerun build_detailed.py",2026-10-02,Region ids extracted by regex from the regions-and-zones page (https://cloud.google.com/about/locations also exists but is not the source named in cloud_regions.csv). arxiv_elsworth_google,"Google serving disclosure: 0.24 Wh per median Gemini text prompt, energy split (accelerator 0.14, CPU/DRAM 0.06, idle 0.02, overhead 0.02), 232 g/M","recompute_v3_lifecycle.py H low 1.43, I central 1.10, PUE low 1.09; recompute_v3_final.py GOOGLE_WH, GOOGLE_G_M","Elsworth et al., arXiv:2508.15734, version v1 (2025-08-21)",arXiv,https://arxiv.org/abs/2508.15734,yes,json_field,annual,"New arXiv version, journal version with changed numbers, or a new Google disclosure of a full serving stack: recheck Table 1 split and 0.24 Wh; update H, I, PUE low and GOOGLE_*; Rerun recompute_v3_lifecycle.py and build_detailed.py; A first-party per-token energy for a served model moves that model to the hybrid method, or to Tier 3 if the provider publishes a CO2e figure (paper s3.9)",2026-10-02,Script checks only the arXiv version number. google_efficiency_page,Google fleet-wide PUE,PUE low 1.09 in recompute_v3_lifecycle.py,2025 fleet-wide average PUE 1.09 (trailing twelve months); page Last-Modified 2026-08-25,Google data centers,https://datacenters.google/efficiency/,yes,content_hash,annual,Fleet PUE restated: update PUE low in recompute_v3_lifecycle.py and the S6 PUE row; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,"Paper cites https://www.google.com/about/datacenters/efficiency/ (older address); datacenters.google/efficiency is the URL used in overhead_evidence.md. Hash is of visible text, so it can change for unrelated edits; read the PUE figure when flagged." latif_host_uplift,Host (non-GPU) server energy on an 8-GPU H100 node,H central 1.50 in recompute_v3_lifecycle.py,"Latif et al. 2025, IEEE Access 13:61740-61747; H = 1.50 taken via SAIG CAIEC-0926 s5.1",IEEE Access,https://doi.org/10.1109/ACCESS.2025.3554728,no,manual,annual,"A new measurement of non-GPU energy on a GPU server, or any first-party full-serving-stack measurement: revisit H (1.43 / 1.50 / 1.667) in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02,UNSURE: H = 1.50 is the value SAIG derives from Latif; it was not re-derived from the IEEE paper in the source notes. arxiv_tokenpowerbench,"TokenPowerBench: non-GPU share of server energy ('over 60%' of GPU energy), a ceiling of 1.667 on H, kept as a consistency check only",H ceiling check in recompute_v3_lifecycle.py (1/0.60); H high is 1.88 from the Schneider TPU ratios,"arXiv:2512.03024, version v1 (2025-12-02)",arXiv,https://arxiv.org/abs/2512.03024,yes,json_field,annual,New version or follow-up with a different host share: revisit H high in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,Spec cites the PDF https://arxiv.org/pdf/2512.03024; the abs page is used for the version check. arxiv_oviedo,Oviedo et al.: provisioned/idle capacity uplift and production serving energy (1.03 Wh/1K comparison),I range 1.05-1.20 (via SAIG); comparison line in numbers_v3_lifecycle.json,"arXiv:2509.20241, version v2 (2026-06-09); Joule 2026, 102430",arXiv,https://arxiv.org/abs/2509.20241,yes,json_field,annual,New version or new measurement of idle/provisioned energy: revisit I (1.05/1.10/1.20) in recompute_v3_lifecycle.py and the Oviedo comparison; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,"UNSURE: the idle range is taken via SAIG, not from the primary text, and it is not recorded which arXiv version (v1 or v2) was read; v2 is dated 2026-06-09, so check what changed from v1 first." saig_caiec_source,"Sustainable AI Group 'Closed AI Emissions Calculation' (source of H central, I range, construction low, development ratio, PUE high 1.45)","recompute_v3_lifecycle.py H, I, K_CONSTR low, D_DEV, PUE high",CAIEC-0926 (29 Sep 2026); web page Last-Modified 2026-09-27,Sustainable AI Group (GitHub),https://raw.githubusercontent.com/sustainableaigroup/ai-emissions/main/index.qmd,yes,content_hash,quarterly,"Source text changes: read the diff of s5.1 (host), idle, s8 (construction), development ratio and PUE; update affected constants in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02,Rendered version: https://reports.sustainableaigroup.com/ai-emissions (paper reference). This is a secondary compilation that several constants route through; re-verify against primaries when it changes. Hash is of the raw .qmd. aws_pue_wue_csv,AWS published PUE and WUE by region (2022-2025); regional on-site WUE rows and AWS fleet PUE,"v4_inputs/water_onsite_selected.csv (AWS rows, read by recompute_v3_lifecycle.py at run time); PUE evidence","AWS aws-pue-wue.csv, years 2022-2025, Last-Modified 2026-09-08; 2025 values used (e.g. eu-north-1 WUE 0.02, us-east-1 0.06, ap-southeast-1 1.57)",Amazon Sustainability,https://sustainability.aboutamazon.com/content/dam/sustainability-marketing-site/12-the-cloud/aws-pue-wue.csv,yes,csv_value,annual,New year published (2026) or any restated value for the listed regions: update recommended_wue_l_per_kwh in v4_inputs/water_onsite_selected.csv; Rerun recompute_v3_lifecycle.py and build_detailed.py; water rows move; recheck AWS PUE (1.15 global 2024 cited in paper s3.3),2026-10-02,"Script compares the regions the dataset uses (GLOBAL, Europe, North America and 9 region ids) for 2025 and the latest year in the file. The paper cites the 2024 Amazon Sustainability Report for PUE 1.15; the CSV gives GLOBAL PUE 1.14 for 2025. eu-west-3 has PUE only, no WUE." microsoft_efficiency_page,Microsoft datacentre efficiency page: WUE definition and fleet PUE/WUE,v4_inputs/water_onsite_selected.csv (Microsoft rows); PUE evidence,Page as of 2026-10-02; only Microsoft source for WUE (the Fact Sheet has none),Microsoft,https://datacenters.microsoft.com/sustainability/efficiency/,yes,content_hash,annual,Restated PUE/WUE or new fiscal-year values: update Microsoft-derived rows in v4_inputs/water_onsite_selected.csv; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,Visible-text hash; the page Last-Modified moves on every rebuild so it is not used. microsoft_fact_sheet,"Microsoft Environmental Data Fact Sheet, Table 15 (FY25 datacentre water and electricity by location)","v4_inputs/water_onsite_selected.csv rows europe-west4, us-south-central, us-central1 (derived site ratios)","2026 Environmental Data Fact Sheet PDF, 3,912,081 bytes, Last-Modified 2026-07-16 (FY25)",Microsoft,https://aka.ms/SustainabilityFactsheet2026,yes,http_last_modified,annual,Reissued or new edition (FY26): re-derive the three site WUE rows from Table 15 (withdrawal ML / MWh x 1000); Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,"The short link carries the edition year; the next edition will probably need a new link (for example aka.ms/SustainabilityFactsheet2027, not verified): look for it manually each autumn." meta_env_data_index,Meta Environmental Data Index: fleet WUE 0.19 (2024) and per-site water,"v4_inputs/water_onsite_selected.csv (us-south, Meta rows)","2025 Environmental Data Index PDF, 2,266,742 bytes, Last-Modified 2025-10-17 (CY2024)",Meta,https://sustainability.atmeta.com/asset/2025-environmental-data-index/,yes,http_last_modified,annual,"Reissued, or a new edition published: re-derive Meta rows (Stanton Springs GA); Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02,"The URL carries the edition year, so the script detects a reissue of the 2025 PDF only. Find the 2026 edition manually on https://sustainability.atmeta.com/ (Meta published the 2025 edition in October); its URL is not guessed here." google_env_report,Google Environmental Report: site and fleet water and PUE,v4_inputs/water_onsite_sources.csv rows 44-54 (Google sites); PUE,Google 2026 Environmental Report (web page),Google,https://sustainability.google/reports/google-2026-environmental-report/,yes,content_hash,annual,New report year or restated values: update Google rows in v4_inputs/water_onsite_sources.csv and selected rows; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,URL carries the report year; the 2027 report needs a manual search. Visible-text hash of a 32 KB page that may be only a landing page for the PDF: UNSURE whether the figures themselves are on the page. other_operator_wue,Other operators' published WUE (reference rows only: Equinix 0.91 for 2025; Oracle publishes none),v4_inputs/water_onsite_sources.csv reference rows; not used in a factor,Equinix 2025 Sustainability Data Summary p.7; Oracle FY25 datasheet has no WUE,Equinix; Oracle,https://sustainability.equinix.com/ ; https://www.oracle.com/a/ocom/docs/social-impact-datasheet.pdf,no,manual,annual,"Oracle starts publishing WUE, or Equinix restates: add or update reference rows; no factor changes unless a region row is replaced",2026-10-02,"The Equinix site returned HTTP 403 to a scripted request, hence manual." uptime_survey,Uptime Institute global average PUE (1.56),Scenario value in paper s3.3 (PUE row); not a model constant,"Uptime Global Data Center Survey 2024, industry average PUE 1.56",Uptime Institute,https://uptimeinstitute.com/resources/research-and-reports/uptime-institute-global-data-center-survey-results-2024,no,manual,annual,New survey with a different average: update the scenario sentence in the paper; PUE high (1.45) stays tied to LBNL unless the high case is redefined,2026-10-02,Survey results page; the link for a later survey is not documented in the sources. lbnl_report,LBNL 2024 US Data Center Energy Usage Report (US average PUE via Fig. 4.6; indirect water and emissions; 176 TWh in 2023),recompute_v3_lifecycle.py PUE high 1.45 (via SAIG); v4_inputs/ewif_sources.md cross-check,"LBNL-2001637 (Shehabi et al., Dec 2024); PDF Last-Modified 2025-06-18",LBNL / eScholarship,https://escholarship.org/content/qt32d6m0d1/qt32d6m0d1.pdf,yes,http_last_modified,on release,Report reissued or superseded: re-read Fig. 4.6 (US average PUE) and Section 5 (indirect water and emissions); update PUE high in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,"PUE 1.45 is quoted from SAIG, which cites LBNL Fig. 4.6; the figure itself was not re-read in the source notes (UNSURE). Landing page https://escholarship.org/uc/item/32d6m0d1. CHANGED, NEEDS REVIEW (2026-10-03): the supplement review reports a 2025 Update report, LBNL-2001758 (18 June 2026), which supersedes the report this row watches; its US average PUE and Fig. 4.6 equivalent must be re-read before PUE high (1.45) is kept. refresh_state.json carries a stale marker so the next run lists this row under CHANGED." dell_pcf_xe9680,Dell PowerEdge XE9680 product carbon footprint,"recompute_v3_lifecycle.py M_MFG, M_TRANSPORT","Manufacturing 2,872 kg, transport 712 kg (M_MFG low 2,872; with the XE9680L gives M_TRANSPORT central 672 (mean of 712 and 632))",Dell Technologies,https://www.delltechnologies.com/asset/en-us/products/servers/briefs-summaries/poweredge-xe9680.pdf,yes,content_hash,annual,PDF content changes or a new PCF names the accelerator: re-read manufacturing and transport lines; update M_MFG / M_TRANSPORT ranges in recompute_v3_lifecycle.py (and EMB_M_RANGE in recompute_v3_final.py); Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,PAIA v1.4.6 report; Dell leaves GPU model and quantity blank and the Use stage assumes PUE 1.0 (overhead_evidence.md section E). The /asset/ URL redirects to /assetlink/doc/; the script follows redirects. dell_pcf_xe9680l,Dell PowerEdge XE9680L product carbon footprint,"recompute_v3_lifecycle.py M_MFG, M_TRANSPORT","Manufacturing 3,065 kg, transport 632 kg (with the XE9680 gives M_TRANSPORT central 672)",Dell Technologies,https://www.delltechnologies.com/asset/en-us/products/servers/briefs-summaries/poweredge-xe9680l-pcf-report.pdf,yes,content_hash,annual,PDF content changes or a new PCF names the accelerator: re-read manufacturing and transport lines; update M_MFG / M_TRANSPORT ranges in recompute_v3_lifecycle.py (and EMB_M_RANGE in recompute_v3_final.py); Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,PAIA v1.4.6 report; Dell leaves GPU model and quantity blank and the Use stage assumes PUE 1.0 (overhead_evidence.md section E). The /asset/ URL redirects to /assetlink/doc/; the script follows redirects. dell_pcf_xe9685l,Dell PowerEdge XE9685L product carbon footprint,"recompute_v3_lifecycle.py M_MFG, M_TRANSPORT","Manufacturing 4,241 kg, transport 530 kg (M_MFG high 4,241)",Dell Technologies,https://www.delltechnologies.com/asset/en-us/products/servers/briefs-summaries/poweregde-xe9685l-pcf-report.pdf,yes,content_hash,annual,PDF content changes or a new PCF names the accelerator: re-read manufacturing and transport lines; update M_MFG / M_TRANSPORT ranges in recompute_v3_lifecycle.py (and EMB_M_RANGE in recompute_v3_final.py); Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,PAIA v1.4.6 report; Dell leaves GPU model and quantity blank and the Use stage assumes PUE 1.0 (overhead_evidence.md section E). The /asset/ URL redirects to /assetlink/doc/; the script follows redirects. The file name contains Dell's own typo 'poweregde'. dell_pcf_xe7745,Dell PowerEdge XE7745 product carbon footprint,"recompute_v3_lifecycle.py M_MFG, M_TRANSPORT","Manufacturing 4,009 kg, transport 427 kg (M_TRANSPORT low 427)",Dell Technologies,https://www.delltechnologies.com/asset/en-us/products/servers/briefs-summaries/poweredge-xe7745-pcf-report.pdf,yes,content_hash,annual,PDF content changes or a new PCF names the accelerator: re-read manufacturing and transport lines; update M_MFG / M_TRANSPORT ranges in recompute_v3_lifecycle.py (and EMB_M_RANGE in recompute_v3_final.py); Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,PAIA v1.4.6 report; Dell leaves GPU model and quantity blank and the Use stage assumes PUE 1.0 (overhead_evidence.md section E). The /asset/ URL redirects to /assetlink/doc/; the script follows redirects. nvidia_hgx_h100_pcf,"NVIDIA HGX H100 product carbon footprint summary (baseboard 1,312 kg)",Cross-check of M_MFG; B200 embodied note in paper,"Datasheet 4117701, July 2025; PDF Last-Modified 2025-08-06, 152,680 bytes",NVIDIA,https://images.nvidia.com/aem-dam/Solutions/documents/HGX-H100-PCF-Summary.pdf,yes,http_last_modified,annual,"Reissued or a server-level footprint naming the H100: reconsider M_MFG central 3,300 and range in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02,"Baseboard only, not a full server." nvidia_hgx_b200_pcf,"NVIDIA HGX B200 product carbon footprint summary (baseboard 2,274 kg)",B200 scenario: embodied term held at H100 values (no B200 server footprint); paper S6,"Datasheet 4069550, July 2025; PDF Last-Modified 2025-07-30, 263,118 bytes",NVIDIA,https://images.nvidia.com/aem-dam/Solutions/documents/HGX-B200-PCF-Summary.pdf,yes,http_last_modified,annual,A B200 server-level footprint is published: replace the held-at-H100 embodied value in the B200 scenario of recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02, arxiv_schneider_lca,"Schneider et al. Life-Cycle Emissions of AI Hardware: construction per TPU (Table 1: 7.66 g/IT kWh), transport share","K_CONSTR_IT central 0.00766 kg per kWh of server (IT) electricity in recompute_v3_lifecycle.py; also the source of H high 1.88 (machine/chip power ratios, Table 1); transport cross-check","arXiv:2502.01671, version v1 (2025-02-01)",arXiv,https://arxiv.org/abs/2502.01671,yes,json_field,annual,New version or a publication giving construction for GPU data centres: recheck Table 1 and App. A.3/A.5.2; update K_CONSTR in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,"K_CONSTR low 0.0053 comes from Carbon Trust, 'The Carbon Impact of AI Video Generation' (2026), read via SAIG only; no URL is given in the sources, so it has no row of its own (UNSURE)." meta_llama31_model_card,"Meta Llama 3.1 model card: location-based training emissions 420 / 2,040 / 8,930 tCO2e (8B / 70B / 405B)",recompute_v3_final.py TRAIN_T,"Llama 3.1 model card 2024: 420, 2,040, 8,930 tCO2e (location-based)",Meta (GitHub),https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama3_1/MODEL_CARD.md,yes,csv_value,annual,"Any of the three values changes, or Meta publishes training emissions for a newer family: update TRAIN_T in recompute_v3_final.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; every factor's training term moves",2026-10-02,"The paper cites the github.com/.../blob/main/... page; the script reads the same file from raw.githubusercontent.com (same repo path, derived, not quoted in the sources). Table parsed with a regex on the card text." anthropic_model_deprecations,"Anthropic model deprecations: 'not sooner than' retirement dates, at least 60 days notice",v4_inputs/model_service_life.csv; S_MONTHS in recompute_v3_lifecycle.py,"Page as of 2026-10-02; S_MONTHS central 12, low 18, high 6",Anthropic,https://platform.claude.com/docs/en/about-claude/model-deprecations,yes,content_hash,quarterly,"New or changed retirement dates, a changed minimum-notice policy, or a withdrawn floor: update v4_inputs/model_service_life.csv, recompute the median (first-party 12.8 months, n=80) and the S_MONTHS rationale; change S_MONTHS in recompute_v3_lifecycle.py only if the central moves; Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02,Dates are floors ('not sooner than'). Visible-text hash. anthropic_api_release_notes,Anthropic API release notes (launch dates used as model release dates),v4_inputs/model_service_life.csv; S_MONTHS in recompute_v3_lifecycle.py,Page as of 2026-10-02,Anthropic,https://platform.claude.com/docs/en/release-notes/api,yes,content_hash,quarterly,"New model launches or retirements: add rows to v4_inputs/model_service_life.csv (release date, status), then update v4_inputs/model_service_life.csv, recompute the median (first-party 12.8 months, n=80) and the S_MONTHS rationale; change S_MONTHS in recompute_v3_lifecycle.py only if the central moves; Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02,"Long page that changes with every announcement, so it will show CHANGED often; skim for new models only." aws_bedrock_model_lifecycle,AWS Bedrock model lifecycle (models launched from 7 Sep 2026 carry per-model dates),v4_inputs/model_service_life.csv; S_MONTHS in recompute_v3_lifecycle.py,Models launched before 2026-09-07 stay available 'at least 12 months before the EOL date',AWS,https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html,yes,content_hash,quarterly,"New or changed retirement dates, a changed minimum-notice policy, or a withdrawn floor: update v4_inputs/model_service_life.csv, recompute the median (first-party 12.8 months, n=80) and the S_MONTHS rationale; change S_MONTHS in recompute_v3_lifecycle.py only if the central moves; Rerun recompute_v3_lifecycle.py and build_detailed.py; The 12-month Bedrock floor supports S_MONTHS central 12.",2026-10-02, aws_bedrock_model_lifecycle_legacy,"AWS Bedrock model lifecycle (legacy models, per-model EOL dates)",v4_inputs/model_service_life.csv; S_MONTHS in recompute_v3_lifecycle.py,Page as of 2026-10-02,AWS,https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle-legacy.html,yes,content_hash,quarterly,"New or changed retirement dates, a changed minimum-notice policy, or a withdrawn floor: update v4_inputs/model_service_life.csv, recompute the median (first-party 12.8 months, n=80) and the S_MONTHS rationale; change S_MONTHS in recompute_v3_lifecycle.py only if the central moves; Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02, azure_foundry_model_retirements,Microsoft Foundry model retirement policy,v4_inputs/model_service_life.csv; S_MONTHS in recompute_v3_lifecycle.py,Page as of 2026-10-02,Microsoft,https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirements,yes,content_hash,quarterly,"New or changed retirement dates, a changed minimum-notice policy, or a withdrawn floor: update v4_inputs/model_service_life.csv, recompute the median (first-party 12.8 months, n=80) and the S_MONTHS rationale; change S_MONTHS in recompute_v3_lifecycle.py only if the central moves; Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02, azure_foundry_retirement_schedule,Microsoft Foundry model retirement schedule (version and retirement dates per model),v4_inputs/model_service_life.csv; S_MONTHS in recompute_v3_lifecycle.py,Page as of 2026-10-02,Microsoft,https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirement-schedule,yes,content_hash,quarterly,"New or changed retirement dates, a changed minimum-notice policy, or a withdrawn floor: update v4_inputs/model_service_life.csv, recompute the median (first-party 12.8 months, n=80) and the S_MONTHS rationale; change S_MONTHS in recompute_v3_lifecycle.py only if the central moves; Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02,URL taken from the source_url column of model_service_life.csv. openai_deprecations,OpenAI API deprecations (snapshot dates and shutdown dates),v4_inputs/model_service_life.csv; S_MONTHS in recompute_v3_lifecycle.py,Page as of 2026-10-02,OpenAI,https://developers.openai.com/api/docs/deprecations,yes,content_hash,quarterly,"New or changed retirement dates, a changed minimum-notice policy, or a withdrawn floor: update v4_inputs/model_service_life.csv, recompute the median (first-party 12.8 months, n=80) and the S_MONTHS rationale; change S_MONTHS in recompute_v3_lifecycle.py only if the central moves; Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02, google_gemini_deprecations,Google Gemini API deprecations (release and shutdown dates; no minimum published),v4_inputs/model_service_life.csv; S_MONTHS in recompute_v3_lifecycle.py,Page as of 2026-10-02; only gemini-3.1-flash-lite has a shutdown date (12.0 months),Google,https://ai.google.dev/gemini-api/docs/deprecations,yes,content_hash,quarterly,"New or changed retirement dates, a changed minimum-notice policy, or a withdrawn floor: update v4_inputs/model_service_life.csv, recompute the median (first-party 12.8 months, n=80) and the S_MONTHS rationale; change S_MONTHS in recompute_v3_lifecycle.py only if the central moves; Rerun recompute_v3_lifecycle.py and build_detailed.py; ",2026-10-02,"Dates are announced earliest shutdown dates, i.e. floors." openrouter_models,OpenRouter models endpoint: created and expiration_date (second service-life evidence line),v4_inputs/service_life_summary.md F.1 (vi); not a model constant,"464 models, 28 with expiration_date, median (expiration_date - created) 12.5 months (2026-10-02)",OpenRouter,https://openrouter.ai/api/v1/models,yes,json_field,quarterly,Median moves by more than 0.5 month or the number of dated models changes by more than 3: update service_life_summary.md F.1 (vi); it supports S_MONTHS only as a cross-check,2026-10-02,"Weak evidence: created is OpenRouter's listing date, dates cluster on shared cut-offs, and expiration_date contradicts Google's own page for Gemini 2.5. Never cite as a provider commitment." openrouter_rankings_models,OpenRouter per-model rankings (month view): platform-wide input and output tokens per request,v4_inputs/openrouter_per_request.csv; tier2b_defaults.csv; v4_inputs/analyze_openrouter.py,"Month to 2026-10-01: 657 models, 26.85 bn requests, 21,162 input and 555 output tokens per request",OpenRouter,https://openrouter.ai/api/frontend/v1/rankings/models?view=month,yes,json_field,quarterly,Input or output tokens per request move by more than 10%: rerun v4_inputs/analyze_openrouter.py and update the OpenRouter lines in v4_inputs/tier2b_defaults.csv,2026-10-02,"Undocumented frontend endpoint: not in OpenRouter's llms.txt, so its stability and terms are unknown and it may disappear. Cited by path in tier2b_evidence.md." openrouter_rankings_apps,"OpenRouter per-app rankings (month view): tokens per request for Claude Code, Codex, Kilo Code, Cline",v4_inputs/openrouter_app_per_request.csv; tier2b_defaults.csv (coding_agentic),"Month to 2026-10-01: Claude Code 96,884, Codex 85,269, Kilo Code 79,379, Cline 119,346 tokens per request (tier2b_defaults.csv)",OpenRouter,https://openrouter.ai/api/frontend/v1/rankings/apps?view=month,yes,json_field,quarterly,"Any of the four app values moves by more than 10%, or an app leaves the top 20: rerun v4_inputs/analyze_openrouter.py and update the coding_agentic default in v4_inputs/tier2b_defaults.csv",2026-10-02,"Undocumented endpoint, top-20 apps only. Values seen on 2026-10-02 differ slightly from the file because the month window rolls daily." openrouter_rankings_daily_documented,OpenRouter documented daily token totals for the top 50 models (the only per-model time series),Not currently used; candidate replacement for per-model volume evidence,Documented endpoint /api/v1/datasets/rankings-daily; keyless call returns HTTP 401,OpenRouter,https://openrouter.ai/docs/api/api-reference/datasets/daily-token-totals-for-top-50-models,no,manual,quarterly,A per-model volume series becomes usable (key obtained): reconsider V_DAY and the class-specific volume sensitivity,2026-10-02,"Needs an ordinary API key; whether a zero-balance key works is undocumented (UNSURE, service_life_summary.md H.2). CC BY 4.0, cite 'Source: OpenRouter (openrouter.ai/rankings), as of {as_of}'." epoch_usage_reports,"Epoch AI usage reports: per-model token volumes disclosed by providers (GPT-5 Codex 2.0e12 tokens/day, 2025-10-06)",V_DAY = 40e12/21 in recompute_v3_lifecycle.py; N_LIFE,"51 rows, 2023-03-25 to 2026-09-29; GPT-5 Codex row 2.0e12 daily tokens (source: OpenAI DevDay 2025)",Epoch AI,https://epoch.ai/data/ai_companies_usage_reports.csv,yes,csv_value,quarterly,New per-model volume rows (especially closed frontier models) or a restated GPT-5 Codex row: reconsider V_DAY and the class-specific volume sensitivity in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; every training term moves with V_DAY,2026-10-02,"Latest date is the newest observation date in the file, not a release date. V_DAY itself uses the DevDay statement '>40 T in first 21 days' (the Epoch row is 2.0e12/day), see provider_token_disclosures." provider_token_disclosures,"Provider statements of token volume (OpenAI DevDay, Google I/O)",V_DAY in recompute_v3_lifecycle.py; paper s3.7 comparisons,OpenAI DevDay 2025-10-06 (>40 T tokens in 21 days for GPT-5-Codex; 6 bn tokens/min API); Google I/O 2026-05-19 (3.2 quadrillion tokens/month),OpenAI; Google,https://openai.com/devday/2025/ ; https://www.youtube.com/watch?v=hS1YqcewH0c&t=1940s ; https://blog.google/innovation-and-ai/sundar-pichai-io-2026/,no,manual,quarterly,Any provider discloses a per-model serving volume: reconsider V_DAY in recompute_v3_lifecycle.py; a disclosed energy per token moves the service to the hybrid method (Tier 3 only if the provider publishes a CO2e figure),2026-10-02,Epoch's CSV is the machine-readable index of these statements; read the source when Epoch adds a row. arxiv_patterson_dev,"Patterson et al. 2022: about 3/5 of ML energy for inference, 2/5 for training (development ratio low 0.67)",recompute_v3_lifecycle.py D_DEV low 0.67,"arXiv:2204.05149, version v1 (2022-04-11)",arXiv,https://arxiv.org/abs/2204.05149,yes,json_field,annual,A new version (unlikely) or a newer first-party Google split: revisit D_DEV low in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,Static paper; the check is a formality. epoch_rd_vs_training,Epoch AI: final training runs are a minority of R&D compute (development ratio evidence),recompute_v3_lifecycle.py D_DEV,"Gradient Updates, 23 March 2026",Epoch AI,https://epoch.ai/gradient-updates/r-and-d-vs-training-compute,yes,content_hash,annual,Page revised or a first-party disclosure of development against serving energy: revisit D_DEV (0.67 / 1.5 / 2.5) in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,Visible-text hash. epoch_openai_compute,Epoch AI: most of OpenAI's 2024 compute went to experiments (USD 5.0 bn R&D vs about 2.0 bn inference),recompute_v3_lifecycle.py D_DEV high 2.5,Data insight (2024 figures),Epoch AI,https://epoch.ai/data-insights/openai-compute-spend,yes,content_hash,annual,Page revised or newer-year figures: revisit D_DEV high in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,Visible-text hash. epoch_company_spending,Epoch AI company spending breakdown (development share; basis of D_DEV central 1.5 via SAIG),recompute_v3_lifecycle.py D_DEV central 1.5,"Data insight as of 2026-10-02 (also cites Dwarkesh Patel interview with Dylan Patel, https://www.dwarkesh.com/p/dylan-patel-3)",Epoch AI,https://epoch.ai/data-insights/company-spending-breakdown,yes,content_hash,annual,Page revised: revisit D_DEV central in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,UNSURE how directly the 1.5 central follows from this page; V3_LIFECYCLE_SPEC.md records it as 'SAIG / Epoch (Anthropic 2025)'. sec_10k_useful_life,Operators' server useful life (hardware life 5 years; range 4-6),recompute_v3_lifecycle.py LIFE_Y,Amazon 10-K FY2025: servers and networking equipment 6 -> 5 years from 2025-01-01,SEC EDGAR,https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/R10.htm ; https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/R31.htm ; https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/R11.htm ; https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/R10.htm,no,manual,annual,Any large operator discloses a change in estimated server life: revisit LIFE_Y (5; sensitivities 4 and 6) in recompute_v3_lifecycle.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; ,2026-10-02,"Filing URLs are per fiscal year (listed in service_life_summary.md; the first is Amazon's, the others are the operators of CIK 1652044, 789019 and 1326801); the next 10-K needs new URLs found on EDGAR." arxiv_li_water,Li et al. 'Making AI Less Thirsty': on-site WUE and EWIF for the global average and Finland rows,v4_inputs/water_onsite_selected.csv rows global-average (0.55) and europe-north1 (0.01),"arXiv:2304.03271, version v5 (2025-03-26)",arXiv,https://arxiv.org/abs/2304.03271,yes,json_field,annual,New version: recheck the 0.55 L/kWh US average and the Finland value; update the two rows; Rerun recompute_v3_lifecycle.py and build_detailed.py; water rows move,2026-10-02,The Finland value is flagged in the file as the weakest row (projected site). Paper cites the CACM version (doi 10.1145/3724499). reig_ewif,Reig et al. (WRI 2020) electricity water intensity factors (EWIF) and generation mixes,"recompute_v3_final.py REIG_F, REIG_MIX (App. 4-6, EPA 2018 mixes for eGRID subregions)","WRI Working Paper, January 2020",WRI,https://www.wri.org/research/guidance-calculating-water-use-embedded-purchased-electricity ; https://files.wri.org/d8/s3fs-public/guidance-calculating-water-use-embedded-purchased-electricity.pdf,no,manual,on release,WRI publishes revised EWIF tables or the paper moves to a newer electricity mix: update REIG_F / REIG_MIX in recompute_v3_final.py; Rerun recompute_v3_lifecycle.py and build_detailed.py; scope-2 water rows move,2026-10-02,Static 2020 publication; the web page Last-Modified changes with site updates so it is not used. aware2_zenodo,AWARE 2.0 water-scarcity characterisation factors (annual non-agricultural CF by country),v4_inputs/aware_factors.csv (read by recompute_v3_lifecycle.py at run time),"AWARE 2.0, Zenodo record 15133241, version 1.0.0, published 2025-04-03 (file AWARE20_Countries_and_Regions.xlsx, sheet CFs_nonagri)",Zenodo (WULCA),https://doi.org/10.5281/zenodo.15133241,yes,json_field,on release,"A newer version of the record exists: download it, compare the annual non-agricultural CF of each country row in aware_factors.csv, update the CSV; Rerun recompute_v3_lifecycle.py and build_detailed.py; scarcity-weighted water rows move",2026-10-02,"The script asks Zenodo for versions/latest of the record. At the first check (2026-10-02) latest was 1.0.2 (published 2026-07-30), newer than the 1.0.0 the dataset uses; refresh_state.json is pinned to 1.0.0 so this shows as CHANGED until reviewed. Record metadata also shows a modification on 2026-09-02 (informational)." exiobase_zenodo,"EXIOBASE 3 'Computer and related services' factor, Austria 2019 (0.1181 kg CO2e/EUR) for the Tier 1 spend-based line","recompute_v3_final.py EEIO, NOTION_EUR; paper Tier 1 and case study","EXIOBASE 3.8.2 (Zenodo 10.5281/zenodo.5589597); factor 0.1181 kg CO2e/EUR, 2019 basic prices, as served by Climatiq",Zenodo (EXIOBASE consortium),https://doi.org/10.5281/zenodo.5589597,yes,json_field,annual,"A newer EXIOBASE version exists or Climatiq restates the factor: look up the new Austrian (or vendor-country) factor, update EEIO in recompute_v3_final.py, confirm the deflator basis year; Rerun recompute_v3_lifecycle.py and build_detailed.py; Tier 1 and the EEIO-to-physical ratios (2.5-9.5x) move",2026-10-02,"UNSURE: 0.1181 was read from Climatiq; an earlier note recorded the record as deprecated and replaced, but on 3 October 2026 its data-versioning status was Current with no deprecation notice, so that note is withdrawn (reference [22] of the paper), not from the raw EXIOBASE tables, so a Zenodo version change is only a prompt to re-source it. At the first check the latest Zenodo version was 3.10.2 (2026-05-13); refresh_state.json is pinned to 3.8.2 so this shows as CHANGED until reviewed." eurostat_hicp,Euro-area HICP annual average index (2015=100) for deflating Tier 1 spend from the spend year to the factor's 2019 basis,"recompute_v3_final.py HICP_2019, HICP_2025","Euro area (EA) all-items: 104.80 (2019), 128.75 (2025)",Eurostat prc_hicp_aind,https://ec.europa.eu/eurostat/api/dissemination/statistics/1.0/data/prc_hicp_aind?geo=EA&coicop=CP00&unit=INX_A_AVG&format=JSON&lang=en&sinceTimePeriod=2019,yes,json_field,annual,A new year appears or 2019/2025 are restated: update HICP_2019/HICP_2025 (and the spend year) in recompute_v3_final.py; Tier 1 case-study line changes,2026-10-02,"The API URL is built from the dataset code and filters named in the paper (prc_hicp_aind, CP00, 2015=100, euro area); it is not a URL quoted in the sources. The paper's databrowser URL is https://ec.europa.eu/eurostat/databrowser/view/prc_hicp_aind/default/table. Values matched the paper on 2026-10-02." ecb_usd_eur,"ECB euro reference rate, US dollar, annual average 2025 (1.1300 USD per EUR)","recompute_v3_lifecycle.py EUR_USD = 0.885 (price conversion in the EEIO comparison, Notion seat example)",2025 annual average 1.1300 USD/EUR (series EXR.A.USD.EUR.SP00.A) = 0.885 EUR/USD,ECB,https://data-api.ecb.europa.eu/service/data/EXR/A.USD.EUR.SP00.A?format=csvdata,yes,csv_value,annual,2026 annual average published (January 2027) or 2025 restated: update EUR_USD in recompute_v3_lifecycle.py; EEIO comparison ratios and the Notion example move (linear),2026-10-02,The API URL is built from the series key the paper cites; the paper's URL is the portal https://data.ecb.europa.eu/. Value matched the paper (1.1300) on 2026-10-02. openai_gpt4o_page,OpenAI GPT-4o API list price (USD 2.50 in / 10.00 out per M tokens),recompute_v3_lifecycle.py GPT4O_USD; EEIO-to-physical comparison,"GPT-4o USD 2.50 / 10.00 per M tokens, accessed 2026-09-29",OpenAI,https://developers.openai.com/api/docs/models/gpt-4o,yes,content_hash,annual,Price change or deprecation of gpt-4o: update GPT4O_USD in recompute_v3_lifecycle.py; EEIO-to-physical ratios move; reconsider whether gpt-4o is still the reference model,2026-10-02,Visible-text hash of a 400 KB page; may change for unrelated edits. openai_pricing_page,OpenAI API pricing page (GPT-4o mini USD 0.15 in / 0.60 out),recompute_v3_lifecycle.py GPT4O_MINI_USD,GPT-4o mini USD 0.15 / 0.60 per M tokens,OpenAI,https://developers.openai.com/api/docs/pricing,no,manual,annual,Price change: update GPT4O_MINI_USD in recompute_v3_lifecycle.py,2026-10-02,"UNSURE: the paper cites one URL (the gpt-4o model page) for both models; the GPT-4o mini page URL is not documented, so the pricing page (cited in token_weighting_evidence.md) is listed for manual reading." model_class_mapping_pages,"Vendor documentation of current model names, default models and pricing tiers (the tier branch of the Section 3.5 class assignment rule)",model_class_mapping.csv (the published pattern to class mapping); main.tex Table S4; model_factors.csv class assignment,"Class by vendor tier: mini/nano/flash/lite/haiku/terra/luna = A, standard/sonnet/pro/plus = B, frontier/opus/ultra/max/sol/astra and any tier priced above the previous flagship = C",OpenAI; Anthropic,https://openai.com/index/introducing-gpt-5/ ; https://platform.claude.com/docs/en/about-claude/pricing,no,manual,quarterly,"A provider publishes a new model name, renames a tier or changes a pricing tier: re-apply the Section 3.5 rule to the affected patterns in model_class_mapping.csv and rerun tests_end_to_end.py",2026-10-02,Table S4 in the paper and model_class_mapping.csv in the dataset must be updated together. Where a provider publishes no tier definition the ordering is inferred from its published prices and the row says so. anthropic_pricing_page,Anthropic API pricing (used in tier2b price conversions),v4_inputs/tier2b_defaults.csv,Page as of 2026-10-02,Anthropic,https://platform.claude.com/docs/en/about-claude/pricing,yes,content_hash,annual,Price or tier change: update the price-based conversions in v4_inputs/tier2b_defaults.csv,2026-10-02,"Visible-text hash; changes with every new model, so expect frequent CHANGED." notion_pricing,"Notion Business plan price (USD 20 per member per month, billed annually)","Paper case study only (seat conversion to EUR 8,000); not a dataset factor",USD 20 per member per month; EUR 19.50 on the euro page; accessed 2026-09-29,Notion Labs,https://www.notion.com/pricing,yes,content_hash,annual,Price change: update the worked example in the paper (37.7 seats); no factor changes,2026-10-02,Illustrative input only. mistral_lca,"Mistral Large 2 life-cycle assessment (1.14 g CO2e per 400-token response, excludes training)",recompute_v3_final.py MISTRAL_G_400; comparison in numbers_v3_lifecycle.json,"Mistral LCA with Carbone 4 / ADEME, 22 July 2025",Mistral AI,https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai,no,manual,annual,Mistral or another provider publishes a new per-response LCA: update the comparison constant and text; a disclosed per-token energy moves the service to the hybrid method (Tier 3 only if the provider publishes a CO2e figure),2026-10-02, arxiv_bistline,"Bistline et al. corporate AI GHG framework (0.88 Wh/1K comparison, default 0.1 kg/M)",Comparison lines in numbers_v3_lifecycle.json; paper s3.6,"arXiv:2608.06733, version v1 (2026-08-07)",arXiv,https://arxiv.org/abs/2608.06733,yes,json_field,annual,New version or revised defaults: update the comparison constants in recompute_v3_lifecycle.py and the paper text; no factor changes,2026-10-02,Comparison only. arxiv_jegham,Jegham et al. 'How Hungry is AI?' (GPT-4o energy basis for the note that Class B is low-side),Paper note below Table 1; model class table remark,"arXiv:2505.09598, version v6 (2025-11-24)",arXiv,https://arxiv.org/abs/2505.09598,yes,json_field,annual,New version with changed GPT-4o energy: revisit the class assignment note for gpt-4o in the paper,2026-10-02,Comparison only. tier2b_azure_traces,"Azure LLM inference traces (2023, 2024): tokens per request",v4_inputs/tier2b_defaults.csv; v4_inputs/tier2b_evidence.md,AzureLLMInferenceDataset2023.md last commit 1a61e6c66a (2024-04-25); 2024.md last commit 829aa9b8d1 (2026-06-03),Microsoft Azure Public Dataset (GitHub),https://github.com/Azure/AzurePublicDataset/blob/master/AzureLLMInferenceDataset2024.md,yes,json_field,annual,New commit to either dataset description or a 2025 trace: rerun the token statistics and update the Azure rows of v4_inputs/tier2b_defaults.csv,2026-10-02,"Script reads the GitHub commits API for both description files (unauthenticated, 60 requests per hour per IP). The 2023 file URL is https://github.com/Azure/AzurePublicDataset/blob/master/AzureLLMInferenceDataset2023.md." tier2b_wildchat,WildChat-1M conversation dataset (tokens per chat request),"v4_inputs/tier2b_defaults.csv; v4_inputs/analyze_conversations.py, fetch_conversation_samples.py","allenai/WildChat-1M, sha 7d6490e462, last modified 2024-10-17",Hugging Face,https://huggingface.co/datasets/allenai/WildChat-1M,yes,json_field,annual,Dataset revision changes: rerun analyze_conversations.py and update the WildChat row of v4_inputs/tier2b_defaults.csv,2026-10-02,Evidence notes mention WildChat-4.8M and nontoxic variants being gated or null-usage; only the 1M dataset is tracked. tier2b_anthropic_econ_index,Anthropic Economic Index (usage composition),v4_inputs/tier2b_defaults.csv,"Anthropic/EconomicIndex, sha 4e5f69fab1, last modified 2026-09-30",Hugging Face,https://huggingface.co/datasets/Anthropic/EconomicIndex,yes,json_field,quarterly,New release of the index: re-read the task and usage shares and update the corresponding rows of v4_inputs/tier2b_defaults.csv,2026-10-02,The dataset is updated often; expect CHANGED at most checks. UNSURE which release the defaults file was built from. tier2b_uk_copilot_pilots,UK and other government Copilot pilots: queries or prompts per user per day,v4_inputs/tier2b_defaults.csv (seat-based defaults); tier2b_evidence.md,"DBT M365 Copilot evaluation; UK AI coding assistant trial; HMRC, DWP, Australian and Pennsylvania pilots",UK government and others,https://assets.publishing.service.gov.uk/media/68adbe409e1cebdd2c96a19d/dbt-microsoft-365-copilot-evaluation.pdf ; https://www.gov.uk/government/publications/ai-coding-assistant-trial/ai-coding-assistant-trial-uk-public-sector-findings-report ; https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report/microsoft-365-copilot-experiment-cross-government-findings-report-html ; https://www.gov.uk/government/publications/evaluation-report-phase-3-trial-of-microsoft-copilot/evaluating-the-impact-of-microsoft-copilot-in-hmrc ; https://www.gov.uk/government/publications/an-evaluation-of-dwps-microsoft-copilot-365-trial/an-evaluation-of-dwps-microsoft-365-copilot-trial ; https://www.digital.gov.au/initiatives/copilot-trial/microsoft-365-copilot-evaluation-report-full/employee-related-outcomes ; https://www.pa.gov/content/dam/copapwp-pagov/en/oa/documents/programs/information-technology/documents/openai-pilot-report-2025.pdf,no,manual,annual,A new published pilot with measured usage per seat: add to tier2b_evidence.md and revisit the seat-based defaults in v4_inputs/tier2b_defaults.csv,2026-10-02,"Several documents, no machine-readable feed; search gov.uk for new Copilot or AI assistant trial reports." tier2b_other_usage_evidence,"Other usage evidence for Tier 2b defaults (Claude Code cost documentation, OpenAI workspace analytics, NBER 'How People Use ChatGPT')",v4_inputs/tier2b_defaults.csv (use_case rows coding_agentic and all),"code.claude.com costs page (991,200 input + 5,300 output over 14 requests); NBER w34255",Anthropic; OpenAI; NBER,https://code.claude.com/docs/en/costs ; https://help.openai.com/en/articles/10875114-workspace-analytics-for-chatgpt-enterprise-and-edu ; https://www.nber.org/papers/w34255,no,manual,annual,A vendor documents measured tokens per request or per seat: update the matching default in v4_inputs/tier2b_defaults.csv,2026-10-02,Pages are not stable enough to hash usefully. model_class_mapping_csv,"The published commercial and open-weight model identifier to model class mapping: 187 rows, 372 pattern alternatives, of which the first 36 rows are Table S4 verbatim",model_class_mapping.csv; ai_inference_emissions.py map_model(); tests_end_to_end.py part 3c; main.tex Table S4,"v3, assigned 2026-09 (Table S4 rows) and 2026-10 (extensions)","Provider model lists and pricing pages (OpenAI, Anthropic, Google, xAI, Mistral, Meta, DeepSeek, Alibaba/Qwen, Cohere, Amazon Nova, Microsoft Phi, Z.ai, Moonshot, MiniMax, NVIDIA, Tencent, Xiaomi, AI21, Perplexity) and model cards for disclosed sizes",https://developers.openai.com/api/docs/models ; https://platform.claude.com/docs/en/about-claude/models/overview ; https://ai.google.dev/gemini-api/docs/models,no,manual,quarterly,"A new model name appears in a provider model list, in a provider pricing page or in the OpenRouter top 150 by tokens: add the pattern, or confirm that an existing prefix already covers it, then rerun tests_end_to_end.py (part 3c fails if any identifier in model_service_life.csv or in the OpenRouter top 150 falls to the Class C default)",2026-10-02,"Maintained by hand; there is no generator, because the inputs are provider web pages. Each row records its own source URL and the month it was assigned. Rows whose tier could not be determined from the provider's own documentation say so and take Class C." provider_model_lists,"The provider model-list and pricing pages themselves, which are the source of every [tier] assignment and of the disclosed parameter counts behind every [size] assignment",model_class_mapping.csv source column; the Section 3.5 assignment rule,"Fetched 2026-10-02. Note that platform.openai.com/docs/models now 301-redirects to developers.openai.com/api/docs/models, openai.com/api/pricing returns 403, and llama.com redirects to dev.meta.ai","OpenAI, Anthropic, Google, xAI, Mistral, Meta, DeepSeek, Alibaba, Cohere, AWS, Microsoft",https://developers.openai.com/api/docs/pricing ; https://platform.claude.com/docs/en/about-claude/pricing ; https://ai.google.dev/gemini-api/docs/pricing ; https://docs.x.ai/docs/models ; https://mistral.ai/pricing/api ; https://api-docs.deepseek.com/quick_start/pricing ; https://www.alibabacloud.com/help/en/model-studio/model-pricing ; https://docs.cohere.com/docs/models ; https://aws.amazon.com/bedrock/pricing/ ; https://azure.microsoft.com/en-us/products/phi,no,manual,quarterly,"A page moves, a provider retires a tier name or a provider starts or stops disclosing a parameter count: the affected model_class_mapping.csv rows change basis between [tier] and [size], which can change the class. AWS's Bedrock pricing page is JavaScript-rendered and needs a browser-capable fetch; Cohere does not publish prices for its current generation",2026-10-02,"Where a provider's tier word and its disclosed size disagree, Section 3.5 gives the disclosed size precedence. Three 2026 cases turn on this: DeepSeek V4 Flash, GLM-5.3-Flash and the MiMo V2.x Flash variants are all 'flash' tiers of 284B to 552B total parameters and so Class C, not Class A." dcm_site_locations,Physical site locations of hyperscale cloud regions where the provider states only a state or country (secondary evidence for the eGRID subregion mapping),cloud_regions.csv (notes and confidence of the US rows mapped from an assumed site),"Microsoft (355 listings) and Amazon AWS (570 listings) operator pages, read 2026-10-14",Data Center Map,https://www.datacentermap.com/c/microsoft/ and https://www.datacentermap.com/c/amazon-aws/,no,manual,semi-annual,"A new US region announced by state only, or a listed campus moving to another eGRID subregion: rerun the point-in-polygon test (EPA eGRID subregion shapefile) and update the note and confidence",2026-10-14,"A voluntary, secondary directory (14,501 listings from 3,614 operators; the site states that outdated records exist), used only to locate a site, never for energy, PUE or water. Provider documentation remains the primary source; where the two disagree the provider wins and the row stays medium." subnational_grids,Generation shares behind the subnational grid rows (Quebec for AWS ca-central-1),v4_inputs/subnational_grids.csv; grid_intensity_all.csv (kind subnational); cloud_region_factors.csv (ca-central-1),"Quebec 2023: 195.5 TWh, hydro 94.0%, wind 5.0%, biomass 0.7% (CER page updated 2026-03-25)","Canada Energy Regulator (Statistics Canada tables 25-10-0020-01, 25-10-0084-01, 25-10-0028-01)",https://www.cer-rec.gc.ca/en/data-analysis/energy-markets/renewable-energy-canada/provinces/renewable-power-canada-quebec.html,no,manual,annual,"New data year on the CER page or in the Statistics Canada tables, or a new Ember release (the per-fuel factors are re-derived from Ember's Canada rows at build time)",2026-10-14,The intensity is not read from a source; build_detailed.py derives it from these shares and Ember's implied Canadian per-fuel factors so the row shares the country rows' life-cycle basis.