Survey Data Verification India: Why Every Number Must Be Re-Derived

Most survey tools do exactly what a junior analyst does: generate an answer in one LLM pass and display it. The LLM reads a table summary, invents arithmetic in natural language, and writes '62% of respondents prefer UPI cashback.' No one checks whether 62% actually equals 247 of 398. In India, where skip logic is common — 'ask Q7 only if Q3 equals bought in last 30 days' — the denominator is frequently wrong without anyone noticing. A Mumbai fintech learned this when their NPS by city tier was averaged naively; Tier 1 had 800 responses, Tier 3 had 90, and a simple mean of city means overstated Tier 3 by 11 points. Poseidon never does that. Built by Jupiter Meta Labs in Bangalore, Hercules Works (hercules.works/ai) treats every number as guilty until proven innocent, and the proof is a fresh DuckDB SQL query that touches the Parquet file directly. That is survey data verification India at its most literal: if a number cannot be re-derived, it does not leave the pipeline.

The business consequence is stark. When Unilever or ICICI Prudential sign off on a concept test that will allocate manufacturing across three plants, they cannot rely on a number that was hallucinated. They need a verification trail. Poseidon's Numerical Claim Verifier extracts every number in the narrative with regex, then fires a separate, LLM-free DuckDB query to recompute it. If the generated 'NPS 42' is actually 38.2 by SQL, the pipeline flags it, retries with a corrected prompt, and only releases the answer when numbers converge within 0.5 tolerance. This is fully deterministic — no LLM judges its own math. For a brand manager in Delhi burned by an agency deck where percentages summed to 113%, that guarantee is ekdum solid. Hercules Works can thus claim 99.1% numerical accuracy versus ground-truth SQL: accuracy is not promised, it is measured per answer and logged.

Verification also fixes the most Indian of problems: multi-select and ranking confusion. A SuperJ survey stores multi-select as 'Feature A;Feature B;Feature C' in one row. Without Poseidon's Survey Knowledge Graph knowing the column is multi_select, a naive COUNT(*) would count respondents, not selections, and percentages collapse. Poseidon expands via UNNEST(string_split(col, ';')) automatically, using the column's declared type, so the denominator is selections, not respondents. Similarly, ranking uses Borda count Σ(n − rank_position), not average rank. These are formula-driven, not prompt-driven choices. For survey data verification India, the verifier recomputes using the same correct formula template — so even if the narrative LLM misremembered Borda, the verifier catches it. Brands from Kochi to Jaipur now treat Hercules answers as pre-audited, moving from fieldwork to decision in hours at Free ₹0/month, Starter ₹1,119/month and Pro ₹30,000/quarter.

The Numerical Claim Verifier is the headline act of survey data verification India because it is the only layer with zero LLM involvement. After the Code Generator assembles DuckDB SQL and executes it, Narrative Synthesis writes 'Digital wallets lead at 47.3% (n=612), ahead of COD at 29.1%.' A regex extracts every numeric claim — 47.3, 612, 29.1 — plus the entity described. For each claim, no verifier LLM is consulted. Instead, Poseidon builds a fresh SQL query from the same column to formula mapping the Survey Knowledge Graph supplied, executes it via a new ephemeral DuckDB connection on the Parquet file, and compares. If absolute difference exceeds 0.5, the answer is rejected and the pipeline retries with the correct number injected. This loop runs up to two cycles before surfacing failure to Prometheus metrics at /health/stats, so no unverified number ever reaches the Hercules chat UI or report.

That determinism is why speed and scale do not erode trust. DuckDB reads Parquet columnarly — for a 50,000-response survey with 40 columns, a two-column aggregation reads under 5% of the file, 10 to 20 times faster than Pandas. Re-computing every number is therefore cheap. In production, the verifier adds only milliseconds for SIMPLE queries that already bypass the LLM via regex templates, and under a second for COMPLEX queries that decompose into parallel sub-tasks. A Bangalore product lead named Vikram once asked 'What is NPS by city tier split by top-ranked feature?' — a COMPLEX query fanning to four sub-analyses. Each sub-result was verified independently before aggregation, so final heatmap cells were all re-derived. Without this, a single sub-task hallucination poisons the whole answer. With it, Hercules shows a confidence score and verification badge, which is why verification on Poseidon is described as 'if it cannot be re-derived, it does not leave.'

Engineers chose regex plus SQL over an LLM judge deliberately. An LLM judging its own arithmetic is like asking a student to grade their paper — optimistic. The Numerical Claim Verifier's SQL templates come from the same formula registry powering Van Westendorp, Gabor-Granger, Borda count, top-2-box and NPS equals (Promoters minus Detractors) divided by Total times 100. The verifier is not approximating; it executes the canonical definition. For a Hyderabad FMCG brand testing price sensitivity across Hindi, Tamil and English, 'too expensive' percentages are recomputed with identical Van Westendorp logic, not paraphrased. Paired with semantic caching at 78% hit rate — once verified, the vector embedding is cached with a schema-hash fingerprint, so repeats return under 200 ms at zero LLM cost — verification is rigorous and economical on every plan from Free ₹0/month with 10 AI chats to Pro at ₹30,000/quarter.

If the Numerical Claim Verifier is the accountant who re-adds the column, the Output Validator is the editor who checks that the story makes sense. This second layer of survey data verification India runs structural, rule-based checks that no number violated its universe. Percentages must sum to approximately 100 within rounding — a frequency adding to 113% is flagged instantly. NPS must be minus 100 to plus 100; 'NPS 142' from a mis-scaled 1-5 rating is rejected. Rating means must lie within declared scale bounds from the Question Node — a likert_5pt mean of 5.8 is impossible. No nulls may appear in primary result columns; if GROUP BY drops Tier 3 because n is too small, the validator surfaces a sample-size warning rather than a blank. These deterministic guards run in milliseconds, catching errors plausible alone but absurd in context.

The validator also understands survey structure via the Survey Knowledge Graph, so it knows when to apply lite mode versus full mode. Lite mode — for low-stakes, descriptive queries like 'how many respondents?' — skips expensive structural checks to save latency, returning sub-second. Full mode — for inferential and market-research computations like ANOVA, chi-square, Lasso or Ridge regression, or Conjoint — runs the complete suite, including denominator correction for routing edges. Consider conditional Q12: 'Why did you churn? (asked only if NPS ≤ 6).' If a user asks 'What are churn reasons?' without specifying base, the validator checks that reported n is the conditional n, not full N. If narrative says 'n equals 400 cited price' but conditional base is 112, validator flags mismatch and injects the correct WHERE clause from the Routing Edge's showIf rule. That auto-fix eliminates Priya's Pune launch error and saves a Lucknow agency a week of Excel.

Crucially, the Output Validator balances speed versus certainty. A SIMPLE query like 'average rating for Q5' hitting the semantic cache at 78% may already be verified and return under 200 ms with lite validation. A COMPLEX query like 'driver analysis of repeat purchase among urban millennials with post-stratification weighting' goes full validation per slice and may trigger a retry loop. Both share the guarantee: structural absurdity never ships. In production, 97.4% pass on first attempt, with retries resolving most of the remaining 2.6% without intervention. That is why Government of Karnataka can publish citizen insights from SuperJ's 20M plus ZK-verified Indians without manual audit, and why an Ahmedabad retail chain can trust a cross-tab of 'purchase channel by city tier' without opening the Parquet file. Verification on Poseidon is default, from Free ₹0/month through Starter ₹1,119/month to Pro ₹30,000/quarter, built on DuckDB-on-Parquet.

The third layer of survey data verification India is the most humanlike: a separate LLM pass that reads the narrative against verified numbers and asks, 'Did we actually earn the right to say that?' The Self-Critique node receives the verified result frame, the narrative paragraph, and the research brief from survey_chats.research_brief, and scores the narrative zero to ten on four axes: unsupported claims, internal contradictions, causal language where only correlation exists, and missing context such as percentages without a base N. If score falls below threshold, the narrative is rewritten with specific revision instructions and re-scored, up to two cycles. For example, data shows 'Tier 2 NPS 38 versus Tier 1 NPS 34, n Tier 2 equals 42,' but draft says 'Tier 2 decisively outperforms Tier 1.' Self-Critique flags causal overreach and small base, rewriting to 'Tier 2 NPS is directionally higher (38 versus 34), but n equals 42, so interpret cautiously and re-test.' That is senior-analyst judgment automated.

This layer also enforces the 'So What' rule that makes Hercules Works reports readable. Every paragraph must answer 'so what does this mean for the business?' with business implication first, then evidence with percent and N, never evidence then implication. Self-Critique checks that structure. BAD: '65% prefer digital platforms.' GOOD: 'Digital platforms are the dominant channel — 65% prefer them with n equals 612, making mobile-first features the highest-ROI investment.' If a paragraph fails, it is sent back to Report Drafting with note: 'Add implication before evidence.' For a Mumbai CMO skimming between meetings, that rule separates data dump from decision. It is also where Poseidon's India training matters: the model knows 'paisa vasool' is implication, not tag, and that 4.2 in Kochi differs from 4.2 in Delhi, so language baselines are checked.

Self-Critique closes the loop because it alone catches narrative errors numbers cannot. Numerical Claim Verifier confirms '47.3% is correct' but cannot say '47.3% proves wallets will win next quarter' is an unsupported forecast. Output Validator ensures 47.3 plus 29.1 plus 23.6 equals 100, but cannot say omitting 'n equals 612' hides the sample is Tier 1 only. Self-Critique catches both. In benchmarks, 94.8% pass on first attempt; the rest are improved, not discarded. Because critique is prompted with the Research Brief as North Star, every fix is framed against the user's original goal, not generic quality. For survey data verification India, final answers are arithmetically correct, structurally sound, and rhetorically honest — a three-layer moat built by Jupiter Meta Labs in Bangalore, powered by Gemini via LangGraph, delivered through hercules.works/ai with streaming SSE that survives refresh.

Verification sounds expensive until you see how Poseidon makes it cheap. For survey data verification India at India scale — thousands of D2C brands asking daily — every redundant LLM call is a leak. That is why Poseidon has lite mode. For SIMPLE queries — single column, single formula, no routing, caught by regex like 'how many respondents' or 'what is average' — the pipeline skips the LLM, uses a rule-based SQL template, executes via DuckDB, verifies via Numerical Claim Verifier, and returns via semantic cache under 200 ms. No Gemini call, zero cost. For MODERATE queries — one or two columns, one formula — it uses one code-generation call and lite Output Validator. For COMPLEX — multiple questions, demographic cuts, conditional logic, joins — it goes full verification with parallel sub-agents and both validators plus Self-Critique. Routing is roughly 78% cache hit, 11% SIMPLE, 7% MODERATE, 4% COMPLEX, so most users rarely pay for verification they don't need, yet no query escapes re-compute.

Retry loops are similarly thrifty. When Numerical Claim Verifier finds a 0.5 plus mismatch, it does not rerun the five-phase pipeline. LangGraph retries only the failed node — typically Code Generator or Narrative Synthesis — with correct number injected and instruction: 'Prior said 47.3%; SQL says 41.9%; rewrite with 41.9% and adjust implication.' That node-level retry saves tokens and time. Same for Section Audit to Redraft and Safety Check to Redraft in the 18-node report pipeline: only the offending section is rewritten, up to two cycles, before assembling Markdown to HTML to PDF. For a Pro user at ₹30,000 per quarter (₹24,000 annual) running six to eight reports monthly, saved tokens are real money. With 78% semantic cache, similar questions like 'Show NPS' and 'What is NPS score?' share the same verified embedding, so the second returns instantly at zero cost. Verification is included on every plan, including Free perpetual.

Consider the alternative: manual verification. A Pune agency once billed ₹1.2 lakhs to audit a 40-question tracker because every cross-tab was recomputed in Excel by a second analyst. On Hercules Works, that audit happens in seconds, every time, for every answer, without a second analyst. SuperJ's 20M plus verified Indians provide raw truth via WhatsApp at 60 to 90% rates, Parquet and DuckDB provide speed at 10 to 20 times faster than Pandas by reading only required columns, and three-layer verification provides trust. The result is survey data verification India fast enough for Rahul in Hyderabad to ask five follow-ups in one sitting, cheap enough for a bootstrapped brand in Indore to stay on Free ₹0/month, and credible enough for Kantar and Unilever to co-sign. If a number cannot be re-derived, it does not leave the pipeline — and that sentence, repeated across every chat and 20 to 50 page report, is the moat you can build your next launch on.