docs: production-grade README + boardroom-grade BUSINESS-PROPOSAL

Two long-form reference docs, written against real V0.10 measured numbers
(no inflated claims, no stock-startup tropes).

README.md (16 sections):
  1. Tagline + status bar (V0.10, 31 tests, AGPL-3.0)
  2. What it does, in one round trip — concrete API call → 80 s response
     with persisted audit chain
  3. Quickstart — offline 60 s, online 2 min
  4. Architecture diagram (Studio → Orchestrator → Persona registry +
     Runtime backend + Persistence)
  5. Persona library — sources (Census 2011 / NFHS-5 / Lokniti / WVS w7 /
     AI4Bharat) + Pydantic schema + Biography deep-dive (top_5_books with
     kind discriminator for non-literate personas) + generator + validator
  6. Sampling — k-DPP (Kulesza-Taskar 2012 exact spectral) with the
     negative-association property stated formally
  7. Debate flow — R0 → R1 → R2 → R3 with model selection per stage,
     concurrency notes, and the V1 ADK SequentialAgent + ParallelAgent
     re-architecture sketched
  8. Synthesis — 7-section editorial prompt + Bradley-Terry conclusion
     prompt; the §6 "argument quality, not majority" guard explained with
     reference to the V0.10-verified MSP debate (institutional skeptic
     beat the four-persona moral cluster)
  9. Audit trail — hash-chain spec, canonical_payload_json, tamper
     detection verified by an executable test
  10. Full API reference — every endpoint, request/response, examples
  11. Configuration — 13-row env-var table
  12. CLI — debate run, library generate / validate / show, eval
  13. Tests — 31 tests grouped by suite with what each verifies
  14. Performance & cost — measured numbers (5/12/30/50-persona debate
      latency + cost; library generation; pytest; Studio bundle sizes)
  15. Project layout (full tree)
  16. Roadmap (V0 → V0.10 shipped; V1 / V2 / V3 with gating milestones)
  17. License (AGPL-3.0 + commercial)
  18. Acknowledgements (every dataset + paper cited)

docs/BUSINESS-PROPOSAL.md (15 sections, ~5500 words):
  1. Executive summary — wedge + value-prop + pilot offer in one page
  2. The problem — current Indian qualitative-fieldwork toolkit
     (Dalberg, McKinsey HII, Sambodhi, Athena, Ipsos, Kantar) and its
     four binding constraints (iteration speed, coverage, defensibility,
     languages)
  3. What Akhada is — concrete, no buzzwords; cites real personas
     (Bhadralok professor's top-5; Bihar farmer's TV-serial + lokgeet
     mix; Odisha fisherman's Bhāgabata + Pala/Daskathia)
  4. The wedge — why multilateral programme offices first, with a
     comparison table for the seven adjacent wedges and why each is
     not first
  5. Value proposition — three pillars with V0.10 measured numbers
     (Pillar 1 speed table; Pillar 2 representativeness with k-DPP
     property; Pillar 3 audit-chain executable property)
  6. Why us, why now — built artefact (not promise) + Indian context
     (not localisation) + audit by default + open-core licensing;
     window 18-24 months
  7. Pricing & packaging — ₹4L pilot offer with deliverables; 5-tier
     subscription model (Free / Pro / Multilateral / Government /
     Enterprise API); year-1 ARR target ₹1.8 cr
  8. Risks & mitigations — 11-row table covering Anthropic+Pol.is
     competitive risk, multilateral procurement speed, persona-bias
     audit, hallucination, DPDP/IT Rules/EU AI Act, ECI risk, Sarvam
     single-vendor, public backlash, partner-exit, open-core fork,
     cost runaway
  9. 90-day GTM plan — week-by-week from pre-pilot hardening through
     first paid pilot delivery
  10. Pre-pilot hardening — explicit list of items shipped (✅ marks
      V0.10 capabilities) vs items still to ship in the V0.11–V0.12
      sprint (multi-tenant, ToS+DPIA, library expansion to 200,
      Wikidata fallback, AkhadaBench v0, SSE streaming, bias-audit v0,
      methodology paper)
  11. Validation evidence — three persisted real debates (table with
      their measured numbers); market-size validation (Indian qual-
      research market ₹1500-3000 cr, multilateral slice ₹400-800 cr);
      methodology-acceptance evidence (Anthropic+Pol.is, Du et al.
      ICML 2024, Stanford generative agents, SYNTHIA, Remesh);
      explicit "what is NOT validated" section (no paying customer,
      no multilateral has seen V0.10 yet, library not externally
      bias-audited, BT weights uncalibrated, 22-lang is V1)
  12. Architecture summary — pruned diagram for non-engineers
  13. Cost economics — gross margin analysis on the ₹4L pilot (~95%)
      and subscription tiers
  14. Data sources & licensing — full table including disqualified
      sources (IMDb non-commercial, Sahapedia CC-BY-NC-SA)
  15. The single contact ask — 30-min working session with a live
      programme question

Both docs are calibrated against measured V0.10 reality (not aspirational).
They are designed to be sent verbatim to a sophisticated technical or
strategic reader without further editing.

Author: sargupta