Codebook

Purpose. This document exists so a researcher can reuse IGRM data without reading the source code or emailing the author: every published file, every field in it, its units, and its construction, in one place. Definitions here are column-by-column and mechanical; the reasoning behind each construction choice lives in the methodology, and the evidence that the constructions measure anything lives in validation. If a published field is not documented here, that is a defect — report it.

Researcher quick-start

Fetch, cite, and reproduce in five lines. No key, no rate limit, CORS open; field stability is promised by the frozen API contract.

# 1. The full daily series, 2017-present (composite + five channels):
curl -s https://igrm.in/data/history.csv -o igrm_history.csv
# 2. Today's finalized scores with metadata:
curl -s https://igrm.in/data/latest.json | python -m json.tool
# 3. The evidence behind today's numbers (per-channel matched articles):
curl -s https://igrm.in/data/receipts.json | python -m json.tool
# 4. The validation record (hit-rate, placebo, robustness, drift):
curl -s https://igrm.in/data/validation.json | python -m json.tool
# 5. Reproduce the index from source (pinned deps, same raw stores):
#    git clone https://github.com/PlausibleDissent9/india-geopolitical-risk-monitor

Citation: Krishna, Ishan (2026). India Geopolitical Risk Monitor. https://igrm.in/ — data CC BY 4.0, code MIT. A ready-made BibTeX entry is at igrm.bib.

Upstream data sources and attribution

IGRM's published values are derived data. The CC BY 4.0 licence above covers IGRM's own derived output; it does not and cannot relicense the upstream sources, each of which carries its own terms.

The GDELT Project — https://www.gdeltproject.org/

Every channel share, the composite, the event aggregates, the tone and theme series and the per-channel receipts are derived from datasets released by The GDELT Project. GDELT makes its datasets available for unlimited and unrestricted use, on one condition, quoted from its terms:

any use or redistribution of the data must include a citation to the GDELT Project and a link to this website (https://www.gdeltproject.org/)

That condition attaches to any use, not only to redistribution, and this section exists to meet it. Anyone reusing IGRM data that originates from GDELT inherits the same obligation and should carry this attribution forward.

Other upstream sources, each under its own terms and cited where used: IMF PortWatch (chokepoint transits), UCDP Georeferenced Event Dataset (conflict events), Wikimedia pageviews, Correlates of War MID 5.0, JODI oil statistics, UN Comtrade, Indian major-port statistics, and the Caldara–Iacoviello Geopolitical Risk index used as a comparison series. Their rights positions are recorded per source in governance/source_rights_registry.json, including which are still under review and therefore not yet drawn on.

data/raw/gdelt_volume.csv

Column	Definition	Units	Range
`date`	Calendar day, UTC (GDELT's day convention)	ISO date	2017-01-01 onward
`pakistan_west` … `shipping`	Share of all GDELT-monitored articles that day matching the channel's query; where a channel needs two sub-queries (length budget), the SUM of the two shares, a slight upper bound on the union	percent of corpus	≥ 0, typically ≪ 1

docs/data/episodes.csv

Column	Definition	Units
`channel` / `label`	Machine key and display name of the spiking channel	text
`start`, `end`	First and last spike day of the cluster	ISO date
`peak_date`	Day of maximum raw coverage share inside the episode	ISO date
`peak_value`	That maximum share	percent of monitored corpus
`n_spike_days`	Spike days in the cluster	count

docs/data/event_study.csv

One row per channel × outcome × window. mean_cum_log_return_pct is the mean cumulative log return (percent) over window_trading_days after episode starts; ci95_lo/ci95_hi bound the bootstrapped 95% interval; n_episodes is the number of episodes with sufficient data; descriptive_only marks configured outcomes (Brent, gold) with no separable India-specific component. available_outcomes and unavailable_outcomes in the JSON disclose what the current source cache actually supports; an unavailable configured outcome has no CSV rows and is never described as reported. Association, not causation.

JSON `_meta` blocks

Every dict-shaped JSON published by the pipeline embeds a _meta object (what the file is, license, citation, codebook link, source URL, generation date — plus units where the payload has a single unit) so a downloaded file explains itself without this website.

Two payloads are bare JSON arrays with nowhere to carry _meta: episodes.json and notes.json. Both are written in the same commit as files that are stamped (history.json, note_latest.json), so their provenance is one file away rather than absent.

One dict-shaped exception is deliberate: ai_gpr_benchmark.json is served byte-for-byte as the hash-pinned output promised by its public registration. Its own _meta carries the analysis registration, script and source hashes, provider citation and raw-value redistribution limit. The generic metadata stamper skips it because changing even an innocuous envelope field would break the published result hash.

public_api_byte_manifest.json is the second deliberate deterministic exception. Its _meta includes its meaning, licence, citation, codebook and source URL, but no generation clock: a wall-clock field would churn the bytes even when none of the 117 hashed endpoints changed. The file hashes one captured Git candidate index/tree, excludes exactly itself to avoid recursion, and is skipped by the generic stamper after generation. It is unsigned. An independently obtained SHA-256 digest of the manifest's own raw bytes is needed for an external comparison; the file does not authenticate a deployment or claim that 117 separate hosted requests were served atomically.

This was a promise before it was a fact. Until 2026-08-07 the licence, citation and codebook link appeared in three payloads out of sixty-four, while this paragraph claimed all of them — so most downloads carried no way to know their own terms or how to cite them. src/stamp_meta.py now stamps the universal fields after every lane has run, without ever overwriting a field a module set for itself. A test checks the claim against what is actually served, so the sentence cannot drift from the files again.

docs/data/latest.json

Field	Definition
`date`	Last day with a full composite
`definition`	The one-line construct definition shown under the headline
`composite`	Unweighted mean of the five channel percentiles, 0-100
`channels.<ch>.score`	Channel percentile vs its own trailing 730 days, 0-100
`channels.<ch>.label`	Display name

docs/data/history.json

Field	Definition
`dates[i]`	ISO date for position `i` in every parallel array
`composite[i]`	Composite percentile that day (null before 180 trailing obs)
`channels.<ch>[i]`	Channel percentile that day
`labels`	Channel display names
`wikipedia.*`	Same structure computed from Wikipedia pageviews (present once the second source is active); levels are never comparable across sources, each is percentile-normalized independently

docs/data/episodes.json

Array of episodes, each:

Field	Definition
`channel` / `label`	Which channel spiked
`start`, `end`	First and last spike day (raw share > trailing 90-day mean + 2σ, baseline lagged 1 day; days ≤ 3 apart cluster)
`peak_date`, `peak_value`	Day and value of the maximum raw share inside the episode; value in percent of corpus
`n_spike_days`	Count of spike days in the cluster

docs/data/event_ledger.json

The Global Event & Episode Ledger is a count-unit firewall over the current event substrate. It must not be cited as a global event census. Its current public artifact is a value-free refusal state: GDELT Events, Web NGrams, IGRM first-party payloads and Natural Earth all remain review_required in the signed source-rights registry, so this endpoint emits no source-derived count, episode or geometry denominator. Candidate validation can run internally, but it cannot override that gate. This new compiled-endpoint gate neither licenses nor silently retracts existing source-specific endpoints; their separate rights decisions remain review_required. The retained GDELT event store also lacks raw rows, stable GlobalEventID values and revision lineage, so deduplicated and canonical event counts remain unavailable even after rights approval.

Field	Definition
`_meta.artifact_status`	Currently `public_release_blocked_rights_review`; an authorized future value release uses `public_observation_foundation`
`rights_gate`	Exact rights-registry and signer-registry hashes, required source IDs and uses, decision states, and blocked IDs
`frame`	`null` while blocked. An authorized release separates the 247-member display geometry, India-as-self/not-applicable, mapped members, reason-unresolved non-observations and provider codes that are unmappable
`count_units.*.public_available`	Always `false` in the refusal artifact; value is `null`. Candidate computation is not public availability
`count_units.aggregate_source_rows`	Future authorized counts use rows that passed the expected layout, not every provider-export row; malformed-row counts are not retained in the legacy store
`count_units.deduplicated_source_events`	Unavailable because stable source-event identity and revision lineage were not retained
`count_units.canonical_geopolitical_events`	Unavailable until cross-source resolution, rights, method and signed-release gates pass
`count_units.detected_salience_episodes`	Rights-gated threshold-defined IGRM channel windows; detector output, not real-world event identity
`aggregate_historical_series`	`null` while blocked. An authorized release defaults to India-row share of valid-layout rows; raw levels are secondary and cannot be treated as like-for-like across unregistered provider regimes
`episodes`	`null` while blocked. An authorized release uses stable detector-window IDs, `provisional_open_window` until the three-day cluster gap is observed, and `canonical_event_ids: null`
`canonical_event_layer`	Target lifecycle vocabulary and the non-negotiable requirements for a future production release
`release_lineage`	No public vintage exists while blocked. Every authorized release is predecessor-bound, complete-state hashed, carries a typed date/episode delta and requires a detached Ed25519 signature from a separately pinned release trust root. Source-rights approval alone cannot sign a public vintage; an agent-created key or self-declared role has no authority. CI replays every first-parent commit that ever changed the archive, refuses removal or byte changes to any predecessor and permits only the next sequential vintage, so an innocent tip cannot hide an earlier rewrite. The delta is recomputed from predecessor bytes. `released_at` must follow every bound evidence/knowledge date, increase strictly, and remain within five minutes of the verifier clock

Authorized release-content and whole-artifact digests use igrm-typed-canonical-f64-v1, not the host runtime's JSON number formatting. Every number is encoded by its finite IEEE-754 binary64 bits, integers outside the JavaScript safe range are refused, negative zero is preserved, and strings and object-key ordering use UTF-8 bytes. Python and browser implementations are checked against the same fixtures in validation/event_ledger_canonicalization.json, including 1/1.0, -0.0, exponent boundaries, non-BMP Unicode and a release-shaped projection. A profile mismatch or fixture failure refuses the authorized UI instead of silently disagreeing about a signed digest.

The current refusal artifact is partial: true because values are withheld. Inside an authorized release, partial: false would mean only that the declared calendar, aggregate-store and display-geometry partitions reconcile. It would not mean all geopolitical events, provider members, sources or canonical event states have been observed. The legacy unavailable-day register has no immutable retrieval receipts; those dates are not described as verified provider outages.

docs/data/event_study.json

Field	Definition
`windows`	Trading-day windows (1, 5, 20), inclusive of first trading day ≥ episode start
`units`	Cumulative log return, percent
`descriptive_only`	Configured outcomes with no separable India-specific component (Brent, gold)
`available_outcomes` / `unavailable_outcomes`	Which configured series have at least one observation in this release
`channels.<ch>.outcomes.<o>.<w>`	`{mean, ci95:[lo,hi], n}` across the channel's episodes; CI from 1,000 episode resamples
`per_episode.<ch>[]`	`{start, outcomes.<o>.<w>}` raw window returns for one episode (no CI; n = 1)

Configured outcomes: nifty_minus_em (strips global equity beta), defence_minus_nifty, energy_omc_minus_nifty, it_services_minus_nifty, ports_logistics_minus_nifty, usdinr_minus_dxy (strips broad dollar moves), brent_ret, and gold_ret (descriptive). Consult available_outcomes and unavailable_outcomes for the current release: a configured unavailable outcome has no CSV result rows. Brent and gold remain descriptive-only whenever available.

docs/data/validation.json

Key	Definition
`hit_rate`	Pre-registered episode list vs detected episodes, ±3 days; overall and per channel, plus per-episode hit/miss
`placebo`	Episodes detected in placebo channels and their overlap with geopolitical episode days
`robustness.narrow/.broad`	Correlation of primary percentile scores with narrower/broader dictionary constructions, per channel and composite (2022 onward)
`cross_source`	Per-channel correlation of GDELT and Wikipedia percentile scores
`drift`	Mean daily monitored corpus by year; sampled domain diversity and top-10 Herfindahl per channel-year; per-channel share-vs-corpus correlation

docs/data/alt_specs.json

Key	Definition
`composite_weightings`	Equal (primary) vs trade-/import-exposure composites; correlation with primary
`episode_thresholds_sigma`	Episode counts at 1.5σ/2.0σ/2.5σ and episode-day Jaccard overlap vs primary
`percentile_windows_days`	Composite and per-channel correlation of 365-/1095-day windows vs the 730-day primary

docs/data/seasonality.json

Key	Definition
`channels.<ch>.partial_r2`	Share of detrended log-share variance explained by smoothed day-of-year effects
`channels.<ch>.top_recurring_dates`	Calendar dates with the largest positive day-of-year effects (log points)
`channels.<ch>.corr_seasonal_vs_deseasonalized`	Correlation between primary and deseasonalized percentile indices

docs/data/priced_risk.json

Key	Definition
`gap.dates[i]`, `gap.<ch>[i]`	Attention percentile − India-VIX percentile, shared trading days; positive = press louder than the market
`divergence_episodes.press_louder/.market_ahead`	Top-decile
`lead_lag.ccf[]`	`{lag, corr, lo, hi}`: correlation of daily changes at each lag (−10..+10 trading days), moving-block bootstrap 95% bands; positive lag = attention leads
`lead_lag.reading`	The associational one-line interpretation

data/raw/events_daily.csv

The second measurement modality (V2): counts of recorded events involving India from the GDELT Events v1 daily archive, one row per day since 2017-01-01. Where the salience series measures how loud the press is, this series counts what the press recorded happening. An event involves India when either CAMEO actor is coded IND or the action geography is India. Events are dated by the file they arrive in, the day GDELT recorded them, matching the salience series' "when did the world's press carry this" framing.

Column	Definition
`date`	UTC day of the GDELT v1 daily export file
`n_global`	Total events recorded worldwide that day (normalization denominator; GDELT's coverage grows over time, so raw counts drift without it)
`n_india`	Events involving India (either actor IND, or action geography India)
`n_verbal_conflict`	India events in CAMEO QuadClass 3 (verbal conflict)
`n_material_conflict`	India events in CAMEO QuadClass 4 (material conflict)
`n_protest`	India events with CAMEO root code 14 (protest)
`goldstein_mean`	Mean Goldstein cooperation-conflict score of India events (−10 most conflictual, +10 most cooperative)
`mentions_sum`	Total press mentions across India events that day

data/raw/events_dyads.csv

The relations layer: India's bilateral event record with every partner country, one row per (day, partner). A dyad event has India as one CAMEO actor and the partner as the other; no dictionaries or curation are involved, only actor codes, so this layer carries no term-selection bias.

Column	Definition
`date`	UTC day of the GDELT v1 daily export file
`partner`	CAMEO 3-letter code of the non-India actor (e.g. PAK, CHN, USA)
`n`	India-partner events that day
`n_coop`	Of those, cooperative (QuadClass 1 verbal / 2 material cooperation)
`n_conflict`	Of those, conflictual (QuadClass 3 verbal / 4 material conflict)
`goldstein_mean`	Mean Goldstein score of the dyad's events (−10 to +10)
`mentions_sum`	Total press mentions across the dyad's events

data/raw/events_states.csv

The internal layer: events geolocated to an Indian state, one row per (day, state). adm1 is the FIPS ADM1 code (IN01-IN36; IN00 marks country-level geocoding without a state).

Column	Definition
`date`	UTC day of the GDELT v1 daily export file
`adm1`	FIPS ADM1 code of the state the event action was geolocated to
`n`	Events located in that state that day
`n_conflict`	Of those, QuadClass 3 or 4
`n_protest`	Of those, CAMEO root code 14

data/raw/portwatch_chokepoints.csv

The physical-flow layer (V2): daily ship transits through the four chokepoints most relevant to India's trade, from IMF PortWatch (portwatch.imf.org), which derives them from satellite AIS with a published methodology. Free with attribution; PortWatch publishes with roughly a five-day lag and revises recent days, so updates re-fetch a trailing window and revisions win.

Column	Definition
`date`	Calendar day (UTC)
`chokepoint`	`suez`, `bab_el_mandeb`, `malacca`, or `hormuz`
`n_total`	Ship transit calls that day
`n_tanker`	Of those, tankers
`n_cargo`	Of those, cargo vessels
`capacity`	Total transiting capacity (deadweight tonnage-based, per PortWatch)

data/raw/chokepoint_salience.csv

Daily GDELT volume-intensity share per chokepoint sub-dictionary (shipping.chokepoints in dictionaries.json v1.1.0), 2019-present. The 2019 floor matches PortWatch coverage: these series exist only for the salience-vs-transits comparison and never enter the composite.

Column	Definition
`date`	UTC day
`hormuz`, `bab_el_mandeb`, `suez`, `malacca`	Share of all GDELT-monitored articles matching that sub-dictionary (sum of sub-query group shares, as for channels)

docs/data/chokepoints.json

Site payload for the analysis-page chokepoint chart: per chokepoint, the weekly-mean salience and PortWatch n_total transits, each ranked as a percentile of its own full 2019-present weekly history, plus a weekly Spearman correlation on levels and the latest salience-minus-transits percentile gap.

docs/data/ai_gpr_benchmark.json

Static output from the public, code-frozen AI-GPR India benchmark. The primary statistic is Spearman rho between month-over-month changes in the monthly mean IGRM daily composite and AI-GPR India_all. sample publishes every eligible month and the no-bridging rule; primary carries the point estimate, registered six- and twelve-month moving-block intervals, seed and mechanically selected decision sentence. descriptive_primary contains the registered level and Pearson checks plus lag-1 autocorrelations. exploratory_correlations is the labelled 6 by 4 matrix; episode_month_ranks is descriptive and deliberately has no AI-GPR hit rate; largest_rank_divergences republishes ranks and IGRM evidence links, never raw AI-GPR values. The source, IGRM inputs and analysis script are all SHA-256 pinned in analysis/ai_gpr_benchmark_registration.json.

docs/data/divergence_register.json

Append-only rows documenting large disagreements between IGRM ranks and independent comparator ranks. Every entry has a stable id, publication and observation periods, register family, the two within-sample ranks, absolute gap, direction, sample definition, evidence URL, source payload and a claim limit. Samples differ by family and are never pooled. Rows identify where to inspect; they do not decide which measure is correct or assign a cause. Corrections retain the original id and add a dated note rather than silently removing history.

docs/data/nowcast.json

PROVISIONAL today-so-far scores, replaced roughly every two hours by the nowcast workflow. Computed from a partial-day sample of the GDELT Web NGrams bridge (same machinery and splice calibration as the heal path), each channel ranked against its trailing 730 days exactly as a finished day would be. n_samples and n_docs_sampled disclose the sample behind the number. Never enters the historical series; the daily run's finalized score supersedes it, and the site renders it only while the payload's UTC date is still today.

docs/data/map_relations.json and map_states.json

Aggregates behind the Maps page, published only once the 2017 events backfill is complete (_meta.partial stays false on anything served). Relations: per partner country (keyed by the dyad actor code), total and trailing-365-day event counts, conflict share of classified events (QuadClass 3-4 vs 1-2), event-weighted Goldstein mean. States: per Indian state, total and trailing-365-day located events with conflict (QuadClass 3-4) and protest (root 14) shares. GDELT still emits some pre-2000 FIPS state codes; they fold into today's boundaries (IN09 into IN35 Madhya Pradesh, IN33 into IN36 Uttar Pradesh, IN25 into IN22 Tamil Nadu, IN06 into IN52 Dadra and Nagar Haveli and Daman and Diu), so a pre-2000 parent's counts include its later split-off child. IN00 (a country-level geocode) and CAMEO regional actor codes (SAS, EUR, AFR, ...) are excluded and listed with row counts in each payload's excluded block.

docs/geo/world.json and india.json

Self-hosted map geometry, baked once by scripts/prepare_map_geometry.py from Natural Earth public domain data (110m admin-0; 10m admin-1 filtered to India) and committed; the site fetches no external geometry. Countries keyed by ADM0_A3, Indian states by FIPS. States too small for 110m polygons carry a point marker instead of a path.

docs/data/stress_gauge.json

The experimental India Stress Gauge (methodology section 9): today's 0-100 gauge, the per-component percentiles behind it (press, events, market, wikipedia), the pre-registered weights, the hit-rate against the pre-registered episode list with per-miss detail, and a 365-day history. _meta.registered_weights records the intended four-component design. _meta.latest_available_components, latest_missing_components and latest_effective_weights disclose any per-date renormalization; _meta.partial is true when either the events history or a latest-date component is missing. The current 2/29 validation result is experimental and the gauge is not eligible as a headline measure. Weights, detection rule, and missing-component rule live in validation/stress_gauge_weights.json, committed before any hit-rate was computed.

docs/data/receipts.json

Per-channel receipts for the latest published day only (not a historical archive). Three lanes, labeled per article in lane: "corpus" articles were matched inside the same sampled ngrams corpus the day's scores are computed from (the estimator actually counted them; matcher, anchor, and tokenizer identical to the series); "corpus-extended" articles matched under the identical matcher in the REST of the day's minute-files, outside the half-hourly scoring sample (they exist in the day's monitored coverage; the estimator did not count them); "artlist" articles come from GDELT's bounded relevance search, restoring wire originals whose syndicated copies the corpus caught. n_matched_in_corpus, n_matched_extended, and n_artlist_supplement count the lanes before URL dedup; n_pool_unique counts the merged pool after syndication dedup; up to 150 publish per channel. Each article carries its source tier from source_tiers.json (null = "unranked", never assumed into a tier) and a match tag ("headline" when a channel phrase is in the title, else "full-text"). Ordering is credibility (tier), then visible aptness, then arrival order. spike_quality_tier12_share is the tier 1-2 share of the published list. _meta.method says which lane(s) produced the file (the artlist-only fallback publishes when the corpus files are unavailable). Tiers order presentation only and never enter any score.

docs/data/receipt_identity.json

An independently retryable source-link surface for the exact completed UTC D-1 day. It is not a score input and never writes or substitutes for receipts.json, receipts_archive.json, the NGram caches, shares, history, or latest scores. Each of the five exact registered channels is either available with at most five deduplicated {title,url,domain} records or unavailable with a typed reason. An empty articles array means a completed request returned zero valid links; an unavailable block is missing evidence, not zero.

The lane is default-deny. The committed profile remains inactive_pending_human_signature, the GDELT DOC source decision remains review_required, and production signer trust is empty, so the current public form is a value-free refusal and no source request occurs. Future activation requires both a trusted human-signed source decision and a separately signed closed profile authorizing exactly cite_metadata, model_processing, and publish_extract. It never requests, retains, or republishes article bodies, snippets, descriptions, bylines, images, story/document identifiers, raw responses, or full source records. GDELT requires attribution for use of its datasets. Headlines here identify and link to publisher-controlled works; GDELT's dataset terms are not presented as a publisher copyright sublicense.

The receipts page renders these links only when target_date exactly matches the current published score date. A stale or future identity payload therefore cannot masquerade as the score day's evidence. Source or rights failure affects only this payload and never blocks the aggregate score.

Same-target retries are monotone over the exact regular Git blob at the current remote parent. Every previously available channel retains its exact ordered article rows, while later retries may recover additional channels. The check is repeated against the exact candidate parent immediately before push, so a rebase cannot silently replace a stronger same-day payload with a weaker or different one. Only a current rights refusal may withdraw all article values; that transition emits the closed value-free unavailable state instead.

docs/data/uncertainty.json

95% sampling bands for sample-estimated daily scores, keyed days.<date>.<channel> = [lo, hi] in percentile-score units (0-100), plus days.<date>.composite. Construction: the day's matched share (sum of sub-query group shares over n sampled documents; the sum-of-groups construction can count a document under two groups, so the effective count inherits that convention) gets a Wilson 95% interval, and both bounds pass through the same splice ratio (ngram_calibration.json) and the same trailing-percentile transform as the published point value. The composite band is the mean of channel bounds — an envelope, conservative because channel sampling errors are independent. The frozen artifact covers exactly _meta.first_banded_date through _meta.last_banded_date (currently 2026-06-30 through 2026-08-07). Earlier and later published days carry no published band; absence is not a zero-width interval or a claim of certainty.

docs/data/robustness_series.json

The overlay behind the validation page's robustness correlations: weekly mean percentile scores of the primary index and its registered narrow/broad dictionary variants (dictionaries_alt.json), keyed channels.<ch>.{primary,narrow,broad} aligned to the shared weeks axis, over the window where all three exist (_meta.window; the variants begin 2022-01-01). null = the variant store has no data that week, shown as a gap, never interpolated.

Conventions

Every percentile is computed against the series' own trailing 730 days, minimum 180 observations, never using future data.
The window is 730 calendar days, ending on and including the day being scored, not the last 730 observations. The distinction is real because the series has genuine gaps, so the two readings select different windows on either side of a gap.
The percentile is the share of that window at or below the day's value, times 100. Ties count as at-or-below, not strictly below and not half-credit.
Published values are rounded half-to-even (Python's round). Values exactly half-way therefore round to the nearest even final digit.
GDELT days are UTC; market data uses exchange trading days (IST for NSE).
All return language is associational. Nothing here is investment advice.

Those conventions are sufficient for src/blind_replicator.py to rebuild the series from this codebook and the published shares.csv without importing the pipeline. Every published daily channel and composite score cell must match, and missing reconstructed cells count as failures rather than disappearing from the denominator. The current cell count and exact result are recomputed nightly at data/replication.json; a methodology change that is not reflected here breaks the exact-agreement test.

docs/data/expert_shelf.json

Latest publications from the registered think-tank roster (think_tanks.json), titles and links only, tagged to channels by the frozen dictionaries; items matching no channel are kept as general. Institutions without a public feed are carried in the registry with that status. As of 2026-08-06 the roster includes ReliefWeb (UN OCHA), India primary-country filter — UN-provenance situation reports joining the shelf under the same titles-only rule (S4); ReliefWeb is an evidence shelf source only and, like every aggregation, never a data input to any score.

docs/data/precision.json

The standing precision audit: machine labels under the versioned rubric (auditor/RUBRIC.md), per-channel precision published as found, never tuned. Marked UNCALIBRATED until the author's independent labels reach the registered threshold (n=100); the agreement statistic then publishes with its n. Abstentions counted, not guessed.

docs/data/reliability.json

The morning contract's measured record, computed from git commit timestamps: day D final by 06:30 IST on D+1 (06:00 before 2026-08-11). Pre-contract days appear as context; misses stay listed forever.

docs/data/predictions.json

The Prediction Archive's registry: every dated, falsifiable expectation, registered before outcomes are known, graded at horizon, append-only.

docs/data/comparators.json

Four comparator country series (India, Pakistan, Indonesia, Vietnam) from one registered shared vocabulary (comparators.json at repo root) anchored per country. countries.<k>.weeks[] and .pct[] are weekly means of the trailing-730-day percentile of that country's own coverage-share history; only complete weeks publish. latest repeats the final complete week's value. Cross-country comparison is rank-vs-rank; levels reflect Anglophone press attention, disclosed.

docs/data/predictability.json

Directed lead-lag study on daily changes: for each pair, the R-squared increment from adding five lags of the candidate leader to five own lags, with a permutation p-value from time-shifted nulls (300 draws). Reported as found; predictability in this sense is association across time, never causality.

docs/data/episode_actors.json

Per-episode driver decomposition from the dyad layer, keyed "channel|start": partners whose India-pair event counts in the episode window most exceed their trailing 90-day baseline, split into the channel's registered partners (attribution) and other unusually active dyads (context, never attribution), with cooperation-conflict splits.

docs/data/api_contract.json

The frozen v2 API contract (scripts/generate_api_contract.py, committed and hand-frozen, never regenerated by the daily pipeline): every served endpoint's format, description, and top-level "frozen fields" (JSON keys or CSV columns) promised stable within major version 2, plus the contract's promise, deprecation policy, and access terms in _meta. Rendered for humans at docs/api.html. See methodology.md section 12 for the versioning rule.

docs/data/exposure_traversal_demo.json

A deterministic synthetic L0 traversal through the canonical exposure edges in the shared Max fixture world. It starts from the fixture event, selects only typed, method-bound paths allowed by the signed synthetic release, retains edge units and denominators, and carries explicit no-path and missingness semantics. Its release, object, method, schema, rights and temporal identities are the same ones certified across the other published Max engine records by max_state_join_demo.json.

This is a fixed conformance vector, not a live lane. It establishes that the traversal contract composes without cross-world identity drift; it does not establish a real event, entity, source, right, dependency, exposure, propagation, forecast, causal relationship, recommendation or adoption.

docs/data/max_state_join_demo.json

The deterministic shared-world certificate for the synthetic Evidence Output, Exposure Traversal, Sensor Fusion and Shock Compiler records. It verifies one release identity, no object identifier bound to competing content digests, one rights position per source, compatible temporal boundaries and exact declared coverage denominators. The certificate computes evidence_class = synthetic_nonproduction and licensed_maturity = L0.

Agreement is not accuracy. This certificate proves composition of four engine contracts over one synthetic fixture; it does not re-derive their results or promote them into real observations, real exposure, production readiness, external utility or institutional adoption.

docs/data/public_api_byte_manifest.json

Deterministic byte inventory for every contract endpoint except itself. Each entry binds the public path, repository path, contract format, Git mode, byte length and SHA-256 digest. universe publishes all three denominators: 120 contract endpoints, 119 hashed endpoints and one exact self-exclusion. The generator reads one stage-0 Git index or named Git tree; mutable worktree overlays are not inputs. The generic publisher refreshes it after rebase and before the committed gate, while the final-publication CAS adds it to both the rights-authorized final class and the exact value-free refusal class.

This is repository-candidate consistency only. It is not a signature, deployment attestation, rights decision, source-truth proof, atomic-hosting claim, penetration test or security certification.

docs/data/daily_brief.json

Withdrawn 2026-08-08; stable-shaped null tombstone served during the v2 deprecation window through at least 2026-11-06. The experimental language-model brief stated stress-gauge values absent from its supplied context, interpreted a tier-sorted displayed-source share as pool quality, called displayed title-key representatives the articles underlying a score, and eventually paired 2026-08-07 scores with 2026-08-06 receipts. Its prediction-language lint did not establish factual grounding.

The workflow and module now stop before any model call or prose-payload write. The endpoint retains its frozen top-level fields, but composite and every channel are null and _meta.status is withdrawn_factual_grounding_failure. Generated prose may return only under a new, versioned design that mechanically ties every numeric and entity claim to cited payload fields and rejects unsupported text. The failed outputs and withdrawal are recorded in the corrections ledger; they are not IGRM evidence.

docs/data/ucdp_context.json

Monthly organized-violence event counts and best-estimate fatalities located in India, from the UCDP Georeferenced Event Dataset (pinned annual release, finalized) extended by UCDP Candidate monthlies (preliminary: true per month; replaced when the next annual release finalizes them). Keyed india.<YYYY-MM> = {events, deaths_best, preliminary}, 2017 on. Context beside the salience index, never an input: IGRM measures press attention, UCDP records violence, and divergence between them is a documented finding. Cite UCDP for these numbers, not IGRM. Raw per-event store (id-keyed, revisable): data/raw/ucdp_events.csv.

docs/data/receipts_archive.json

Available recent receipts archive: per channel and day (up to 7 days from the committed corpus day-caches), the matched articles assembled identically to receipts.json (same matcher provenance, tier sort, syndication dedup, lane labels), capped at _meta.per_day_cap per channel-day (n_matched still reports the day's full match count). Today's full list lives in receipts.json. _meta.available_days and complete_window disclose whether all seven cache days exist. The archive never implies a complete week when it does not have one, and depth comes from days, never from padding.

docs/data/episode_terms.json

Term-level attribution per channel-day: how many matched-article HEADLINES in the day's corpus cache contain each registered phrase (days.<date>.<channel>.term_headline_hits), plus n_matched_titles. Derived from the same corpus the scores come from; reproducible from the repo alone. Headline attribution undercounts body-text phrases (stated), and the record starts at _meta.first_day — earlier episodes carry no attribution rather than a refetched guess.

docs/data/aptness.json

Machine aptness labels for every article in the current receipts.json, keyed channels.<ch>.<url> = "ON"|"OFF"|"ABSTAIN", produced by claude-haiku-4-5 under the registered rubric (auditor/RUBRIC.md, version in _meta.rubric_version) from headline and domain only. Display-layer only: the receipts page collapses OFF articles behind a visible count with a toggle — nothing is deleted, and no score is touched anywhere. _meta.founder_agreement publishes the machine's measured agreement with the founder's own rulings (the human labels remain the calibration gold standard). _meta.date must match receipts.json's date for labels to apply; on mismatch the site shows unfiltered receipts.

docs/data/status.json

Operational status, derived entirely from committed files: sources[] — each data source's latest data day (latest_data), its age in days, and the ex-ante freshness window (expect_within_days) it is judged against, with ok computed by rule, plus a note when data is absent for an honest reason. A derived file's build date may appear separately as output_generated, but is never substituted for an unavailable upstream data date; lanes[] — each selected registered operational lane's last write stamp (last), its explicit measured_day where one exists, and the file both come from; alignments[] — registered cross-payload measurement-day joins, with the two evidence files and fields, the observed day_difference, one of aligned|lagging|ahead|unavailable|invalid, and the reader-visible effect of a mismatch; morning_contract — the on-time rate summarised from reliability.json. The windows are operational expectations, not promises: the morning contract is the only promise. A valid older measurement day is published as state and does not fail publication; features that require that join must refuse or fall back as stated in the alignment record. Rendered at status.html, regenerated by the daily run (src/status_data.py).

docs/data/permanence.json

The permanence lane's record (M2): the result of the latest Internet Archive Save-Page-Now request for each key page (wayback[] — url, HTTP status, ok) and the latest Software Heritage save-code-now registration of the repository (software_heritage). Runs twice daily after the publish slots (.github/workflows/permanence.yml), fail-soft: an archive outage is recorded, never hidden, and never blocks the pipeline. Captures are browsable at web.archive.org/web/*/igrm.in; the repository's archived history at archive.softwareheritage.org. Citability is designed to survive this project's own infrastructure.

docs/data/exposure_sectors.json

The registered sector-to-channel exposure map (sectors.json v1.0.0) with the deterministic weighting rule the exposure page computes with (_meta.weight_rule): shares normalize to 1, each sector's share divides equally across its registered channels, channel weight sums across sectors. Per sector: label, channels, rationale, nse_index (null until the measured-sensitivity lane ships). Retrieval and description only — an exposure profile re-cuts published numbers by the reader's own frame and adds no new claim, no forecast, and no risk score. Rendered at exposure.html; profiles encode in the URL (?s=sector:share,...), no accounts, nothing stored. Amendments to the map are dated changelog entries by the founder, never silent.

docs/data/tone.json

Tone as a second axis (G-track layer 1): per channel per day, the mean GKG V2Tone of the channel's MATCHED articles — the same articles the registered dictionaries selected — joined by URL in BigQuery (days.<date>.<channel>.mean_tone, GDELT's conventional roughly −10..+10 scaling), with the join rate disclosed (n_tone_found vs n_matched; GKG does not carry every URL). Tone is an annotation, never a filter, and enters no score, percentile, or composite (methodology changelog v1.7.0). Absent without the CI lane's GCP credentials — fail-closed, the axis is simply missing, never guessed.

docs/data/spike_breadth.json

Spike breadth (G-track layer 3): per channel per day, the number of distinct source domains among matched articles (days.<date>.<channel>.n_domains), total matches, and the largest single domain's share (top_domain_share; near 1.0 means a one-outlet story, low values a broad-based narrative). Computed from the committed receipt day-caches with the identical union-plus-anchor arithmetic the day's estimator uses — the registered dictionaries chose every article counted. Descriptive context beside spike-quality; breadth enters no score. Depth grows automatically as extended day-caches accrue.

docs/data/sector_sensitivity.json

Measured sector sensitivities (E2, on the founder-signed sectors.json v1.1.0 mapping): for each sector with a signed NSE index, the index's cumulative return relative to Nifty over 1/5/20 trading days from the start of each episode in the sector's registered channels (sectors.<key>.channels.<ch>.<window>: mean, ci95, p_boot, n, bh_significant_10pct). Seeded bootstrap; Benjamini-Hochberg at 10% across the full grid, same discipline as the event study — on first computation NO cell survived correction, and that negative result publishes in _meta.multiple_comparisons. Association, never causation; sensitivities change no weight, score, or exposure reading. Sectors without a signed index publish nothing. Yahoo prices are used but not redistributed.

docs/data/sector_<key>.json (fifteen per-sector feeds)

E4: one machine-consumable feed per registered sector (sector_textiles_apparel.json … sector_petroleum_refining.json), regenerated nightly as a pure re-cut of already-published payloads: the sector's registered channels with today's score, 95% sampling band, weight under the registered exposure rule, and a receipts URL per channel; measured_sensitivity appears only where a signed NSE mapping exists (see sector_sensitivity.json). Nothing is computed here that is not already published elsewhere; the salience-not-risk statement travels in every file's _meta.

docs/data/alerts.json

Registered alert events (design/alerts_webhook.md, founder-signed 2026-08-06), Phase 0 poll model: alerts[] newest first, 90-day retention, stable ids for consumer dedupe. Types: T1 band_separated_move (a channel's 95% sampling band does not overlap the previous day's — the arithmetic that separates real moves from sampling noise), T2 composite_percentile_crossing (trailing 90th percentile, 730-day window, and both band edges on the crossing's side), T3 integrity_event (the morning contract's own misses, because the instrument's failures alert the same channel its successes do). Every field is a number or the direction of a completed move; nothing forecasts. Webhook push (Phase 1) follows a four-week observation period per the signed design.

docs/data/forecasts.json

The V11 forecast experiment's public record (validation/forecast_registration.json, founder-signed 2026-08-06). RESEARCH SURFACE ONLY, rendered at research/forecasts.html under its mandatory header: an experiment about the index, not advice, presumed null until the registered criterion resolves. questions[]: weekly mechanical questions (any spike day within 14 days, the registered episode rule), each committed before its window opens, with both arms' probabilities recorded at generation (p_climatology — trailing 104-week frequency; p_salience — frozen-coefficient logistic on two registered inputs, validation/forecast_logit_frozen.json), outcome graded mechanically at window close, voids counted. cumulative: Brier per arm. The fork criterion is frozen in the registration; there is no third outcome.

docs/data/back_extension.json

The historical attention proxy, 1979–2019 (M1; registration frozen before first computation in analysis/back_extension_memo.md). A DIFFERENT construct from the live instrument — monthly shares of global event-mentions by frozen actor-pair filters from GDELT's 1.0 and 2.0 event streams — never spliced into or drawn with the 2017+ series. anchor_grades[]: nine pre-registered episodes graded top-decile-or-miss against trailing-10-year percentiles (Kargil 100.0; Blue Star's 52.3 miss is the registered construct-edge finding). overlap_audit: raw-share correlation vs the live instrument over 36 shared months — border channels track (0.893, 0.848); us_trade and gulf_energy fall under the frozen 0.4 threshold and therefore publish NO historical series (series carries only channels that pass; the refusal is the finding). Shipping is excluded entirely: chokepoint salience has no actor coding.

docs/data/cow_mids.json

Correlates of War MID 5.0 aggregates (S12): distinct militarized interstate disputes involving India by start year and opponent (disputes_by_start_year.pakistan/.china), 1816–2014 — historian-coded conflict as independent context beside the M1 historical attention proxy. Static dataset, fetched once; only aggregates publish with attribution (Palmer et al. 2022, Correlates of War Project). Context only; enters no score.

docs/data/energy_context.json

India's monthly crude-oil imports in thousand barrels per day, as reported to JODI (the IEA/OPEC-backed official oil-data exchange; open annual CSVs, verified endpoints) — official energy statistics as context beside the gulf_energy channel (S3). Trailing 48 months; latest names the latest reported month, because JODI reporting lags one-to-three months and recent values revise. Attribution in _meta; context only; enters no score.

docs/data/shares.csv and shares.json

The quantity, not the transform. Each channel's daily share of the GDELT-monitored article corpus, in percent — the input from which the published percentile series is computed. Promoted to a first-class artifact so a reader who disagrees with the normalization can apply their own and recompute everything. shares.json carries the metadata, including known_gaps: the days absent from the store entirely (21 as of 2026-08-07), listed rather than interpolated, because a user computing monthly means deserves to know which months are short. Shares are commensurable across channels and across years; percentiles are not (they are within-channel ranks against each channel's own trailing two years), and percentiles saturate during sustained crises where shares do not — see methodology v1.9.0. The source column names the instrument that measured each day: doc_api for GDELT DOC API counts (3,446 days) and ngram_bridge for days measured by the Web NGrams feed and divided by an estimated per-channel splice ratio to reach the API's scale (38 days, 2026-06-30 onward). These are two instruments, not one series. Any statistic computed across the boundary inherits the linking constant's uncertainty, so the label travels in the data rather than in a footnote; shares.json._meta.instrument_is_not_constant carries the splice ratios, the recorded-versus-reconstructed basis of each label, and the bound on reconstruction impurity.

docs/data/syndication.json

Was that spike more newsrooms, or more copies? Per channel per day, articles divided by distinct normalized headlines among them — the multiplier by which republication inflates a channel's article count. The index counts articles, so one wire story carried by two hundred outlets moves a channel exactly as far as two hundred newsrooms independently reacting; this is the measurement of how often that happens. On 2026-08-05, pakistan_west ran 160 articles across 46 distinct stories (×3.478) while china_east ran 17 articles across 17 stories (×1.000). It corrects nothing — republication is a real editorial decision, not noise, and no defensible deflation exists; nothing here enters a score. The estimator is biased downward at small samples, because a duplicate is visible only when both copies land in the scan: the same channel-day reads 1.923 at 28,575 documents sampled and 3.478 at 119,349, so only scans of at least 60,000 documents enter the history and n_docs_sampled travels with every row. A companion and deliberately non-identical measure, articles_sum / sources_sum, is computed over the whole GDELT event population in data/raw/events_daily.csv; the two are reported separately and never averaged, because dividing by sources also counts one outlet running several distinct pieces on the same event.

docs/data/freshness.json

Is every number on this site actually current? For every non-exempt JSON payload, the audit publishes the write date (generated, with additive aliases written_date/write_age_days) and any explicitly declared measurement day (measured_date/measured_age_days). A file rewritten today can still describe yesterday; those clocks are never treated as synonyms. Dated CSVs carry their last measured row but no embedded write timestamp, recorded as write_time_status: not_embedded_in_csv. Every exemption remains listed with its reason. The write-time age is checked against the cadence its lane is supposed to run at. A lane that stops writing does not break — it keeps serving its last value, and a stale number is quoted exactly as confidently as a fresh one. status.json watches upstream sources; this watches every payload enumerated in freshness.json, which is where the silence actually lives.

A payload with no _meta.generated counts as a failure, not a pass. Three payloads (alt_specs, seasonality, priced_risk) had no timestamp at all when this was built, and validation.json — the credibility payload — had none either, because it is assembled block by block and no block owned the file. All four now carry one. An auditor that silently skips what it cannot read is worse than no auditor, because it reports green.

Exemptions are listed with a written reason each: the frozen API contract, the author-paced notes, the append-only Prediction Archive, episodes.json (a bare array with nowhere to put _meta, written in the same commit as history.json), and this file itself. The audit fails the enrichment run when anything is stale or undatable — never the 06:00 contract lane, because a staleness report must not be able to stop a publish. A valid but older measurement day is not itself a crash; status.json.alignments[] publishes whether it can safely join to the current reference day and what the affected reader surface does instead.

docs/data/vintages.json

What did the index say on the day it said it? Every daily publish is a commit, so docs/data/history.csv at a given SHA is exactly what the site served that morning. This file diffs every such vintage against what is served today, across the composite and all five channels on their overlapping dates — the question that decides whether a backtest built on the current file carries look-ahead bias.

Finding: 24 of 26 vintages show zero changed values. The series is append-only in practice — days are added, not rewritten. The single revision episode is the two 2026-07-27 vintages (101 values across 17 days, max 99.4 percentile points), and it sits exactly on methodology v1.0.1 of 2026-07-29, which changed the percentile computation and is recorded in the changelog. A revision with a dated published cause is a version; a revision without one is a defect.

It is also a tripwire, which is the real reason it runs nightly. Any vintage that differs without its date appearing in REGISTERED_REVISIONS fails the build, so a future silent rewrite — a rebuilt store, a dictionary applied retroactively, a splice ratio quietly adopted — surfaces the next morning instead of being found by a user whose results stopped replicating.

Retrieve any vintage yourself, with no tooling and no cooperation from this project:

git log --format='%H %ad' --date=short -- docs/data/history.csv
git show <SHA>:docs/data/history.csv > history_as_of_that_day.csv

docs/data/detector_blindness.json

The detector cannot see a second crisis for about three months. Episode detection fires when a channel's share exceeds its trailing 90-day mean plus 2σ. A crisis therefore enters the very window its own future threshold is computed from: the mean rises, the standard deviation rises faster, and the bar climbs for ninety days — peaking after the crisis is over in 521 of 523 analysed episodes. The payload's excluded_episodes lists any detected episode the diagnostic cannot evaluate and gives the reason; the current one-row exclusion is caused by an absent pre-episode source day rather than being silently dropped.

Worked example, pakistan_west through Pahalgam and Sindoor: the threshold was 0.0244 on 2025-04-21, and 0.6912 on 2025-07-20 — a 28× higher bar. On 2025-05-20 the channel's share was 0.1340, five and a half times the pre-crisis threshold and an unambiguous spike by the standard that applied three weeks earlier, and the detector stayed silent.

n_days_suppressed counts days after an episode whose share cleared the threshold in force the day before that episode began but not the live one — days a pre-crisis detector would have flagged and this one did not. Across the whole series that is 2,158 distinct channel-days, 12.4% of all channel-days, deduplicated because episode windows overlap (summing per-episode counts would exceed the length of the series). Read largest_episodes_by_peak_share, not the median: small blips barely move the bar, so the median multiple is ~1.4, and the damage is entirely in the tail — the detector is least able to see a second event exactly when the first was most serious.

The registered rule is not changed, because every obvious fix is worse: excluding spike days from the baseline makes the analyst define spikes before measuring them, a fixed threshold abandons the per-channel adaptivity that makes five channels comparable, and a longer window delays first detection, which is what the detector is for. The cost is now measured rather than inferred.

docs/data/wiki_hindi.json

Does this measure Indian attention, or Anglophone attention about India? The index is built from an English corpus and was cross-validated against English Wikipedia, so every leg of the apparatus was Anglophone and none of it could tell those two apart. This file is the test that can: the same registered article set read on Hindi Wikipedia — titles resolved through Wikipedia's own interlanguage links, never chosen by hand — correlated against the same daily shares over 3,394 overlapping days. english_wikipedia_same_statistic carries the identical computation in English, because the Hindi numbers are uninterpretable without it.

Finding: English leads on 5 of 5 channels, in both levels and changes, with no channel where Hindi leads. Day-to-day changes 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). us_trade is the study's only negative level correlation (−0.136). The thin-traffic objection fails on the project's own data: us_trade has the highest median Hindi traffic and the worst agreement, gulf_energy the lowest and the best.

Six registered articles have no Hindi counterpart (missing_english_titles) and are reported, never substituted; they are the foreign-policy-apparatus topics, while the two India-adjacent channels resolve completely. changes_pearson is the load-bearing figure — levels can correlate through a shared trend alone. Undefined correlations publish as null rather than NaN. Nothing here adjusts a score; it is evidence about what the index measures, and anyone citing this as a measure of Indian public salience should read it first.

docs/data/monthly.csv and monthly.json

The series at the frequency economists actually work in. One row per calendar month: share_<channel> are means of the daily percent-of-corpus shares, composite_mean_of_ranks is the mean of the daily composite percentiles. It ships because twenty users aggregating this themselves would make twenty different silent decisions about the same three traps, and their results would stop being comparable.

Short months. coverage is share-data days present over days expected (for the current month, over days elapsed). A month below 80% publishes its row with null values and a stated refused_reason rather than a mean over the survivors — 2025-06 has 14 of 30 days and is the case that forces the rule. The row is never silently absent: a gap you can see is a limitation, a gap you cannot see is an error.

Mixed instruments. mixed_instruments, with n_days_doc_api and n_days_ngram_bridge, marks months spanning the DOC API/NGrams-bridge boundary, where the mean crosses two instruments joined by an estimated splice ratio. 2026-06 is the first such month. Users who need one instrument can drop these rather than discover them.

Averaging ranks. composite_mean_of_ranks is the mean of daily percentiles, which is not the percentile of the monthly mean; the two diverge most in skewed months, which are the months of interest. Both are published so the difference is visible rather than resolved by whoever aggregates first. The share_* columns mean the same thing in every year and are the recommended input to anything estimated; percentiles are ranks against a moving two-year window and are offered as convenience. n_days_composite_present and composite_coverage separately report how many ranked daily composite scores exist; a monthly composite below the same 80% floor is null with a composite_refused_reason. JSON uses null for missing numeric values; CSV represents the same values as blank cells.

docs/data/episode_themes.json

Narrative composition (G4): per channel per day, the GKG theme codes carried by the day's MATCHED articles — the registered dictionaries choose every article and BigQuery only reads its V2Themes, so themes are an annotation, never a filter, and enter no score. top_themes[] gives each code with the share of themed matched articles carrying it; codes stay raw GDELT identifiers for auditability. n_themed vs n_matched discloses the join rate. Trailing 30 days.

docs/data/historical_intelligence.json

Registered readings of the 1979–2019 back-extension (back_extension.json), and of nothing else. The archive is a different construct from the live 2017+ index and is never spliced to it: it counts actor-pair event mentions, not press salience, so no statistic here is comparable to a published IGRM score. Eligibility is inherited from the source payload's overlap_audit verdicts, not re-decided here: pakistan_west (r = 0.893 over 36 overlap months) and china_east (r = 0.848) are marked tracks; us_trade (r = 0.216) and gulf_energy (r = 0.153) carry the verdict DOES NOT PUBLISH (negative finding), and shipping was excluded at registration for having no defensible historical event analog for its construct. Each refusal is published in channel_eligibility.refused with its reason rather than omitted. Frozen: the archive ends 2019-12, and the payload is exempt from the freshness clock for that reason, not because it is unmaintained.

regime_baselines.rows[] — mean, median, p90, min and max per channel, series and registered period. Periods are the four calendar decades 1979-1989, 1990-1999, 2000-2009 and 2010-2019 and are explicitly not political regimes; the word names a slice of the calendar. Every row carries its own denominators: n_months_in_period (calendar count), n_observed (non-null count) and coverage_fraction = n_observed / n_months_in_period. A period with fewer than min_observed_months (24) observations is refused — available: false with unavailable_reason — never computed over whatever happened to be there. regime_baselines.non_comparability carries three standing warnings, including that pctl_10y is undefined for the archive's first 120 months and that cross-channel comparison of raw_share is not meaningful, each channel's share being relative to its own filter rather than a shared denominator.

structural_breaks.rows[] — a descriptive scan, not a change-point model. Every split with at least min_segment_months on both sides is scored by the absolute Welch t between segment means, and the largest is reported as the single candidate. The null is a seeded permutation (n_permutations 2000, seed 20260809, deterministic): values are shuffled without replacement, which destroys time ordering while preserving the value distribution, and p_value is the fraction of permutations whose maximum statistic met or exceeded the observed one. sensitivity_sweep[] repeats the scan at minimum segment lengths 12, 24, 36 and 48; stable_across_all_settings and distinct_candidates_across_settings report whether the answer survives the setting. It frequently does not — pakistan_west returns 1980-01 and 1984-01 across the sweep and is published as unstable. structural_breaks.language_rule binds the prose: a result is a candidate break in the measured series, never a historical cause, a turning point, a regime change, or an explanation of any event. The series measures attention, so a break is a break in attention and nothing more.

analog_retrieval.by_channel.<ch>.<YYYY-MM> — the nearest other months in the same channel by Euclidean distance over standardised registered features (level_pctl_10y, change_12m_raw_share, volatility_12m_raw_share), excluding a ±exclusion_window_months (6) neighbourhood of the query so adjacency is not mistaken for similarity. Missingness is named, never filled: a feature null on either side of a pair is dropped from that pair's distance, reported in features_excluded_as_null, and counted in n_features_used; a pair sharing fewer than min_features_for_match (2) usable features is refused rather than distance-imputed. No null is ever replaced by zero or by a mean. Retrieval is a similarity lookup over one measured series: it carries no claim that similar months share causes, and no claim about what follows them.

event_archetypes.rows[] — an archetype is attached to a month only where a human-authored registered anchor exists in the source payload's anchor_grades. machine_generated_permitted is false; a month with no registered anchor reports available: false with reason no_registered_human_authored_archetype and is never given an inferred one. A knowledge-cutoff rule (_meta.knowledge_cutoff, archive_end 2019-12) refuses any label, annotation or category whose evidence postdates the archive, and publishes the refusal rather than dropping it silently, so present-day categories cannot leak backwards into a historical view.

_meta carries contract_sha256, source_sha256 and implementation_sha256, so a citation can pin the registration, the input and the code that produced the numbers. Regeneration is byte-identical and CI asserts it. Rendered for humans at docs/history-lab.html; flat downloads at downloads/igrm-historical-regime-baselines.csv and downloads/igrm-historical-analogs.csv. Registered in governance/historical_intelligence_contract.json. Nothing in this payload is a forecast, and no field states or implies what happens next.