EXHIBIT R-2 · MEASURED, NOT ESTIMATED
Parka Cost Curve — What an Agent-Operated Entity-Month Actually Costs
Date: 2026-06-10
Method: grounded in MEASURED production costs from a live agent fleet (30 days of per-session token telemetry + real dollar costs from the production cost ledger), not estimates. State fees and registered-agent pricing from /research/asymmetry-ledger (verified, primary-sourced).
Question under test (charter scope v2): can an agent-operated personal entity run all-in under ~$20/month ex-state-fees, priced at cost?
Verdict: YES at frontier-model pricing, with ~2× headroom at routed-model pricing. The binding cost is the registered agent, not the AI.
1. The entity-month ops loop
The recurring work Parka renders for one real one-person business, per month:
| # | Task | Freq/mo | What it is |
|---|---|---|---|
| 1 | Bookkeeping ingest + categorization | 4 | Pull bank/processor transactions, categorize, post to entity ledger, flag anomalies |
| 2 | Veil-integrity invariant check | 1 | Machine audit of separateness: every txn categorized, no personal-looking spend in entity account, capitalization ratio, contract-name compliance — the evidence trail that defeats piercing |
| 3 | Formality calendar + compliance check | 1 | Annual-report/franchise-tax/registered-agent status, license renewals, deadline lookahead |
| 4 | Document generation | 1 (avg) | Resolutions, contract papering in entity name, operating-agreement amendments |
| 5 | Owner Q&A | 4 | "Can I deduct this?" / "How do I sign this lease?" — interactive, small-context |
| 6 | Quarterly estimated-tax calc | 1/3 | Schedule C/SE projection, safe-harbor check, voucher prep |
| 7 | Annual close + tax-prep package | 1/12 | Year-end books close, 1040 Schedule C/1120-S package for the (human) preparer |
≈ 11 agent sessions per entity-month, most lighter than the fleet average.
2. Measured analogs (production fleet, last 30 days)
Each Parka task is costed from a real production job class that does the same class of work:
| Parka task | Fleet analog (job class) | Measured avg tokens (in / out) | Runs sampled |
|---|---|---|---|
| Bookkeeping ingest | daily financial sync | 480K / 1.8K | 38 |
| Veil-integrity check | scheduled data-health audit | 600K / 3.5K | 31 |
| Formality calendar | light responsive checks | 400K–575K / 1.2K | 35–37 |
| Doc generation | long-form content generation | 920K / 8K | 8 |
| Owner Q&A | light interactive sessions | ~200K / 2K | 174 |
| Quarterly tax calc | weekly data-health deep pass | 1.59M / 8K | 5 |
| Annual close (heavy) | multi-session build class | ~1.6M / 14.5K ×2 sessions | 33 |
Real dollar cost per cloud session (production cost ledger, LLM calls with amount>0):
- prior Sonnet generation: $0.61 avg/call (524 calls)
- current Sonnet generation: $0.86 avg/call (407 calls — output-heavier sessions)
These are blended production numbers including prompt-cache effects — the honest unit, not a token-price spreadsheet.
Important bias, stated: fleet sessions carry general-platform overhead (shared boot context and full tool catalogs — often 30-50% of input). A purpose-built Parka harness with one business's slim context would run lighter than these analogs. All compute figures below are therefore upper bounds.
3. The compute cost curve
| Scenario | $/session (basis) | Entity-month compute (≈11 sessions, weighted light) |
|---|---|---|
| A. Frontier today (Sonnet-class, fleet-measured blended) | $0.60–0.86 | $5.50–8.00/mo |
| B. Routed small-model (flash-lite-class model at ~0.06× incumbent — measured in a 13.5K-row production routing evaluation; categorization/calendar/templated-doc tasks are exactly the task types that evaluation showed promotable) | $0.04–0.08 | $0.40–0.90/mo |
| C. Local inference (workstation GPU, research scenario only) | ~$0 marginal, capacity-bounded | ≈$0/mo + amortized hardware |
Realistic blend (frontier for judgment-adjacent tasks 2/5/7, routed for mechanical tasks 1/3/4/6): ≈$2–4/mo.
4. Irreducibles (from the verified ledger)
| Line item | Annual | Monthly | Notes |
|---|---|---|---|
| Registered agent | $100–125 | $8.30–10.40 | Retail pricing; THE dominant line item. Compression path: volume/partner pricing or operating as RA directly (requires physical presence per state) → ~$25–50/yr plausible at scale |
| State annual — Wyoming | ~$60 | $5.00 | License tax floor, Wyo. Stat. §17-29-210 |
| State annual — Delaware | $300 | $25.00 | Franchise tax |
| State annual — California | $800 (+$20 SOI) | $68.30 | FTB minimum franchise tax — applies via foreign qualification even to WY-formed LLCs operating in CA |
| State annual — New York | ~$4.50 (biennial $9) | $0.40 | But one-time publication $300–1,500 by county |
| Formation (one-time, amortize yr 1) | $70–200 | — | WY $100 / CA $70 / NY $200 |
| Insurance, CPA/attorney judgment events | pass-through | — | Priced at occurrence, not retainer |
5. All-in, at cost
| Configuration | Compute | RA | State | All-in/mo |
|---|---|---|---|---|
| Frontier models, WY entity | $8.00 | $10.40 | $5.00 | $23.40 |
| Blended routing, WY entity | $3.00 | $10.40 | $5.00 | $18.40 |
| Blended routing, RA at volume ($50/yr) | $3.00 | $4.20 | $5.00 | $12.20 |
| Floor (routed + volume RA, ex-state) | $3.00 | $4.20 | — | $7.20 ex-state |
| Same person in California (reality check) | $3.00 | $4.20 | $68.30 | $75.50 — CA's franchise tax is 5–9× the entire rest of the stack |
Against the charter target (<~$20/mo all-in ex-state-fees): PASS in every scenario — even pure-frontier compute lands at $13–18 ex-state. The founding intuition holds quantitatively: the professional-services floor ($2–5K/yr) collapses to low-double-digit dollars, the registered agent is the only stubborn line item, and the AI is nearly free relative to it. The deepest irony in the table: the largest cost a Parka user faces is the state itself — California's $800 minimum is 23× the marginal compute.
6. What this does NOT yet measure (next instrumentation steps)
- Real transaction volumes — analogs assume fleet-typical context sizes; a high-volume business (150+ txns/mo) needs a measured ingest run. Build: one synthetic entity-month executed end-to-end, fully metered. (Done — see Synthetic Months 01–03.)
- Judgment-event frequency — how often per year does a real one-person business hit a question requiring a licensed human (CPA sign-off, attorney letter)? Determines the pass-through tail. No data yet; needs persona interviews or pilot.
- Purpose-built harness delta — how much lighter than fleet analogs a slim Parka context actually runs (claimed upper-bound bias; should be measured, not asserted).
- Setup month — formation, EIN, operating agreement, bank account opening: one-time heavier month, est. 3–5 extra sessions (~$2–4 frontier), plus state formation fee.
7. Provenance
- Token/session data: production fleet telemetry (30 days, 18 job classes, n=38/31/35/8/174/5/33 per analog row), queried 2026-06-10.
- Dollar costs: production cost ledger, LLM-call rows with amount>0 (n=524 + 407).
- Routed-model ratio: 13.5K-row production routing evaluation; small routed model ≈0.06× incumbent cost.
- State fees / RA pricing:
/research/asymmetry-ledgerCosts section (primary-sourced, adversarially verified 2026-06-10). - Isolation note: all measurement was read-only analysis of existing production data — no production tasks were created for Parka, per the charter's isolation boundary.