IGRMPublic research instrument
Open data

Method

Methodology

Data day
See latest published payload
Freshness
Source-by-source status

Version 1.11.0 · five channels frozen 2026-07-24, V2 layers added 2026-07-31, comparators and predictability 2026-08-01, receipts drill-down 2026-08-02, API contract frozen 2026-08-04, corpus receipts and sampling bands 2026-08-06, 7-day headline 2026-08-06, raw shares and the independent splice audit published 2026-08-07, construct naming locked and AI-GPR benchmark published 2026-08-07 · changelog at the end.

1. What the index measures (and what it does not)

The index measures press salience: for each of five channels, the share of all articles GDELT monitors globally that match the channel's term set, expressed as a percentile against that channel's own trailing 730 days. A composite score of 79 means exactly this: today's matching-coverage share, averaged across the five channels' percentiles, is higher than it has been on 79% of days over roughly the past two years. Nothing more.

IGRM keeps its established name, but the construct is geopolitical salience, not the probability or severity of geopolitical events. It is distinct from the Caldara-Iacoviello GPR family. The standing comparison uses their monthly India series, GPRC_IND, over the 110 months shared by the two instruments. A separate, code-frozen benchmark against the newer AI-GPR India_all series finds Spearman rho 0.256 in 106 month-over-month changes, with a registered six-month moving-block 95% interval of [0.050, 0.407]. The registered interpretation is that the measures share a common component but are far from redundant. This does not establish that either is more accurate. The two constructs can diverge for ordinary reasons: an under-covered crisis scores low in IGRM, an anniversary retrospective scores high (§7), and editorial fashion can move the series with no change in the world. The index answers "how much is the press writing about this?", never "how dangerous is this?". It is not a forecast and not investment advice. See the standing comparison.

2. Term selection and the ex-ante rule

Each channel's dictionary draws from four categories, geography where tension physically manifests, institutions and standing mechanisms, recurring doctrine/press vocabulary, and structural chokepoints. Geography alone under-detects diplomatic coverage; doctrine vocabulary typically carries the most recall. Every term carries a one-line rationale inside dictionaries.json itself; that file, not this page, is the term-level record. Channels hold 10-14 terms: below ~8 the series is thin and noisy, above ~15 marginal terms import contamination faster than signal.

The ex-ante rule. No term may be a retrospective event name (Galwan, Pulwama, Balakot, Sindoor, Kargil, Uri, 26/11, Doklam, …). The reason is circularity: an index whose queries contain event names by construction spikes at those events, and its "validation" against them would be meaningless. Spikes must be detected by structural vocabulary that existed before, and will exist after, any particular event. The rule is enforced in CI by tests/test_dictionaries.py, which fails the build on a banned name. One borderline call is documented: "surgical strikes" entered common Indian-press usage after 2016 but is doctrine vocabulary reused across subsequent escalations, not the name of an event; it is admitted on that basis.

Query grammar. GDELT's DOC API permits OR only inside a single un-nested parenthetical that may not mix AND. Every term is therefore one quoted phrase, and disambiguation is done by a single channel-level anchor word, e.g. India ("Line of Control" OR "ceasefire violations" OR …). Quoted phrases match exact token sequences (hyphens tokenize to spaces; no stemming, hence singular and plural forms where both are common in copy). Two generic phrases are accepted with open eyes and disclosed here: "energy security" (broader than its channel; kept for recall of policy coverage) and "Suez Canal" (includes routine transit coverage; the percentile normalization absorbs its baseline).

Cross-channel bleed, decided and documented: "Russian oil" (India) sits in US & Trade Policy, not Gulf & Energy, the risk vector is sanctions policy, not physical supply. Red Sea tanker coverage belongs to Shipping; Persian Gulf tanker incidents to Gulf & Energy. No term appears in two channels, so the composite never double-counts an article across the declared boundaries.

Excluded by choice: leadership rhetoric. Statements by military and political leadership are not construct vocabulary: adding them would make rhetoric frequency part of the measure, and prolific official communication would move a channel even when its underlying geography is quiet. Such stories count only when they co-occur with structural vocabulary; pure rhetoric does not count, and that is a disclosed design decision (2026-08-04), not an oversight.

Freeze. Dictionaries froze 2026-07-24 (_meta.frozen_on). Any later change appears in the changelog with a date and a reason. An index whose definition moves silently is not reproducible.

3. Normalization

Each channel's raw series is GDELT's timelinevol measure: matching articles as a share of all monitored articles that day, which already nets out GDELT's secular corpus growth. Where a channel's term set exceeds the API's query-length limit (measured ~250 characters), the channel series is the sum of two sub-query shares; an article matching both sub-queries counts twice, making the series a slight upper bound on the union share. The partition is fixed and versioned with the dictionaries. From July 2026, during a DOC-API disruption, the recent tail is computed from GDELT's raw Web NGrams files at the maintainer's direction, the same share construct from half-hourly samples of the raw feed, and ratio-spliced to the API series on overlap days, with each channel's splice ratio and its dispersion published alongside the data (see changelog).

3.1 Independent audit of the splice calibration

The production ratios remain frozen, but they are not described as validated. A proposed 38-day stability check was rejected before publication because it was circular: on 37 of the 38 cached days, the store value used as the denominator had itself been produced by dividing the cached NGrams value by the production ratio. Re-estimating from that pair mechanically recovers the old constant.

The table below uses a genuinely independent denominator instead: the last pre-bridge DOC-API store preserved in git commit 091c25e, paired with the retained NGrams day caches. The recovered window is 2026-06-30 to 2026-07-21. It supplies 18 independent days for the first three channels, but only the original single day for US Trade and Shipping.

Channel Frozen ratio (original n) Independent audit ratio (n) Change Audit log-SD
Pakistan / Western Border 1.9547 (5) 2.4634 (18) +26.0% 0.3296
China / Eastern Border 3.3612 (5) 2.6998 (18) -19.7% 0.3954
Gulf & Energy 1.7910 (5) 1.8098 (18) +1.0% 0.0976
US & Trade 2.5616 (1) not independently re-estimable (1) not estimated 0.0000
Shipping & Chokepoints 2.9747 (1) not independently re-estimable (1) not estimated 0.0000

Bridge shares equal the NGrams share divided by the ratio. Relative to the independent audit, the frozen Pakistan ratio is smaller, so primary post-bridge Pakistan shares are higher; the frozen China ratio is larger, so primary post-bridge China shares are lower. Under the audited alternative, the latest Pakistan score therefore moves down 8.0 points and China moves up 5.1. That is the direction of this sensitivity, not a claim that either independent ratio is known without uncertainty.

This is a sensitivity finding, not a replacement calibration. It shows material uncertainty in the Pakistan and China links and leaves the two one-day links untested. Production values and published history remain unchanged to preserve the v1 vintage. The score-level effect is material, so it is published rather than summarized away:

Series Channel Median absolute shift Maximum absolute shift Shift on 2026-08-06
Daily Pakistan / Western Border 6.6 10.9 -8.0
Daily China / Eastern Border 13.0 18.7 +5.1
Daily Gulf & Energy 0.0 0.4 0.0
Daily Composite 1.3 3.0 -0.6
Trailing 7-day Pakistan / Western Border 9.4 11.4 -4.5
Trailing 7-day China / Eastern Border 21.7 25.6 +23.5
Trailing 7-day Gulf & Energy 0.0 0.4 0.0
Trailing 7-day Composite 2.6 4.1 +3.8

The table is the fixed 37-day study window from 2026-07-01 to 2026-08-06. The machine-readable artifact was extended through 2026-08-09 (40 days) before the retained-identity rights boundary was hardened, and is now frozen at docs/data/splice_sensitivity.json. It will not be recomputed unless a current signed source decision permits that evidence use. Neither artifact is a corrected history or may be substituted silently for the primary.

The published sampling-band artifact is likewise a fixed historical window: docs/data/uncertainty.json covers 2026-06-30 through 2026-08-07. It remains available for those dated scores but does not claim coverage of later days and will not be recomputed from retained identity-bearing caches without a current signed decision covering the complete study window.

A future measurement version may adopt new ratios only after at least 14 independently observed overlap days exist for every channel, the score-level impact is published, and the old vintage remains downloadable. Code and the recovered API snapshot are in analysis/splice_overlap_audit.py and analysis/splice_overlap_api_091c25e.csv.

The published score is the percentile rank of today's share within the channel's trailing 730 days (inclusive of today; the window never contains future data). Days with no observed value stay missing rather than scoring zero.

Why percentile rather than z-score: news-volume shares are fat-tailed and drift with editorial fashion. A z-score inherits both problems, single extreme days distort the mean and variance for months. The percentile is robust to outliers, invariant to monotone changes in the level of coverage, and yields a directly interpretable sentence ("higher than X% of the last two years"). Its cost is saturation, and the cost is large enough to state with a number rather than a phrase: between 2025-04-23 and 2025-05-14 the pakistan_west percentile moved 2.2 points (97.8-100.0) while the underlying share moved 5.5-fold (0.214% to 1.173%). During a sustained crisis the percentile therefore carries almost no intensity information — it cannot distinguish the peak day from the twentieth day — and the share carries all of it. Two consequences follow and are honoured elsewhere in this document: episode detection runs on raw shares, not scores (§5), and the raw shares are published as a first-class artifact (docs/data/shares.csv) so any reader can apply a non-saturating transform of their own.

Why 730 days: long enough to span more than one editorial cycle and both halves of a typical escalation-and-decay arc; short enough that "the last two years" remains a claim about the current coverage regime rather than a different era of the corpus. §8's stability check (365- and 1095-day recomputations) tests that nothing below hangs on this choice.

Minimum observations: no score is emitted until a channel has 180 trailing observations; early-window days are null rather than percentiles against a thin baseline.

4. The composite convention

The headline composite is the unweighted mean of the five channel percentiles. This is a transparency convention, not a claim that the five channels matter equally to India, no defensible weighting exists (trade-weighted? casualty-weighted? by what?), and any chosen weighting would smuggle in an editorial judgment the data cannot support. The components are the primary product; the composite exists so the site has one number to anchor the day. Readers who dislike the convention can recompute any weighting from the published per-channel series in docs/data/history.json.

5. Spikes and episodes

A spike day for a channel is a day whose raw volume share exceeds the trailing 90-day mean plus two standard deviations, with the baseline lagged one day so that a spike cannot inflate the threshold that must catch it. Detection runs on raw shares, not percentile scores, because a bounded series compresses at 100 and can make a 2σ exceedance arithmetically unreachable exactly when coverage is most extreme.

Spike days separated by three or fewer calendar days cluster into one episode (start, end, peak). Episodes rather than raw spike days are the unit of analysis because multi-day coverage waves are one event journalistically, and treating each day as independent would let long episodes dominate every downstream average. The 2σ/90-day/3-day parameters are conventions; §8 reports a 1.5σ secondary specification so readers can see the findings are not threshold-dependent.

6. Event-study design

The event study reports India-specific relative returns around episode starts, never outright returns:

  • Nifty 50 − MSCI EM, strips global equity beta; what remains is the India-specific equity move.
  • Defence basket − Nifty, the sharpest India-specific hypothesis: if border salience means anything to markets, it should appear in defence names relative to the broad index. (Basket: HAL, BEL, BDL, Mazagon Dock, Cochin Shipyard; equal-weight, daily-rebalanced.)
  • Energy OMC, IT services and ports/logistics baskets − Nifty, the three registered sector-transmission hypotheses, each using its frozen equal-weight basket.
  • USDINR − DXY, strips broad dollar moves from the rupee.

Brent and gold are configured descriptive-only outcomes: no India-specific component of a globally-priced commodity is separable, so they carry no interpretation beyond context. A configured outcome appears only when its current source cache contains observations; the JSON explicitly lists available and unavailable outcomes. Windows are 1, 5, and 20 trading days, inclusive of the first trading day on or after the episode start. Every estimate carries a bootstrapped 95% interval (1,000 resamples over episodes). The language rule is absolute: episode starts are associated with subsequent relative returns. Coverage and prices respond to the same underlying events; nothing in this design can separate the two, so "caused" and "predicts" never appear (§7).

7. Known limitations

Named here before a reader must raise them. Each with its mitigation and its residual.

  1. Salience ≠ risk. The permanent one. Mitigation: this page, §1, and a definition line under the headline number. Residual: total, the index never becomes a risk measure; it measures attention.
  2. No causal identification. No natural experiment or instrument exists in this design. Mitigation: association-only language, enforced by review. Residual: total, at any level of statistical sophistication.
  3. Thin sample. India has had tens of geopolitical episodes since 2022, not thousands. Mitigation: bootstrapped intervals reported everywhere; the series and event backfill now extend to 2017. Residual: intervals stay wide forever; findings stay descriptive.
  4. Single-source dependency. The score rests on GDELT's corpus and its English-language, Western-outlet skew. Mitigation: Wikipedia-pageview cross-validation now publishes beside it but never enters the score. Residual: agreement between two biased attention measures is not unbiasedness.
  5. Hindsight in dictionary construction. The dictionaries were written in 2026 by people who know 2022-26 history. Mitigation: the ex-ante structural-terms rule bounds the leak, no event names, only vocabulary that predates and outlives specific events, and the robustness harness (§8) shows results survive reasonable re-wordings. Residual: bounded, not eliminated; disclosed.
  6. Anniversary and editorial-cycle effects. Retrospectives count as salience by construction. Arguably a feature (attention is attention); either way, planned work quantifies it with day-of-year effects.
  7. Coverage-drift. GDELT's source list itself evolves; a step-change in monitored outlets can move shares with no change in the world. Partially absorbed by the share denominator and the trailing percentile; residual disclosed.
  8. Timezone convention. GDELT days are UTC; Indian market days are IST; the daily run, final by 6:30 AM IST, treats "today" as the UTC date. A same-day Indian-evening event lands on the correct UTC day but after the NSE close, event-study windows therefore start at the first trading day on or after the episode start, never before.
  9. Composite arbitrariness. §4. Mitigation: labelled a convention; components published. Residual: no weighting is privileged.
  10. Phrase brittleness. Exact-phrase matching misses paraphrase ("infiltration attempt" vs "infiltration bid") and non-English coverage entirely. Mitigation: doctrine terms chosen from wire-service vocabulary; robustness harness. Residual: recall is partial and skewed toward English-language convention.

8. Validation

Where credibility lives. Four checks, all runnable from the repo.

8a. Pre-registered historical detection (python -m src.validate hit-rate). Twenty-one episodes across the five channels, 2017-2025, were frozen in validation/validation_episodes.json before the first validation run (thirteen at the 2026-07-24 freeze; eight pre-2022 episodes appended, dated, under the file's append-only rule when the backfill extended to 2017, still before any validation ran). None of their names appears in any query term, that is the ex-ante rule doing its work. A hit is a detected episode in the same channel within ±3 days. The per-channel hit table is published to docs/data/validation.json and is this project's key figure: 18 of 21 detected (86%) on first run. A second tranche of 8 episodes (2018–2022, concentrated in the thin channels) was drafted blind from external chronology, registered with ex-ante criteria and recorded exclusions (validation/validation_episodes_tranche2_SIGNED.json), founder-signed on 2026-08-06, and graded only after that signature: 6 of 8 detected, bringing the combined record to 24 of 29 (83%). The two tranche-2 misses (the 2022 BrahMos accidental launch, the 2022 OPEC+ production cut) are listed in the payload like every miss — the sequence draft-blind, sign, then grade is the figure's warrant.

83% is not the number to be impressed by, and this section will not present it alone. The same payload (docs/data/detection_baselines.json) publishes what 24 of 29 should be read against: 6.8 hits expected by chance, 19 of 29 under the stricter start-based timing rule, and 26 of 29 for a naive detector that asks only whether any channel moved, ignoring which one. The naive rule beats the registered one by two events. So detection alone is cheap, and the honest reading of this section is not "the index detects events" — it is that the index detects them in the right channel, which is the only part a naive rule cannot do. Anywhere the 24 of 29 figure is quoted without those baselines, the quote is flattering the index.

8b. Dictionary robustness (python -m src.validate robustness). Broader and narrower constructions of every channel are frozen in dictionaries_alt.json. The full index is recomputed under each and correlated with the primary. Correlations above 0.9 mean "why these terms?" has no purchase; anything lower is reported as term-dependence, prominently. The robustness and placebo fetch windows cover 2022 onward (request budget), narrower than the 2017 primary series; each published table states the window it covers.

8c. Placebo channels (python -m src.validate placebo). Two channels with no India-geopolitics content (IPL cricket, Bollywood) run through the identical pipeline. They must not spike around geopolitical episodes; their overlap fraction is published. One disclosed imperfection: Indian sport is not perfectly insulated from geopolitics (India-Pakistan fixtures), so the cricket terms avoid Pakistan-linked phrasing.

8d. Normalization-window stability. The index recomputed at 365- and 1095-day windows must be qualitatively unchanged (rank correlation with the primary reported alongside 8b's table). Episode detection is additionally reported at 1.5σ beside the primary 2σ.

A note on what validation cannot do: passing 8a-8d shows the instrument detects what it claims to detect and is not an artifact of one term list. It does not, and cannot, convert salience into risk (§7.1).

9. The India Stress Gauge

Experimental secondary object; not a validated headline measure. As of 7 August 2026 the registered detection rule succeeds on 2 of 29 episodes. That negative result is published rather than tuned away. Each release also lists the latest available and missing components and the effective weights after any registered renormalization.

The gauge fuses four sources into one daily 0-100 line: the composite press-salience percentile, a conflict-event intensity percentile from the GDELT Events stream ((verbal + material conflict events) / global events), a market-stress percentile (the mean of the India VIX level percentile and the USDINR 10-day realized-volatility percentile), and a Wikipedia attention percentile. Each component is ranked against its own trailing 730 days with the same 180-observation minimum as the index; the gauge is their weighted mean.

These weights belong to the stress gauge, not to the index. §4 refuses any weighting of the five press channels because none is defensible; the gauge is a different registered object combining four distinct measurement modalities, where the components are not interchangeable and a stated weighting is unavoidable. A reader who finds §4 and §9 contradictory has found an ambiguity in this document rather than in the construction: the index composite is unweighted and always has been. The gauge's weights (press 0.35, events 0.25, market 0.25, wikipedia 0.15), the detection rule (gauge at 90 or higher within 3 days of an episode date), and the missing-component rule (press required, at least two of the other three present, weights renormalized) were registered in validation/stress_gauge_weights.json with per-component rationale and committed before any hit-rate was computed; the repository history is the proof of ordering. The hit-rate against the pre-registered episode list publishes with the gauge in docs/data/stress_gauge.json, along with the per-component percentiles behind each day's number. Whatever the hit-rate is, it is reported as found; the gauge measures attention and stress, and predicts nothing.

10. Comparator countries and predictability

Three comparator series (Pakistan, Indonesia, Vietnam) plus India run through one deliberately simple instrument: a single shared geopolitical-risk vocabulary of ten structural phrases, anchored per country, registered with per-term rationale in comparators.json before the first fetch. Cross-country lines are comparable by construction because the instrument is identical; levels still reflect Anglophone press attention, disclosed. The five-channel index remains the primary product; the comparators exist for context and for the cross-country questions V8 will formalize.

The predictability study (docs/data/predictability.json) asks the directed lead-lag question on daily changes: five own-lags with and without five lags of the candidate leader, R-squared increment, and a permutation p-value from time-shifted nulls. In the published run, no salience-leading direction clears the conventional 0.05 threshold; events-leading-salience is the closest of the registered directions. Exact values live in docs/data/predictability.json rather than this prose so a registered rerun cannot leave a stale number here. This negative result is consistent with the instrument's stated scope: a salience monitor, not a risk predictor.

11. Receipts and source tiers

docs/data/receipts.json publishes a bounded latest-day evidence display, not a census and not the complete evidence behind the 7-day headline. The primary path reuses the sampled Web NGrams corpus and may add separately labelled, query-matched GDELT DOC supplemental URLs. It deduplicates URLs, then publishes one representative for each case-folded 120-character (title or url) key, up to 150 representatives per channel. If the corpus path is unavailable, the explicitly labelled artlist-only fallback publishes at most 75. On the primary path, every row states its lane; corpus rows distinguish the scoring sample from the extended scan.

Each displayed domain is looked up in source_tiers.json. Registered tiers affect presentation order and descriptive source-mix fields only; unregistered domains remain unranked. The legacy spike_quality_tier12_share field is the tier-1/2 share of the displayed, tier-sorted list and must not be read as the source mix of the underlying retrieval pool. Most importantly, source tiers never enter a channel score or the composite.

12. API contract

docs/data/api_contract.json freezes, as of 2026-08-07, every endpoint IGRM serves for machine consumption: the file's format, a plain-language description, and its top-level "frozen fields" (JSON keys or CSV columns) promised not to be removed, renamed, or repurposed within major version 2. New fields may be added to any payload at any time without notice; only a removal, rename, or type change requires a major version bump, which would be announced in the contract's deprecated list and in this changelog before it ships. The contract versions independently of this document (which tracks construct changes) and of the igrm Python package (which tracks code) -- docs/api.html renders it for human readers. The contract file itself is committed and hand-frozen, not regenerated by the daily pipeline: a promise that rewrote itself every night would not be a promise.

13. Historical Intelligence (the 1979-2019 archive)

The back-extension is a separate instrument, and this section exists to keep it separate. It counts actor-pair event mentions in the GDELT Events archive; the live index measures press salience over a matched-article corpus. The two are never spliced, never plotted on one axis, and no statistic computed on one is comparable to a statistic computed on the other. Where the archive is read at all, it is read as a series in its own right.

Three readings are registered in governance/historical_intelligence_contract.json and published in docs/data/historical_intelligence.json (rendered at docs/history-lab.html, defined field by field in the codebook): regime baselines over four calendar decades, structural break scans, and analog retrieval. Only the two channels the source payload's overlap_audit marks tracks are eligible; the refusals are published with their correlations rather than dropped.

Four rules bind all three, and they are the reason this section is short rather than a set of findings:

  1. Every statistic reports its own denominator, and a statistic over fewer than the registered minimum is refused with a reason rather than computed over a thin window. No unavailable value is ever rendered as a dash, a zero, or an imputed mean; the interface prints "unavailable" and why.
  2. A break is a candidate break in the measured series. It is never described as a historical cause, a turning point, a regime change, or an explanation of any event, and its stability across minimum-segment settings is published beside it. Where the candidate moves with the setting, the instability is the result.
  3. An analog is a similarity lookup, not a precedent. It carries no claim that similar months share causes and no claim about what follows them. Features missing on either side of a pair are named and excluded, never filled.
  4. No present-day knowledge is backfilled. The archive ends 2019-12, and any label, annotation or category whose evidence postdates that is refused and the refusal published. Archetypes exist only where a human-authored registered anchor does; nothing here is machine-labelled.

The decade labels are calendar slices, not political regimes, and the word "regime" in regime_baselines should be read that way and no other. Nothing in this section is a forecast.

Changelog

  • 2026-08-07, v1.11.0 (AI-GPR benchmark; Divergence Register). Before computing any joint statistic, the exact Iacoviello-Tong country-monthly file, IGRM history vintage, event list, sample rule, transformations, primary statistic, block bootstrap, decision text and analysis-script hash were frozen in public commit 58ca6c0. The primary result across 106 consecutive-month changes is Spearman rho 0.256, with a six-month moving-block 95% interval of [0.050, 0.407] and a twelve-month robustness interval of [0.082, 0.364]. Under the frozen rule, the two measures share a common component but are far from redundant. Two initial invocations failed before producing output because SciPy was absent; the dependency was installed and declared without changing the script or any registered choice. The failures are in the public run log. A new append-only Divergence Register publishes the five largest benchmark rank gaps and two salience-versus-physical-flow gaps with receipts and claim limits. No raw AI-GPR values are redistributed, and no superiority claim is made.

  • 2026-08-07, v1.10.2 (construct name locked; public/private boundary). IGRM retains its established name, while the paper title and construct line use the scientifically accurate term geopolitical salience. A dedicated comparison page identifies the exact Caldara-Iacoviello India series already tested and states that the newer AI-GPR products remain untested. An operational author queue, accidentally served as a public page and endpoint earlier the same day, was removed because it is not research data. The analytical record remains public: registrations, amendments, corrections, validation misses and sensitivity results. Removing the out-of-scope endpoint triggers API contract v2.0.0; the removal is recorded in the contract itself.

  • 2026-08-07, v1.10.1 (independent splice audit; primary unchanged). A proposed 38-day calibration-stability result was rejected before publication because 37 comparison values had been constructed from the same numerator and production ratio. The independent replacement recovers the last pre-bridge DOC-API store from git, reports 18-day re-estimates for Pakistan, China and Gulf, refuses to claim a new estimate for the two n=1 channels, and publishes the full score-level sensitivity at docs/data/splice_sensitivity.json. Pakistan and China shifts are material; the homepage carries a notice. Production ratios and every primary history value remain unchanged to preserve the v1 vintage. The rejected result and correction mechanism are recorded in the append-only corrections ledger.

  • 2026-08-07, v1.10.0 (the construct is partly Anglophone, measured). The index is built from an English-language corpus, read by an English-language matcher, and was cross-validated against English Wikipedia. Every leg of that apparatus was Anglophone, so nothing published could distinguish "salience of India-relevant geopolitical risk" from "what the international English-language press covers about India". src/wiki_hindi.py runs the test that can: the same registered article set read on Hindi Wikipedia, with Hindi titles resolved through Wikipedia's own interlanguage links rather than chosen by hand, correlated against the same daily shares over 3,394 overlapping days.

The index tracks English-language attention more closely than Indian-language attention, on five channels of five, in both levels and day-to-day changes, with no channel where Hindi leads. Changes correlations fall from 0.223/0.340/0.589/0.117/0.564 (English) to 0.160/0.132/0.163/0.050/0.030 (Hindi) for pakistan_west, china_east, gulf_energy, us_trade and shipping respectively; us_trade is the only negative level correlation in the study at −0.136.

The obvious objection is that Hindi Wikipedia simply carries less traffic, so the correlation attenuates through noise. That does not survive the data: us_trade has the highest median Hindi traffic of any channel (602 views/day) and the worst agreement, while gulf_energy has the lowest (161) and the strongest. The ordering of traffic and the ordering of agreement do not match.

Six of the twenty-nine registered articles have no Hindi counterpart at all — CAATSA, Sanctions against Iran, Houthi movement, Piracy off the coast of Somalia, Trade policy of the United States, Energy policy of India — and their absence is reported as a result rather than a coverage caveat: they are the foreign-policy-apparatus topics, and the two India-adjacent channels (pakistan_west 6/6, china_east 5/5) resolve completely while the other three lose a third of their articles each.

This changes no score, weight or published value. It is a limitation of the instrument, and anyone citing this index as a measure of Indian public salience should read docs/data/wiki_hindi.json first. Section 8c's supply-side/demand-side caveat stands; this is a sharper and less flattering version of it.

  • 2026-08-07, v1.9.0 (the quantity publishes; referee findings). Raw channel shares — the input the percentile series is computed from — are promoted to a first-class published artifact (docs/data/shares.csv, shares.json), with the store's 21 known missing days disclosed in the payload rather than left for a user to discover. This answers three findings at once, all now stated as limitations rather than implied: (a) ceiling saturation — across 2025-04-23 to 2025-05-14 the pakistan_west percentile moved 2.2 points (97.8–100.0) while the underlying share moved 5.5-fold (0.214%→1.173%), so during a sustained crisis the percentile carries almost no intensity information and the share carries all of it; (b) cross-channel incommensurability — percentiles are within-channel ranks against different reference distributions, so averaging them into a composite has no clean interpretation, whereas shares are fractions of one common daily corpus; (c) cross-time incomparability — a score of 90 in 2019 and in 2026 are ranks against different two-year baselines, while a share means the same thing in every year. The percentile series is unchanged and remains the headline because it answers "loud for this channel?", which a bare share cannot; publishing the quantity beside it lets any reader apply their own normalization and recompute the index from scratch. No published value changed in this entry.
  • 2026-08-06, v1.8.0 (the headline becomes the 7-day; founder-signed in chat the same day). The front page now leads with composite7/score7: the trailing-7-day mean of each channel's raw share passed through the identical percentile transform. Reasons, in order: (1) a rank transform amplifies the dense middle of the distribution, so ordinary news-cycle oscillation printed as 40–70 point daily swings on thin channels (china_east 33.5 → 74.3 → 4.5 across three days whose raw shares were 0.0134% → 0.0232% → 0.0070% against a two-year median of 0.0165%); (2) seven days is the minimal window that cancels global press volume's weekly periodicity; (3) the construction was verified on real onsets before signing — the 7-day read 99.5 the morning after the Pahalgam attack and 96–98 through Article 370, so nothing urgent is lost. The daily series is unchanged and fully published as the tape (chart underlay, episodes, receipts, and alerts remain daily; the separately bounded historical sampling-band artifact has the frozen window disclosed above); composite/score keep their frozen contract meaning forever, and the weekly fields are additive. No historical value of any series changed.
  • 2026-08-06, v1.7.0 (tone as a second axis; G-track layer 1). Each channel's matched articles — the same articles the registered dictionaries selected and the receipts pages enumerate — are annotated with GDELT GKG's V2Tone by URL join, and the per-channel daily mean publishes as docs/data/tone.json with the join rate disclosed (GKG does not carry every URL). Tone introduces no new article selection: the registered dictionaries remain the only selector, and tone is an annotation, never a filter. Context axis only: tone enters no score, no percentile, and no composite, and may never do so without a founder-signed memo recorded here. No score construction changed anywhere in this entry; series values are bit-identical.
  • 2026-08-06, v1.6.0 (corpus receipts and sampling bands; founder MI work order of 2026-08-05). Receipts construction changed: the primary lane now enumerates matched articles from the same sampled ngrams corpus the day's scores are computed from (matcher, anchor, and tokenizer identical to the series -- an article in this lane is one the estimator actually counted), with a bounded artlist supplement restoring wire originals whose syndicated copies the sample caught; every article is lane-labeled and the artlist-only path remains as fallback (src/receipts_ngrams.py, gated by tests/test_receipts_ngrams.py). Sampling uncertainty published and displayed: Wilson 95% intervals on each sample-estimated day's matched share, passed through the identical splice ratio and trailing-percentile transform as the point value, shaded on the composite chart and disclosed in the at-a-glance strip when a headline move sits inside two days' bands (docs/data/uncertainty.json, src/uncertainty.py, gated by tests/test_uncertainty.py). Robustness made legible: ex-ante reading conventions on the validation page, per-channel plain readings, a public discussion of the gulf_energy narrow correlation (0.527), and weekly primary-vs-variant overlay series (docs/data/robustness_series.json). The founder's first 16 firm calibration rulings now count toward the registered n=100 threshold (author-machine agreement 0.875 on the overlap; series remains flagged UNCALIBRATED). No score construction changed anywhere in this entry: series values are bit-identical; receipts, uncertainty, and validation displays are evidence layers.
  • 2026-08-04, v1.5.0 (API contract). Section 12 added: the v1.0.0 API contract (docs/data/api_contract.json, docs/api.html) freezes 27 endpoints across docs/data/*.json, three CSVs, and feed.xml, with a stated promise and deprecation policy. Nothing deprecated yet. Gated by tests/test_api_contract.py (every served payload appears in the contract; every frozen field is still present in the live payload).
  • 2026-08-02, v1.4.0 (receipts drill-down). Section 11 added: clicking any channel score on the homepage opens receipts.html, showing the exact query and a tier-sorted sample of matched articles for the latest day (docs/data/receipts.json, gated by tests/test_receipts.py). Source tiers (source_tiers.json, registered 2026-08-01) order the list credible-first and produce spike_quality_tier12_share; disclosed everywhere as never entering any score. Not a historical archive -- only the latest published day is kept.

  • 2026-08-01, v1.3.0 (comparators and predictability). Section 10 added: four-country comparator series from one registered shared vocabulary, and the directed lead-lag study whose negative result (salience predicts nothing; events marginally lead salience) is published as found.

  • 2026-07-31, v1.2.0 (stress gauge). Section 9 added: the India Stress Gauge, four pre-registered components fused into one daily 0-100 line, gated on the completed events history, hit-rate published as found. Registration precedes computation in the commit history.

  • 2026-07-31, v1.1.1 (nowcast). A provisional "today so far" score now publishes to docs/data/nowcast.json roughly every two hours, computed from a partial-day sample of the Web NGrams bridge with the v1.0.1 splice calibration and ranked against each channel's trailing 730 days exactly as a finished day is. It is labeled provisional everywhere it appears, discloses its sample size (n_samples, n_docs_sampled), never enters the historical series, and is superseded by the daily run's finalized number. The historical construction is unchanged.
  • 2026-07-31, v1.1.0 (chokepoint sub-dictionaries). The shipping channel gains four sub-dictionaries (shipping.chokepoints in dictionaries.json: Hormuz 5 terms, Bab el-Mandeb 5, Suez 3, Malacca 3, each with per-term rationale). They exist only for the salience-vs-transits comparison on the analysis page, where each corridor's weekly press salience is set against IMF PortWatch transit calls, both as percentiles of their own 2019-present weekly history. Sub-dictionary series never enter the composite. A sub-dictionary may repeat a parent-channel term (it is a decomposition of shipping, not an addition) but no term appears in two sub-dictionaries; they carry no anchor word because they measure global corridor salience, not India-linked salience. The ex-ante rule and query grammar apply unchanged and CI enforces both on the new terms (tests/test_dictionaries.py). Store: data/raw/chokepoint_salience.csv; site payload: docs/data/chokepoints.json.
  • 2026-07-29, v1.0.1. (1) Percentile computation now returns missing for days with no observed value; previously a missing day scored as 0th percentile, deflating the composite when one channel's tail lagged. (2) During the July 2026 DOC-API disruption, and at the GDELT maintainer's direction, recent days are computed from the Web NGrams v5 feed (half-hourly samples, per-document matching, English only) and ratio-spliced to the API series on overlap days. Splice ratios (log-sd, overlap days): pakistan_west 1.95 (0.38, n=5), china_east 3.36 (0.46, n=5), gulf_energy 1.79 (0.15, n=5), us_trade 2.56 (n=1), shipping 2.97 (n=1); the thin-overlap channels are calibrated on the single day both sources cover and marked accordingly. (3) The gulf_energy sub-query "crude oil supply" has not yet been fetched by any source at historical depth; the channel currently carries its main sub-query, and the series will be re-based when it lands. (4) Event-study cells now publish bootstrap p-values with Benjamini-Hochberg FDR flags (10%) across the grid of relative-outcome tests, as §6 specifies.
  • 2026-08-05, dictionaries v1.2.0 (precision amendment, founder- approved). Two full-text leak classes removed after practitioner and founder review. Shipping: bare "shipping lanes" and "maritime security" dropped (a charity swim story reached the receipts via body-text mentions of dodging shipping lanes); standing corridor geography "Gulf of Aden" and construct vocabulary "merchant vessels" added to carry the recall. China/East: bare "Arunachal Pradesh" dropped (as a state name it matched monsoon, flood and administrative coverage across the Northeast); the standing boundary term "McMahon Line" added. Both channels remain single sub-queries, so no series-construction change rides along. Effects on recall and precision will be measured by the standing auditor and published as found. Ex-ante rule enforced by CI on the amended lists as on the originals.
  • 2026-07-31, v1.1.0-dev (V2 data layers). (1) Sector event study: three sector baskets added beside defence (energy_omc, it_services, ports_logistics; members and channel hypotheses pre-registered in validation/sector_hypotheses.json before any cell was computed). Outcome grid doubled, so the 10% FDR threshold tightened: 4 cells now flag significant (was 5); the new sector hypotheses are largely NOT confirmed at this threshold, a pre-registered negative reported as such. (2) Events stream: daily GDELT Events v1 counts for India (national, bilateral-dyad, and state layers; data/raw/events_*.csv), backfilling to 2017. (3) Physical flow: IMF PortWatch daily transit calls for Suez, Bab el-Mandeb, Malacca, and Hormuz (data/raw/portwatch_chokepoints.csv, 2019-present, revisions upserted). Attribution: IMF PortWatch (portwatch.imf.org).
  • 2026-07-24, v1.0.0. Initial dictionaries frozen (five channels, 10-14 terms each, per-term rationale in dictionaries.json). Robustness variants and placebo channels frozen the same day. Validation episode list pre-registered (13 episodes, 2022-2025). Parameters: 730-day percentile window, 180-observation minimum, 2σ/90-day/3-day episode rule, 1/5/20 trading-day event windows, 1,000-resample bootstrap.