State AI Policy Tracker

A traceable collection of reported government policies across thirteen Indian states.
Research window: 1 January 2022 to 10 November 2026.

I built this around a simple question: when five headlines describe an AI announcement, what should count as one record? The tracker keeps the reports behind each record, explains the rules used to group them, and records what was dropped. It contains 716 records, 1,306 contributing reports and 17,253 excluded items with reasons.

Figure 1. The live dashboard [1]
Figure 1. The live dashboard [1]

An announcement, a signed agreement and a working public service need different evidence. The tracker labels the reported stage; it does not verify delivery, spending or benefit. Seven central-government actions, 22 judiciary actions and seven local actions are labelled separately from 680 state-government records.

Dashboard | Trend charts | GitHub repository | Hugging Face dataset | Dataset CSV | Submission folder

What the collection shows

The same initial query was run across all 28 states. The selection used that discovery volume and geographical spread. Andhra Pradesh, Madhya Pradesh and Rajasthan were added for comparison with a separate document-led reference collection. These thirteen states are a deliberate sample, not a representative sample or a league table.

Figure 2. Records by state and report year [2]
Figure 2. Records by state and report year [2]

The useful findings

Karnataka has 99 retained records, Kerala 95 and Telangana 86. Those counts tell me where this collection found more reportable material. They also depend on English-language reporting, search indexing and the rules used to retain a headline. A small count cannot establish that a state did little.

The annual count rises from 23 in 2022 to 389 in 2026 through 10 November. Older search coverage and different screening coverage make a policy-growth estimate unsupported. The year chart is useful for finding periods to inspect, not for measuring a rate of adoption.

QuestionWhat the data supportsWhat it does not establish
Where was activity reported?State and report-date countsA complete census of state action
What stage was described?Headline-based stage labelsWhether a programme was delivered
How often was it reported?Contributing reports and outlet labelsIndependent corroboration by each outlet

How a report becomes a record

Google News RSS is the main discovery source because it can be queried without a paid key. Four query shapes cover government actors, policy instruments, public services, and infrastructure or rules. Each runs by state and month. Overlapping results are expected; the pipeline handles them explicitly. [3]

Figure 3. Collection and processing pipeline [3]
Figure 3. Collection and processing pipeline [3]
Source or stepSaved releaseWhat to account for
Google News RSS3,068 query responses; 36,855 hitsNo response reached the 100-item warning threshold. This is not proof of complete coverage.
GDELT and Bing RSS896 and 172 hits respectivelyGDELT was substantially incomplete. Across sources, 391 attempts failed or were skipped.
Page extractionLocal trafilatura; cached redirect resolutionTry up to three reports per cluster. Blocked pages can leave only a description or headline.
Keyed providersGNews, NewsAPI and NewsData adaptersNot live-tested in this release. Tavily fallback was unused; Exa, Firecrawl and Crustdata are not integrated.

After repeated discovery items are collapsed, the build has 18,559 items. Relevance rules retain 1,662, clustering forms 967 groups, and saved exclusions leave 716 records. These are different processing stages: 36,855 Google hits are not 36,855 distinct articles. The offline replay reads caches and makes no page or optional model calls.

Keep the original meaning of names and dates

A search for Uttar Pradesh can return a story about another state. The query therefore never supplies the state label. The headline is checked first, then the description when it names no state. The rules favour a full state name over a city and require context for ambiguous abbreviations. [4]

Input or contextStored resultRule
Uttar-Pradesh; UttarPradesh; U.P.Uttar PradeshAliases, spacing and whole-word matching
Karnatka; Maharastra; OrissaKarnataka; Maharashtra; OdishaExplicit spelling and historical-name aliases
MP govt; Congress MPMadhya Pradesh; no MP state matchAmbiguous abbreviations need government context
Bengaluru with no state namedKarnatakaCity evidence has lower weight
Tamil Naddu in a state-label cellTamil NaduEdit similarity >= 0.84 for labels of at least six letters; not a general headline correction

Report date and action date are separate facts

The published date is the earliest report timestamp in a cluster, converted to India time. All 716 released records have day precision because the feeds supplied timestamps. That does not verify the day a government order was issued.

InputStored date and precisionReason or limit
2026-07-07T20:30:00Z2026-07-08 / dayConvert to IST before taking the date
03/04/20262026-04-03 / dayDay-first assumption; ambiguous input
Sep 2026; Q3 20262026-09-01 / month; 2026-07-01 / quarterFirst day is a storage anchor, not an exact date
FY 2025-262025-04-01 / yearFinancial-year anchor
recently; invalid date; no usable dateEmpty / unknownReject as no_date if feed and URL provide no date

The general parser handles month spelling errors and relative expressions, but relative hours can fail across midnight and week phrases are approximate. Those paths were not used for the published feed timestamps. Missing dates are never replaced with the collection date.

Match reports without losing a later reversal

The matcher compares reports attributed to the same state and at most five days apart. It removes common words, applies light stemming and a small synonym map, and gives rare shared terms more weight. Repeated words in one headline count once. A cancellation headline cannot merge with a non-cancellation headline. [5]

Figure 4. Duplicate matching score [5]
Figure 4. Duplicate matching score [5]

A and B are sets of normalised headline tokens. N is the number of compared headlines; df(t) counts headlines containing t; delta d is the date gap in whole days. Empty token sets score zero.

The first term measures shared weighted content relative to both headlines. The second measures how much of the shorter weighted set is shared. Averaging them lets a short headline match a longer version without relying on a common state name or the word AI.

Date gap in days0 or 12345
Required score0.250.330.410.490.57

A small worked example

For illustration, let W(A) = 9, W(B) = 16 and W(A intersect B) = 8. Then S = 0.5 x 8/12 + 0.5 x 8/9 = 0.778. A same-state, same-reversal-status pair three days apart clears 0.41. A reversal mismatch still prevents the link. These are illustrative weights, not a measured article pair.

Linked pairs form connected groups using union-find. Identical URLs can also join reports beyond five days, but only if state and reversal status agree. The representative headline is chosen for similarity to the rest, with outlet and time tie-breakers. The earliest report supplies the date and ID seed. Every contributing report stays in record_sources.csv.

This design can still chain related actions into one group, leave late reports as duplicates, or lose detail when decimal amounts are truncated. A newly found earlier report can change an ID. Similarity is a grouping rule, not proof of event identity.

Show how much evidence each record has

Figure 5. Summary basis and headline stage [2]
Figure 5. Summary basis and headline stage [2]

Most summaries use publisher descriptions: 629 of 716. Only 47 use extracted article text; 40 repeat the headline. All representative links are direct publisher links, but 1,206 of the 1,306 supporting links still use Google redirects. A report from another outlet may also be syndicated, so it is not automatically independent evidence.

CheckObserved resultBoundary
Release and regression tests97 tests passed in the release checkout and extracted code ZIPStructural consistency, not verification of article claims
Review history545 legacy_headline_screen; 171 autoHistorical screening and pair labels were assistant-produced, not independent human article review
Duplicate-pair diagnostic136 of 140 labelled pairs found; reporting precision 0.96, recall 1.00Development diagnostic only: 31 headlines overlap halves; changes used errors across the set
Offline replayPublished CSV outputs reproduced in the release checkXLSX and SQLite binary bytes can differ; compare rows and values

Precision = TP / (TP + FP); recall = TP / (TP + FN). Here, a positive prediction means two labelled headlines end up in the same cluster. These formulas do not fix the evaluation design: shared headlines and assistant labels prevent an independent accuracy claim. [6]

The broad Government deployment category contains 317 records because it is also the fallback category. It needs further review. I would not report those rows as 317 operating services.

An agreement and its cancellation are different events

The final check found two defects: the reversal regex contained control characters instead of word boundaries, and the shared-URL shortcut bypassed state and reversal checks. Repairing both separated two events that had been merged. The total changed from 714 to 716; all original IDs were retained. [7]

CaseRetained report dateWhy it matters
Uttar Pradesh Puch AI cancellation27 March 2026The report describes cancellation on 26 March after the 23 March agreement. The agreement cannot stand for both actions.
Safe Kerala AI-camera payment halt21 June 2023The report describes the preceding day's court order. A payment halt belongs in reversed, not announced.
Figure 6. Evidence to collect in follow-up research
Figure 6. Evidence to collect in follow-up research

The separate reference appendix contains 85 events, 113 sources and 137 event-source links. It uses different inclusion and date rules and can overlap the tracker. I have kept it separate rather than presenting an inflated combined count.

Reference leadWhat it establishesWhat to find next
Rajasthan GPU-server RFP dated 24 March 2026Procurement requirements, not installed capacityContract award, acceptance and operational records
Odisha policy approval and Gazette publicationDifferent formal milestones in the inherited notesReopen the Gazette and link the notified text
Tamil Nadu AI mission government orderA formal-instrument lead beyond a general headlineReopen the order, then track funded activities

Rajasthan's RFP was fetched in the release review. The Odisha and Tamil Nadu links timed out during that check; their inherited notes are not newly verified findings. The source links remain in the appendix so a reviewer can follow them. [8]

Use the charts to find records worth checking

I kept the existing dashboard because the reader needs a short route from a pattern to the reports behind it. State, category, sector and government-level filters narrow the list; the monthly heat map selects a period; source links open the evidence. The separate trends page makes the time series easier to inspect.

Figure 7. Records by half-year from the trends page [1]
Figure 7. Records by half-year from the trends page [1]

Rebuild the saved release

Use Python 3.11 or later and install requirements.txt in a virtual environment. The cached release needs no paid API key. These are PowerShell commands from the repository README. [3]

python -m pip install -r requirements.txt
python -m pytest -q
$env:PIPELINE_OFFLINE = "1"
python -m pipeline.collect
python -m pipeline.build
python -m scripts.validate_release
python -m http.server 8790 --directory docs

The weekly workflow collects recent candidates into a review artifact. It does not add them to the published dataset. SQLite is an export of the current tables, not a persistent observation history. The saved release and a fresh live collection are different runs and should not be expected to have identical coverage.

Files to inspect and claims to check

DeliverableLocation in the submission folder
Answer document01_Submission: native Google Doc, editable Word and PDF
Dataset02_Dataset: records.csv, records.xlsx, record_sources.csv, rejected.csv and tracker.sqlite
Runnable source and history03_Code: code ZIP; 04_Git_History: Git bundle
Reasoning and method05_Method_and_Research: one-page rationale, README, methodology and research appendix
Checks and visuals06_Verification_and_Figures: validation outputs and chart assets

What I would improve next

First, independently review a fixed set of whole events, with no event appearing in both tuning and evaluation. Then add official orders, tenders and budgets, keeping report dates separate from action dates. Resolve supporting redirects, record later changes to articles, and test regional-language discovery. Additional providers should be judged by useful new records and failure rates on the same query set.

Source index

[1] Live dashboard and trend charts. Screenshots captured 11 November 2026; published counts cross-checked against the local release.

[2] Released records and source table. 716 records; 1,306 contributing reports. Figures are computed from this release.

[3] Collection and reproduction. Source status, cached-run counts and build commands.

[4] State and date rules. Input examples, precedence, precision and known parser limits.

[5] Duplicate matching implementation. Tokenisation, squared rarity weights, similarity, guards and connected groups.

[6] Development diagnostic. Overlapping headlines and development-only precision and recall.

[7] Corrective release notes. Two separated reversals, preserved IDs and regression checks.

[8] Document-led research appendix. Official-document links, access limits and follow-up evidence.