The Team Receipts

THE TEAM-EQUIVALENT STUDY

Indelible, one founder + AI, ~5 months (2026-02-18 → 2026-07-15) — measured 2026-07-15 from real artifacts

Claim references: [A#] = code census, [B#] = velocity census, [C#] = operations census. Every number below traces to a census line.

1. THE QUESTION

How many people — and which people — would a conventional software organization need to build this codebase and simultaneously operate this production footprint over roughly five months?

Not "how many people does 100k LOC imply" (a naive and discredited heuristic), but: which distinct engineering functions are demonstrably being performed, at what depth, and what would the smallest honest org chart owning each one look like? The system in question: an encrypted on-chain memory protocol (MCP server + published npm CLI), a 7-node federation bridge fleet, two self-run BSV validators with a sovereign chain indexer, a production web app, a local-only semantic/embedding search stack, a 26-agent review swarm, and the release/monitoring machinery around all of it [A1, C1].

2. THE INVENTORY

3. THE ROSTER

Heuristic (stated up front): subsystem-based staffing, not LOC-per-engineer. For each function: (a) is the evidence strong that the function is genuinely performed (not just present as files)? (b) what is the smallest fraction of a competent specialist that could own that subsystem's build + evolution over 5 months? "Conservative" assumes strong seniors wearing multiple hats; "fair-value" assumes the staffing a real company would actually hire to ship this scope at this cadence.

Total FTE equivalent: ~9 (conservative) to ~13 (fair-value).

One structural note the fractional sum hides: real orgs can't hire 0.5 of a security engineer, and an on-call rotation covering a 7-node fleet needs at least 2 humans [C17]. The ops census's own estimate for operations alone was 4–6 FTE with minimum-viable indivisible staffing at 6–8 heads [C17]; the build side roughly doubles that. A conventional company attempting this scope at this cadence would plausibly post 12–15 job openings.

4. THE HONEST CAVEATS

Brooks's law in reverse. A real 9–13 person team pays a heavy coordination tax — meetings, interface negotiation, knowledge silos, integration bugs at team boundaries. This operation has zero coordination overhead: one brain holds the entire map (literally — MAP.md, auto-regenerated [C16]), and the "team members" share perfect context. Some fraction of the FTE equivalence is therefore overhead a conventional org would need that no individual here performs. The honest reading: the output matches a 9–13 FTE team; the coordination structure is that of one person, which is precisely why the output was possible.

Quality variance vs. a real team. Bus factor is 1. Test coverage is uneven — the bridge carries 9,680 LOC of tests [A8] while the 29,632-LOC frontend has 70 LOC of in-src tests [A11, A12] and the 10,686-LOC server has one (albeit a 23-check tenancy gate) [A14]. Roughly ten ships were adversarially review-gated [C12], but review by AI agents the operation built itself is not independent audit; several subsystems have never seen outside eyes, and the census itself flags a known plaintext-WIF risk [C14]. A real team would have more redundancy and more independent scrutiny; it would also have shipped less.

"Replacing" is the wrong word. Nobody was displaced. A solo pool-company owner could never have purchased this labor — a 9-to-13-person specialist team even for five months, roughly four to five and a half engineer-years; the counterfactual wasn't a hired team, it was this not existing. The correct frame is capability expansion, not substitution.

AI is the multiplier; the human is the management layer. The AI assistants wrote most of the lines, but the census shows the human performing the roles orgs pay most dearly for: engineering management (117 scoped-and-accepted goals [B1], every publish and deploy behind an explicit human gate), product direction (the pivots and vetoes are logged), and QA direction (live acceptance criteria like preserving a real customer's data gap as the deploy test). The 5.6-vs-10.4 goals/week split [B3, B4] and the 54 handoff documents [B6] show a managed process, not an unattended one.

Measurement honesty. LOC are physical lines (~20–30% above logical SLOC) [A caveats]; category buckets are keyword heuristics [B caveats]; "LIVE fleet-wide" operational claims were verified against on-disk artifacts but not re-probed live in the census session [C caveats]; a few hundred lines are deliberately vendored twice across trees; and five additional trees (g295 worktree, published CLI, hearchain, node dashboard) were excluded, so total authored output is somewhat higher than counted [A caveats].

5. THE ONE-PARAGRAPH ANSWER

Measured from real artifacts on 2026-07-15, Indelible's five months of solo-founder + AI output is the equivalent of a 9-to-13 FTE product team: 2 protocol/backend engineers, 2 blockchain-infra engineers, 1–1.5 SREs, 1–1.5 frontend engineers, a full-stack engineer, a fractional ML engineer, plus QA, security, technical writing, product, release management, and support — and a conventional org, unable to hire fractions or run a one-person on-call, would post 12–15 openings to cover it. The three evidence lines that carry the claim: 102,294 lines of scrubbed, authored production code plus 58,799 lines of technical documentation [A1, A3]; six staged fleet-wide production rolls across 7 servers in two and a half weeks, each with canary ordering and per-box rollback [C3]; and an automated security gate that has issued 14,329 credential verdicts — rejecting 5,733 saves before secrets could be etched permanently on-chain [C11]. Nobody was replaced; this is output one person could never have bought, produced with AI as the multiplier and the founder as engineering manager, product lead, and final QA gate.