Frozen financial stress study / 20 September 2026
A reproducible record of 14,000 synthetic startup-cohort financial projections: 10,000 design trials and 4,000 separate-seed holdout trials. Every published count, capital quantile and distribution below is checked against the stored per-trial records at build time.
These are local JavaScript startup-finance calculations, not operating-agent runs, transactions, independent human observations or observed market outcomes. They do not call paid LLMs. A conditional trial share is not an empirical success probability. Capital covers a projected funding gap; it does not cause the planned member growth.
Explore the assumptions
Each profile changes only the listed inputs from the balanced reference. Do not pool their outcomes into one success estimate: these are different financial choices evaluated with paired random draws.
| Profile and changed assumptions | Trials | Capital median | Capital P95 | All goals met | No cash break-even |
|---|---|---|---|---|---|
| Member value first$2.50/member/month fee; $6,500,000 committed capital | 1,000 | $5,214,067 | $6,298,570 | 960 / 1,00096.00% conditional trial share | 14 |
| Balanced referenceUnmodified reference inputs below. | 1,000 | $4,614,349 | $5,678,821 | 987 / 1,00098.70% conditional trial share | 0 |
| Higher fee$3.25/member/month fee | 1,000 | $4,325,640 | $5,388,962 | 998 / 1,00099.80% conditional trial share | 0 |
| Slower member ramp36-month member ramp; cash break-even target month 42 | 1,000 | $5,157,704 | $6,400,812 | 823 / 1,00082.30% conditional trial share | 0 |
| Lower member scale100,000 planned mature members | 1,000 | $4,305,483 | $5,325,510 | 892 / 1,00089.20% conditional trial share | 108 |
| Higher acquisition cost$30.00 acquisition cost/new member | 1,000 | $7,637,966 | $9,151,980 | 20 / 1,0002.00% conditional trial share | 1 |
| Lower acquisition cost$8.00 acquisition cost/new member | 1,000 | $3,203,521 | $4,149,809 | 1,000 / 1,000100.00% conditional trial share | 0 |
| Higher staffing cost$150,000 loaded staff cost/person/year | 1,000 | $5,205,282 | $6,381,320 | 864 / 1,00086.40% conditional trial share | 1 |
| Lower staffing cost$100,000 loaded staff cost/person/year | 1,000 | $4,218,202 | $5,231,799 | 1,000 / 1,000100.00% conditional trial share | 0 |
| Faster member ramp18-month member ramp; cash break-even target month 30 | 1,000 | $4,304,820 | $5,309,099 | 999 / 1,00099.90% conditional trial share | 0 |
Goal categories overlap. Trials not meeting all goals are retained, not treated as execution errors. A null break-even month remains in every capital and all-goals denominator.
| Profile | No cash shortfall | Reserve + contingency | Cash break-even target | Member price protected | Not all goals |
|---|---|---|---|---|---|
| Member value first | 1,000 | 973 | 986 | 1,000 | 40 |
| Balanced reference | 1,000 | 987 | 1,000 | 1,000 | 13 |
| Higher fee | 1,000 | 998 | 1,000 | 1,000 | 2 |
| Slower member ramp | 1,000 | 860 | 919 | 1,000 | 177 |
| Lower member scale | 1,000 | 998 | 892 | 1,000 | 108 |
| Higher acquisition cost | 575 | 20 | 999 | 1,000 | 980 |
| Lower acquisition cost | 1,000 | 1,000 | 1,000 | 1,000 | 0 |
| Higher staffing cost | 1,000 | 864 | 999 | 1,000 | 136 |
| Lower staffing cost | 1,000 | 1,000 | 1,000 | 1,000 | 0 |
| Faster member ramp | 1,000 | 999 | 1,000 | 1,000 | 1 |
The balanced reference plans 150,000 mature members over 24 months, starting with 0 members. It charges $3.00/member/month, commits $6,000,000 of capital, and hires from 3 to 10 staff at $120,000/person/year. Incremental acquisition costs $15.00 per new member. The horizon is 60 months, with a month-36 cash break-even target, 3 reserve months and 15% contingency. Other inputs stay fixed across profiles unless listed in the table.
“Member value first” lowers the service fee and increases committed capital. “Balanced” means the unmodified reference, not an optimized or validated business plan. Ramp variants also change the cash break-even target as listed. The complete input record contains every parameter, including unchanged procurement, staffing and spending assumptions.
Check selected profiles with fresh draws
Only member value first and balanced were carried forward from design. The holdout seed differs from the design seed while profile inputs and the financial method stay fixed. This checks sensitivity to another pseudorandom sample under the same assumptions. It is not independent market validation, preregistered selection or validation of the other eight profiles.
| Profile and changed assumptions | Trials | Capital median | Capital P95 | All goals met | No cash break-even |
|---|---|---|---|---|---|
| Member value first$2.50/member/month fee; $6,500,000 committed capital | 2,000 | $5,242,906 | $6,308,425 | 1,924 / 2,00096.20% conditional trial share | 17 |
| Balanced referenceUnmodified reference inputs below. | 2,000 | $4,637,665 | $5,684,698 | 1,963 / 2,00098.15% conditional trial share | 0 |
Goal categories overlap. Trials not meeting all goals are retained, not treated as execution errors. A null break-even month remains in every capital and all-goals denominator.
| Profile | No cash shortfall | Reserve + contingency | Cash break-even target | Member price protected | Not all goals |
|---|---|---|---|---|---|
| Member value first | 2,000 | 1,939 | 1,983 | 2,000 | 76 |
| Balanced reference | 2,000 | 1,963 | 2,000 | 2,000 | 37 |
Bins count actual per-trial capital requirements, including the operating reserve and contingency. Each bin includes its lower bound and excludes its upper bound. Bar widths are relative to the largest bin within each profile, not a shared probability scale. All counts are available as text without color, interaction or JavaScript.
The median and P95 use linear interpolation at sorted index (n - 1) * p. P95 is a percentile of these declared assumptions, not a guarantee of sufficient capital. Unrounded cent-denominated quantiles are in the aggregate JSON; interpolation can produce fractional cents even though each trial's money is recorded in whole cents.
Growth, correlated costs and churn use symmetric triangular multipliers centered on 1, with widths 0.25, 0.15 and 0.3. Additional launch delay is discrete uniform from 0 to 6 months inclusive. These are declared stress assumptions, not fitted distributions.
A single cost multiplier is shared across the modeled cost categories in each trial. Procurement price and service-equivalence assumptions are not empirically calibrated by this study.
Design uses startup-design-v1; holdout uses startup-holdout-v1. Replication indices run from 0 to 999 for each design profile and 0 to 1999 for each holdout profile.
Draws use counter-keyed xoshiro128** with a startup:replication:factor key. Within a phase, profiles reuse the same seed and replication keys: comparisons are paired, not independent profile samples. Holdout is seed-independent of design only in this pseudorandom sampling sense.
“No cash break-even” means no six-month qualifying sequence within the horizon, not an execution failure. Cash break-even is not cumulative repayment of startup costs or an investor-return calculation. A qualifying early sequence alone does not guarantee all later months remain profitable.
14,000 planned; 14,000 completed; 0 missing; 0 execution failures; 0 excluded.
Generation aborts on a projection or serialization error. The publication build rejects missing, duplicate, out-of-order, stale or mismatched records and counts. Unfavorable outcomes are retained. Null break-even results are included in the goal and capital denominators; only the optional break-even-month quantiles in the downloadable aggregates exclude them, and report their remaining count.
The model applies an assumed member ramp and spending plan even if capital is insufficient. It does not estimate demand, causal growth from funding, service quality, general migration safety or nonmonetary welfare. Staff costs, acquisition costs and growth are uncalibrated. Taxes, debt service, required investor returns and legally approved capital repayment are not modeled.
The original aggregates retain Wilson Monte Carlo sampling intervals conditional on the declared distributions. Those intervals do not measure real-world business uncertainty and do not correct profile selection or model misspecification. This page reports conditional counts and shares, not investment confidence.
This frozen study is separate from the editable Simulation Lab and its current Follow the Money reference. No live edits silently change these published trials.
Inspect, download, reproduce
The repository stores this study in agentic-model/experiments/viability-2026-09-20/. Only the explicitly listed study artifacts are copied into this website's study-data/ directory. Other model folders and repository-only documents are not part of this download.
The funded, no-capital and member-value-first plan files are separate nominal projections with full monthly financial rows, not extra sampled trials or transaction/event traces. They supply optional frozen source profiles for Follow the Money. Its default reference is generated separately from the current Simulation Lab configuration.
Raw means per-trial summaries here. The gzipped UTF-8 JSON Lines file has one completed projection per line, with phase, profile, replication, seed reference, realized factors, financial summary and outcome flags. These cohort projections have no operating-agent transaction or event traces. The archive does not contain full monthly trajectories.
To reproduce in a checkout with Node.js 22 or later, run node scripts/analyze-startup.mjs, then node --test test/trial-evidence.test.mjs. Run the normal website build to publish validated artifacts. Gzip timestamps are zeroed; identical JSONL and compatible Node/zlib versions yield deterministic archives. A different compression implementation may change compressed bytes without changing decompressed records.
GitHub links follow the main branch. The hashes below and in the manifest identify the exact local source method used for this frozen study; a moving branch link is not an immutable revision citation.
Source hashes remove an optional UTF-8 BOM and normalize CRLF and CR line endings to LF before hashing. Data files are hashed as exact stored bytes. The build recomputes every raw-row aggregate and checks the source method, recorded inputs and file hashes; it refuses a stale or corrupt publication.
| Source file | Normalized LF SHA-256 |
|---|---|
| simulation/startup.mjs | 0b1f8222e41d02b4daf74dadc458dc47fe44e93eecd3e792a4cb9e3cc15b7415 |
| simulation/math.mjs | 719a1089cd260a2bd9b4b5b1c97239f8687320ba5ba4c592acab7e355d111454 |
| scripts/analyze-startup.mjs | 39d94ecf5737fb72eb44f809c27ef66beb6a1e36c883fe0b1eeceaf0f55ca823 |
Compressed SHA-256: 16b5c38541cf3fe6fd05eb09dd4612a723b63c3c6dcc56f0606034a2664f2b81
Decompressed SHA-256: 216a0c642063d967238941ab721d8883cb479adf636c126e64473a78ce427172
1,825,421 compressed bytes; 13,880,246 decompressed bytes.
Manifest SHA-256: d54bb541ad85195bfe7b2953260e0b3183d4b283632fbfc7064d9815c925942d
