CID: Amnesty International Annual Report — India entries, longitudinal series

Document: Amnesty International Report (annual), India country entry — sliced from the global yearbook Organization: Amnesty International CID ID: CID-0045 (provisional; longitudinal series) Document type: TYPE 6, Advocacy Document (TYPE 7 alternative noted below — identical score) Rubric version: v0.3.2 Scored: 2026-05-29 Series span: 1973/74 – 2024/25, 49 substantive India entries (25 in Era A, 14 in Era B, 10 in Era C)

---

Verdict

Every substantive India entry in the Amnesty annual-report series scores Deficient. The methodology that produces these entries does not change across fifty years: there is no published codebook, no sampling frame, no inter-coder reliability, and no per-event verification pathway inside the document. What changes is the events described and, in the modern era, the directional emphasis of who is named as victim. None of that touches the methodology score. The series is a textbook case of structural invariance — the same finding the CID has documented for USCIRF and Freedom House, now confirmed for a third major human-rights institution.

The score sits in the lower third of the Deficient band (4.60–4.87) and is stable across all three weighting schemes. This is a firmer placement than the Freedom House India chapter (5.93, borderline Deficient/Adequate): Amnesty’s yearbook entry has no companion methodology document and no scored framework to lift its definitional precision or verification scores.

---

Type classification rationale

Amnesty International is a campaigning organization. Its annual report opens each country entry with a normative human-rights assessment and selects the cases it documents to advance that assessment. That is the defining marker of TYPE 6 (Advocacy Document): a normative conclusion stated up front, with empirical claims chosen to support it. The India entry is not a think-tank synthesis of existing literature (TYPE 7 Policy Report), nor a rolling incident dataset (TYPE 2), nor an examination of a named organization (TYPE 3). It is an advocacy organization’s narrative country assessment built on its own field documentation.

D2 (Classification Rigor) and D3 (Case Capture & Sampling) are N/A for TYPE 6. The excluded 33% redistributes across D1, D4, D5, D6, D7, D8, giving effective weights of 17.9 / 22.4 / 14.9 / 26.9 / 7.5 / 10.4.

TYPE 7 alternative. If classified as a Policy Report (TYPE 7) — the classification used for the Freedom House India chapter — the effective weights are identical and only the D8 applicability label changes (Full for TYPE 7, Adapted for TYPE 6). At the D8 score assigned here (3), the band is the same under either reading. The grade does not move. The TYPE 6 classification is retained as the more honest description of what an Amnesty yearbook entry is. Anang owns the final pre-registration type assignment.

---

The three structural eras

The longitudinal comparison identifies two methodology-relevant inflection points and therefore three eras. The inflections move only D6 (Verification Standards); every other dimension is invariant across the full series.

Era	Years	Entries	Format	In-text sourcing	D6
A — Narrative	1973/74 – 1999	25	Prisoner-of-conscience prose; no thematic template	None (0 AI-Index refs, 0 URLs)	3
B — Structured + sourced	2000 – 2013	14	”Head of state / Background / Impunity / Discrimination / …” thematic template	”AI country reports and visits” lists with AI-Index numbers (4–12 refs/entry)	4
C — Modern	2014/15 – 2024/25	10	Same thematic template	In-text AI-Index refs gone (moved to footnotes outside the captured text)	3*

* Era C D6 is scored as presented in the available slice text (assertion-level, no captured citations). Modern Amnesty reports do carry endnote sourcing; the text extraction did not capture those footnotes. If captured, Era C D6 would likely rise to 4, matching Era B. This caveat is flagged rather than silently resolved.

---

Dimension scores

Scores are invariant across the series except D6. Each is grounded in the per-year analysis, not in general knowledge of Amnesty International.

D1: Definitional precision — 4 / 10

Effective weight: 17.9%. Invariant across series.

Amnesty applies institutionally-defined terms of art — “prisoner of conscience,” “extrajudicial execution,” “arbitrary detention,” “torture” — that carry established Amnesty meanings. But the yearbook entry republishes no codebook, no operational definitions, and no borderline-case guidance. The document review finds a definitions/glossary section in 0 of 49 entries. Terms are applied editorially inside the entry; a reader cannot reconstruct the inclusion rule that put a given case in the report. This is the band-4 reading: terms used without in-document operational specificity, reliance on an implicit “we know it when we see it” standard. The score sits two points below the Freedom House India chapter (6), which is backed by a published 25-question framework and a 4,900–7,500-word companion methodology document. Amnesty International’s entry has no such companion document.

D2: Classification rigor — N/A

Not applicable for TYPE 6 Advocacy Documents.

D3: Case capture and sampling — N/A

Not applicable for TYPE 6 Advocacy Documents.

D4: Coverage symmetry — 6 / 10

Effective weight: 22.4%. Invariant across series.

The parent document is titled The State of the World’s Human Rights and genuinely walks every country A–Z, so the cross-national coverage is broad, not particularist. Within India, Amnesty’s evaluative criteria — torture, arbitrary detention, unlawful killing, unfair trial — are identity-neutral and applied to both state actors and armed/non-state groups; entries across the series carry explicit “Abuses by armed groups” sections. The criteria pass the Swap Test: removing identity markers does not change how they function.

The directional counts shift over time (Muslim-as-target: 6 in 2022/23, 7 in 2023/24, 5 in 2024/25; Dalit-as-target: 3–4 from 2014/15). This reflects the events Amnesty chose to foreground, not a change in the criteria.

On reading the directionality numbers. These counts are a target-verb co-occurrence proxy — how often a group’s name falls within ~200 characters of a word like attacked, targeted, or persecuted. They are swap-test prep, not a measure of coverage or importance. A group can be covered heavily while the proxy reads low, because the framing did not place its name beside a target verb. The length-normalized mention counts give the coverage trend the proxy cannot: per 1,000 words, Muslim mentions rise 0.92 (Era A) → 1.65 (Era B) → 2.79 (Era C); Dalit rises 0.05 → 1.10 → 2.07; Sikh runs the opposite way (0.76 in Era A, the Punjab years, fading to ~0.25); Adivasi/tribal peaks in Era B (3.50). The increasing focus on Muslims and Dalits is therefore real in the underlying text, not an artifact of the proxy — but it is a shift in editorial emphasis tracking events, not a change in the scoring criteria.

This holds the score below the 7–9 band: narrative emphasis tracks rights-reducing events, coverage distribution is not benchmarked against base-rate data, and recent entries lean heavily on a single victim-group framing. Score of 6 reflects genuinely neutral criteria offset by un-benchmarked, increasingly directional emphasis.

D5: Source independence — 6 / 10

Effective weight: 14.9%. Invariant across series.

Amnesty is institutionally independent of the Indian government and Indian political parties, and famously refuses government funding for its research — a strong structural-independence signal. No circular advocacy loop with other NGOs is visible at the entry level. The evidence base, however, is Amnesty’s own field documentation, and at the entry level there are essentially no external citations (URLs = 0 throughout; the Era-B AI-Index references are self-citations to Amnesty’s own fuller reports). Under the rubric’s source-type rule, a primary-research institution citing its own documentation is internal documentation rather than circular sourcing, which keeps this out of the 4–6 band. Score of 6 reflects independence from the subject combined with an entirely self-referential evidence base.

D6: Verification standards — 3 / 4 / 3 by era

Effective weight: 26.9% (heaviest dimension). Era-dependent — the only moving score.

Adapted TYPE 6 criterion: citation accuracy replaces dataset replication; statistical claims are checked against their cited sources. The problem is that there are almost no citations to check.

• Era A (1973/74–1999): 3. Factual assertions — numbers detained, killed, tortured — carry zero in-text sourcing. No data access, no reference trail. Claims rest entirely on institutional trust. Band 1–3.
• Era B (2000–2013): 4. Each entry closes with an “AI country reports and visits” list carrying AI-Index numbers (e.g., AI Index: ASA 20/10/2000). This is a consistently-published, documented pathway to Amnesty’s own fuller reports — still internal, but a real reference trail. Lifts the entry to the bottom of band 4.
• Era C (2014/15–2024/25): 3*. The in-text reference lists disappear from the entry; the slice reverts to unsourced assertion. Scored as-presented (see era caveat); would be 4 if modern footnotes were captured.

Under the D6 < 7 rule, Research-Grade is blocked for the entire series regardless of other dimensions — moot here, since no entry approaches 8.0.

D7: Transparency and governance — 7 / 10

Effective weight: 7.5%. Invariant across series.

Amnesty is a large membership organization that publishes audited financial statements and, critically, declines government and corporate funding for its research — funding itself through members and individual donors. Governance is a genuine board structure, not a founder-controlled operation. The entry itself carries no per-report funder disclosure and no in-document data-ethics policy, which holds it below the top band. Score of 7 reflects strong organizational transparency with no entry-level disclosure apparatus.

D8: Counter-evidence — 3 / 10

Effective weight: 10.4%. Invariant across series. (Adapted for TYPE 6: limitations and corrections weighted above counter-evidence engagement.)

No entry across the series contains a limitations section, a corrections policy, or engagement with perspectives that complicate Amnesty’s assessment. The entries do record positive developments where they occur (“in a welcome step, the Supreme Court suspended…”), which keeps the score off the floor — Amnesty is not impervious or conspiratorial about criticism, it simply does no methodological self-examination inside the document. Band 1–3.

---

Weighted totals

Dimension	Score (A / B / C)	Effective weight	Weighted (A)	Weighted (B)
D1	4	17.9%	0.72	0.72
D2	N/A	—	—	—
D3	N/A	—	—	—
D4	6	22.4%	1.34	1.34
D5	6	14.9%	0.89	0.89
D6	3 / 4 / 3	26.9%	0.81	1.08
D7	7	7.5%	0.52	0.52
D8	3	10.4%	0.31	0.31
Total			4.60	4.87

• Era A (1973/74–1999): 4.60 — Deficient
• Era B (2000–2013): 4.87 — Deficient
• Era C (2014/15–2024/25): 4.60 — Deficient (4.87 if modern footnote sourcing were captured)

Non-compensatory checks

• D3 sampling cap: N/A (D3 not applicable for TYPE 6).
• D6 Research-Grade gate: D6 = 3–4, below the 7 threshold. Research-Grade blocked across the entire series. Moot at 4.60–4.87.

---

Sensitivity analysis

Scheme	Era A	Era B	Era C	Grade
Standard (redistributed)	4.60	4.87	4.60	Deficient
Equal weights (6 × 16.7%)	4.83	5.00	4.83	Deficient
Verification-heavy (D6 emphasis)	≈4.60	≈4.87	≈4.60	Deficient

Grade stability: STABLE. All three schemes place every era in the Deficient band. The standard scheme already weights D6 at 26.9%, so the verification-heavy scheme barely differs. Equal weighting lifts Era B to exactly 5.00 — still mid-Deficient, nowhere near the 6.0 Adequate threshold. This is a more stable placement than the Freedom House India chapter, which crossed into Adequate (6.00) under equal weighting. Amnesty’s entries sit low enough in the band that no defensible weighting reaches Adequate.

---

Calibration context

The Amnesty India entry (4.60–4.87) scores below the Freedom House India chapter (5.93) and below the estimated USCIRF India chapters (~5.2–5.4). The ordering is methodologically coherent: Freedom House publishes a 25-question scored framework with a companion methodology document and Tier-1 downloadable scores; USCIRF carries a designation framework; Amnesty’s yearbook entry has none of these at the document level — no codebook, no scored framework, no data download, no per-event sourcing. It scores above the CSOH/India Hate Lab floor (~2.1) and just above SASAC (~4.3).

The Deficient grade is a property of the document as presented, not of Amnesty as an institution. Amnesty’s organizational independence and governance transparency are genuine strengths (D5, D7). The methodological thinness is a feature of the genre: a global annual report cannot carry a codebook and per-event verification for every country it covers. The CID scores the document a reader actually encounters, not the organization’s total research infrastructure — the same standard applied to USCIRF and Freedom House excerpted chapters.

---

Structural invariance note

Across all 49 substantive India entries (1973/74–2024/25), the dimensions that measure methodology — D1, D4 (framework neutrality), D5, D7, D8 — are invariant. The core methodology gap (no codebook, no sampling frame, no inter-coder reliability, no in-document per-event verification, advocacy orientation) does not change across five decades. Only D6 moves, and only by one point, tracking whether Amnesty placed report references inside the entry (Era B) or not (Eras A and C).

Year-to-year variation in word count, quantitative-claim count, and directional emphasis reflects changing events in India and changing editorial emphasis — not methodological reform. This mirrors the CID’s USCIRF finding precisely: score variation across a long institutional series reflects citation-infrastructure and formatting changes, not a change in how the knowledge was produced.

The per-year data (word count, claims, directionality, era assignment, assigned scores) underlies the era assignments and scores above.