Survey Quality Rubric & Linter India: 5 Checks Plus SD-032 to 044 Integrity

Check 1 — Accessibility, ethics and neutrality. The rubric's first lens is whether the survey treats respondents respectfully and avoids bias. Ethics review (services/ai/review_bot.py run_ethics_review) blocks HIGH-severity patterns: solicitation of PII beyond structured contact forms (Aadhaar, PAN, UPI, bank account), caste- or religion-based questions unless explicitly requested and the user asked, explicit/racist content, unsafe media, and missing required stimuli for SD-036→039 concept/pack/shelf/video studies. Neutrality checks catch leading phrasing flagged by INDIA_SURVEY_DESIGN_RULES — "Don't you think…", courtesy-inviting statements, unbalanced batteries that invite acquiescence — and require rewriting to fully labeled, balanced forms. Accessibility covers mobile-first readability on the SuperJ app: font scaling, image alt text for concept cards, video forced-exposure timers (SD-039) rather than auto-advance, and keyboardType correctness for text inputs.

Check 2 — Scale balance and acquiescence correction. Indian respondents show elevated acquiescence and positive-end scale use, a pattern documented in INDIA_SURVEY_DESIGN_RULES and corrected here. The rubric scores whether agree/disagree batteries have been replaced by fully labeled 5-point unipolar scales, whether semantic differentials (SD-018) randomise statement order via Block/Random rotation (SD-009→011) and balance positive/negative statements, and whether long batteries (12+ attributes) avoid rating grids that invite SD-034 straight-lining in favour of MaxDiff (SD-015) for 8–30 items or simple Ranking only for 3–7 items (simple ranking beyond 7 is flagged as unreliable). On Hercules Works in Hyderabad, this correction is verified via the Survey Knowledge Graph and DuckDB scoring, not assumption, keeping approval rates honest across NCCS and tier.

Check 3 — Keyboard, option-count and device reality. The rubric validates that keyboardType is correctly set for text inputs — numeric keyboards where needed with sane maxChar caps enforced in parsers.py (fallback sets keyboardType:"numeric" when a text question asks for a number and sanity-caps digits), physical keyboards elsewhere — so respondents on the SuperJ app for rewards do not abandon due to wrong input chrome. Option-count discipline matters: every MCQ is checked for MECE (mutually exclusive, collectively exhaustive) with "Other" (is_explainable:true) and "None of the above" (deselect_all:true) where appropriate; long lists (>8) without shuffle:true are flagged for primacy bias; ordered scales never receive shuffle. Device reality extends to break screens (1–3 survey_break cards for 21–30→1, 31–45→2, 46+→3 questions) at thematic transitions, image_multiple_choice sizing for thumb taps, and forced exposure timers for stimuli rather than skippable embeds.

Checks 4 and 5 — India language/phrasing and routing hygiene. Check 4 applies INDIA_SURVEY_DESIGN_RULES to language: simple direct English under 15 words for Tier-2/3 skews, free of Western idioms, with Indian brand/currency/festival/category references and Indic media support (Hindi and other Indic images/PDFs via PyMuPDF/docx/pptx/openpyxl/Pillow) accepted without breaking routing. Check 5 scores routing hygiene after route_compiler compilation: 8 top-level fields (next/showIf/showIfElse/showIfLogic/requiredOptions/filterFrom/filterOut/shuffle) must be declared correctly, branch_blocks must have showIf protection, loops via Loop & Merge (SD-024) must have stable rejoin, dynamic_lists (SD-025) must have exact option parity to their source, and validation must pass validate_compiled_routes() with repair→fallback→strip ladder. The scorecard surfaces these as severity-weighted findings, not binary pass/fail, so authors can prioritise fixes (e.g., a missing showIfElse on an SD-030 screener is higher severity than a missing "Other" on a brand list). When the scorecard passes, the survey is linter-clean for deployment; when it warns, the author fixes in minutes rather than after 300 polluted completes — and Poseidon later reads the same cleaned questions_json from survey_version.

What SD-032 requires and why it works in India. SD-032Attention Check Questions are embedded items with known correct answers — instructional manipulation checks ("Select Strongly Agree for this statement regardless of your opinion") or directed counts ("How many times did the video mention the price? Select 2"). Framework_designs.md defines auto-detection triggers ("attention check", "trap question", "honesty check") and primary use as mandatory for any online survey above 10 minutes LOI and for high-stakes concept tests; implementation labels SD-032 as Low complexity / P0 critical because it appears in virtually every consumer study. On the SuperJ app — 20M+ ZK-verified Indians rewarded for answers — attention checks are not adversarial; they are brief, plainly worded, and placed after the Hook and mid-Core, never inside a scaled battery where they could distort straight-lining detection. Failures are flagged for respondent review or removal, and the flag propagates as routing-aware metadata so Poseidon's null-correlation Survey Intelligence knows to adjust denominators if flagged respondents are excluded from a conditional insight — so "62% among attentive theatre-goers" cites the correct base.

How Hercules authors and scores them. The question-generation prompts quota at least one attention check for LOI above thresholds, and compute_survey_quality_scorecard() scores presence, correctness of the designated correct answer, and placement. Admin review tolerates SD-032 presence as a quality signal; its absence on a 14-minute sequential monadic with 3 stimuli triggers a MEDIUM-to-HIGH warning in the scorecard before the instrument can go live. This is traceable to Survey-Creator/docs/framework_designs.md SD-032 and analytics_process.md §5/7 trust passages, not invented thresholds, and it pairs directly with the next SD layers (speeder/straight-lining/open-end/ZK) that together make Poseidon's 99.1% three-layer verification meaningful — verification proves numbers were computed correctly; attention checks prove the respondents were paying attention. If you run brand tracking or product testing, you will see an attention check suggested automatically — accept it; the cost of one added question is dwarfed by the cost of defending unverifiable data to leadership. Pricing is Free ₹0, Starter ₹1,119 (₹895 annual), Pro ₹30,000/quarter (₹24,000 annual) with 20% annual off and 100 free first-month responses.

Placement and flag propagation on SuperJ with scorecard gating. On Hercules Works built in Hyderabad, SD-032 checks are brief, plainly worded, placed after Hook and mid-Core, never inside scaled batteries, and failures propagate as routing-aware metadata so Poseidon adjusts denominators if flagged respondents are excluded — keeping "62% among attentive theatre-goers" precisely cited with the correct base. This traceability to framework_designs.md SD-032 and analytics_process.md trust layers, plus the quality scorecard's pre-field gating, is why attention failures never silently pollute a report. Framework trace: SD-032 Low/P0 in framework_designs.md; this is why attention failures never silently pollute a report on Hercules Works in Hyderabad.

SD-033 Speeder Detection — the 40–50% median rule. SD-033 flags respondents who complete the survey below a defined minimum threshold of median length-of-interview (LOI), typically 40–50% of median completion time for that instrument length and stimuli burden. Framework_designs.md labels it Medium complexity / P1 (all surveys above 8 minutes LOI) with triggers "speeder", "minimum time", "LOI threshold", and Poseidon's report pipeline consumes it as part of Data Quality & Respondent Validation sections in the 18-node graph, displaying speeder rate and exclusion count when exclusions are applied. On the SuperJ app, where surveys pay rewards, speeder risk is real but distinct from attention risk: speeders may pass an attention check by tapping quickly and still fail speeder bounds, so Hercules arms both.

SD-033 Speeder Detection — the 40–50% median rule. compute_survey_quality_scorecard() verifies that speeder rules are declared for eligible LOIs and that median estimation will have a burn-in window before flagging, rather than a fixed seconds threshold that would misfire on short Tier-2/3 instruments versus long Tier-1 concept batteries. When a flagged respondent is excluded, the pipeline re-runs verification (deterministic claim verifier at 0.5 tolerance, structural validator sum-to-100, self-critique) on the filtered Parquet so narrative claims reflect the cleaned base, not the raw n. This traceability is why the rate is cited as 40–50% rather than a single number — the rubric applies the instrument's actual median, not a guess.

SD-034 Straight-lining / Pattern Detection — grids are guilty until proven balanced. SD-034 detects respondents who select the same scale position across all items in a grid (straight-lining) or show repetitive patterns (e.g., 1-2-1-2). It is Medium complexity / P1 for any survey with rating grids or large attribute batteries, with triggers "straight-lining" and "pattern detection." The rubric scores whether the instrument invites straight-lining in the first place: grids with 12+ rows and a single scale are flagged to switch to MaxDiff (SD-015) for 8–30 items or simple Ranking for 3–7, with semantic differentials (SD-018) randomised and balanced instead. On the SuperJ app mobile flow where max 5 rows are visible at once, straight-lining is rarer than on desktop grids, but the detector still runs within the 18-node report's Open-Ended Intelligence and Data Analysis parallel nodes and surfaces a straight-liner rate (e.g., "3.2% of grid respondents flagged"). The correct essay link for battery design is conjoint analysis India when batteries become trade-offs and maxdiff survey tool when they become hierarchies; the linter will suggest them automatically.

How both feed Poseidon's denominator correction. Neither SD-033 nor SD-034 is a cosmetic flag. When a speeder or straight-liner is excluded from analysis, the Survey Knowledge Graph's routed_from edges and the filtered Parquet together re-derive denominator-aware frequencies, crosstabs and Borda/MavDiff scores, and verification re-runs before narrative drafting — so "top-2-box 41% (n=266 of 320 attentive completes)" is precisely reproducible via DuckDB SQL. This is the handshake between authoring quality (SD-032→034 armed before fielding) and analytic integrity (Poseidon 5-phase with Survey Knowledge Graph traversal) that makes the "trusted by…" claim auditable. On Hyderabad-built Hercules Works, respondents remain ZK-verified throughout, so exclusions reflect quality, not duplication.

What SD-035 scores and how Hercules automates it. SD-035 Open-End Quality Scoring is labelled High complexity / P1 for all surveys with mandatory open-ends, triggered by "open end quality", "gibberish detection", "verbatim check". Traditional checks count words or characters; Hercules routes verbatim responses through AI-based scoring that detects gibberish ("asdf", repeated characters), copy-paste (identical responses across questions or near-duplicate Levenshtein), single-character entries ("a", "."), and off-topic responses (e.g., answering a biscuit likeability open-end with "I don't know" or an unrelated phrase). This runs in Poseidon's Open-Ended Intelligence node, which is no-op if the survey has no text questions and active otherwise: it codes verbatims into themes, clusters paraphrases, and flags low-quality open-ends for review with per-respondent flags rather than survey-level rejection.

What SD-035 scores and how Hercules automates it. compute_survey_quality_scorecard() scores whether the instrument invites this quality in the first place: text questions are placed at inflection points (after a rating, at end of each thematic block, always final), have sane maxChar aligned with keyboardType (numeric caps for numeric asks, longer for reflective asks), and are never double-stacked without a behavioural anchor between them. The rubric also checks for piping hygiene — piped text (SD-023) that inserts prior open-end content into a later question is validated so truncation does not break narrative in the SuperJ app rendering. The result is not "open-ends exist" but "open-ends are usable" — theme extraction in 8+ Indian languages (Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati, Kannada, Malayalam, Punjabi) depends on verbatim quality, and SD-035 is the gate.

Why SD-035 matters for India — Hinglish, transliteration and brevity are not gibberish. A naïve gibberish detector would flag "mast hai, paisa vasool" or "Tanglish" as low quality. Hercules avoids that by using India-trained normalization that treats Hinglish code-switching and transliterated verbatims as signal, not noise: "price thoda zyada hai but quality mast hai" is coded as a durability-value trade-off rather than flagged. Single-word answers of substance ("crunchy", "dulhana", "₹99") are retained when they match the question's dictionary, while repeated "good good good" across batteries is flagged even if English-valid, because repetition is the tell rather than language. This matters for SuperJ fielding across Tier 1/2/3 where brevity is common on mobile; the scorer distinguishes between a concise value judgment and gibberish by comparing response content to the theme dictionary built from the survey's own coded battery. When low-quality rates are high (e.g., >10% of open-ends flagged), the scorecard recommends trimming one battery or moving the open-end earlier in CLOSE before fatigue, a remedy that respects the 8–12 minute burden rather than simply collecting more text. Explore adjacent depth at open ended survey analysis India and natural language survey analytics India.

SD-044 Panel Deduplication — ZK proof that each voice is one human. SD-044 is labelled High complexity / P0 critical and mandatory on SuperJ, triggered by "deduplication", "unique panel", "exclusion window". SuperJ's identity is Zero-Knowledge-Proof (ZKP) verified: cryptography proves each respondent is a unique real human without exposing personal data, yielding zero bots and cross-survey deduplication within a configurable exclusion window (e.g., no repeat of the same study or conflicting studies). This runs platform-side as part of distribution on the SuperJ app — 20M+ verified Indians, 60-90%+ completion for rewards — not as a questionnaire question, and it is the reason Hercules can cite "real humans only" without qualifying. compute_survey_quality_scorecard() validates that the survey declares its exclusion window and that audience interlocks (e.g., interlocking quotas SD-027 on city×gender×age) do not undermine deduplication by over-targeting the same cell repeatedly. When Poseidon ingests via POST /api/v1/ingest/ (X-Secret-Key) it fetches respondent answer JSON from SuperJ keyed by super_j_survey_id; the Parquet is written with analytics_row_count that already reflects ZK-unique respondents, and sample weighting or post-stratification can be layered deterministically. No Western panel rent, no professional takers — deduplication is infrastructural, not declared.

compute_survey_quality_scorecard as the living gate — before field, after analysis, always. The scorecard is not a one-time badge before deploy. It runs at creation (local quality + domain + public-interest linting), persists per version in survey_versions, blocks or warns on deploy via admin review policy (SAFE/REVIEW_RECOMMENDED/UNSAFE), and is re-consulted by Poseidon's Quality Checker in the 5-phase analysis pipeline (phase 3 of 5: Data Profiler, Column Selector, Quality Checker) to flag completeness, outlier risk, sample-size and attention failures that affect interpretation. That continuity is why engineering cites three independent layers elsewhere in the analytics stack: the scorecard catches design defects, while Poseidon's Verification (Numerical Claim Verifier at 0.5 tolerance via fresh DuckDB SQL, Output Validator with sum-to-100 and NPS bounds, Self-Critique at 99.1%/97.4%/94.8%) catches computation defects — together they are the reason no hallucinated % reaches a chart. The scorecard also enforces the three-phase Hook→Core→Close (20/60/20) and anti-defaults (no sequential monadic default, no A/B without stimuli, no conjoint for <5), so the Data Quality & Respondent Validation section of the 18-node report can literally say "attention pass rate 96%, speeder 2.8%, straight-line 1.1%, ZK dedup window 60d, N=320 attentive completes" and mean it.

What to ship tomorrow — a checklist the linter already runs. Before you next hit deploy: ensure at least one attention check on any ≥10-minute instrument; confirm speeder and straight-lining logic are armed for eligible LOIs/batteries (the linter will flag absence); move any 20-item rating grid to a MaxDiff (8–30) or Turk-style TURF where portfolio matters; place open-ends at inflection points with correct keyboardType and maxChar; keep the 40–50% median speeder rule literal rather than a fixed seconds guess; and state an exclusion window for SD-044 rather than assuming deduplication. Each recommended action from the scorecard maps to a specific SD and a Poseidon behaviour, so when you field on the SuperJ app from Hyderabad at Free/ Starter/Pro (annual 20% off, 100 free first-month responses) the pipeline you booked ends with a report that cites bases, flags caveats, and recommends next steps — the kind of rigour that turns data into decisions within 48–72 hours instead of six weeks.