The plan that survives a bear

Read me first #

This wiki is the third document in a lineage. The v1 feasibility wiki (May 2026) made the sector case and proposed a 24-month, $120–180M plan. A mixture-of-experts review (June 2026) audited it with an independent Python model and found the per-customer math sound but the headline framing inflated. This document is the rebuilt plan: bearish assumptions, GAAP statements, five years, three scenarios, and an enterprise value answer.

The mantra is pragmatism. Every number in here traces to gaap_model_5yr.py or financial_model.py, both shipped alongside this file. If you disagree with an assumption, change it in the script and rerun. The model is the argument.

The verdict #

Lean Informatics is a credible Houston services business wearing, in its v1 draft, a $120–180M costume. Take the costume off and what remains is worth building: a firm that introduces working AI to a critical, under-digitized industry, one operator at a time, with a named human signing every output.

Method and the comb #

The first sentence of the method: nothing in this wiki rests on a number that was not recomputed independently. The process ran in three passes.

• Pass 1, mixture of experts. Four lenses applied to the v1 documents: financial modeling, market and competitive strategy, technical architecture, and a red team whose only job was to kill the idea. Findings in REPORT.md.
• Pass 2, GAAP rebuild. A monthly 60-period simulation with ASC 606 revenue recognition (onboarding fees deferred over 12 months), delivery labor in cost of services, appliance capex depreciated straight-line over 36 months, 21% federal tax with NOL carryforward, and subscriptions collected 30 days after billing. Three scenarios: bear, base, bull.
• Pass 3, the adversarial comb. A final Knuth-style line edit: every figure quoted in prose checked against model output, every claim checked for an honest counterclaim, and every known discrepancy between the analysis passes documented rather than smoothed over. The comb log is chapter 15.

The standard applied is the one Knuth applies to programs: a document is wrong until each statement has been checked against the thing it describes. Where two of our own models disagreed, the disagreement is printed, explained, and resolved in the log.

Unit economics, audited #

The v1 unit economics memo claimed an 88–92% gross margin on the hybrid architecture. The number is arithmetically honest and accountably wrong: it excludes the FDE labor that delivers the service. Standard services accounting puts delivery labor in cost of services. Recomputed at the memo's own midpoints, here is the truth table.

So the memo's directional claims hold. Hybrid does win, by about $1,700/mo of contribution over cloud, and the $5–7K/mo contribution claim is confirmed at $6,175. What does not hold is the margin's category: roughly 52% is a good consultancy, not a software company. The v1 wiki's own Addendum B.6 quietly agrees, reporting 50–55% Year 1 gross margin on the appliance tier. Both numbers were in the source documents. Sophisticated counterparties recompute this in minutes, so we lead with the honest one.

One more audit finding worth a sentence: the memo models a 500–1,000 well operator at $10–14K/mo, but Texas has only on the order of one to two hundred such operators. The 1,500–3,000 operator wedge segment (5–200 wells) prices at $1.5–4.5K/mo, where contribution per customer is $1–3K. The blended business is poorer than the memo's headline table, and the GAAP model in chapter 08 is built on the poorer blend on purpose.

What actually moves the number #

The tornado analysis settles an identity question: is this a tokenomics business or a labor business? It is a labor and pricing business. Price realization ($8K versus $14K) swings per-customer contribution by $6K/mo. FDE loading swings it by about $2.6K. The entire local-versus-frontier inference decision, even pushed to a 60% frontier share, swings it by $1–2.4K. The "buy the token cheap, sell it dear" framing from the unit economics memo is real but third-order.

Mission and posture #

The mission: bring AI into a critical industry that runs on PDFs, spreadsheets, and retiring experts. Carefully. With a named human accountable for every output, at prices a 30-well family operator can justify.

This matters beyond the P&L. Upstream oil and gas is critical infrastructure with a demographic cliff. The engineers who hold the tribal knowledge are aging out faster than they can be replaced, and the junior pipeline thinned through the 2014–2020 down cycles. AI is the only realistic absorber of that throughput loss. Someone will introduce these tools to these operators. It should be someone who shows up in person, prices honestly, and signs their name. Not a cloud platform with a 90-day cancellation trap.

The outsider stance

We are not petroleum engineers and we do not pretend to be. Frontier models score 50–70% on SPE certification subsets out of the box; the model speaks fluent petroleum. Our job is the part the model cannot do: sit with the ops director in Midland, learn how their filing actually works, ship the integration, and take the 2am call. We learn and we teach. Every engagement begins with us learning the operator's workflow and ends with their staff trained on the tool. The cross-vertical record says outsiders win when they do the discovery work: Veeva, Toast, Procore were all built by people from outside their verticals. Determined operators do not want a disruptor. They want a competent tradesperson. That is the FDE: a blue-collar knowledge worker.

Houston

Not a slogan, a cost line and a trust mechanism. Drive distance to the customer base, the Texas Railroad Commission three hours away in Austin, NAPE and CERAWeek and Fuze on home turf, and the densest AAPL membership in the country. Geographic concentration keeps delivery travel at $100–400 per customer per month in cost of services instead of airfare, and lets one FDE carry more accounts. This is the one Epic Systems discipline we keep without apology, alongside hiring implementers before salespeople.

What we sell, what we refuse #

Refused, with reasons

• Government contracts. The procurement cycle outruns a bootstrap runway. Revisit only from profitability.
• Supermajor anchors. No SOC 2 Type II, no references, no procurement standing. Chasing one is a 12-month distraction priced as a lottery ticket.
• Appliance fleet as default. Capex plus fleet operations a small team cannot carry. The two-track gate from v1 survives because it was the most disciplined idea in the document: no on-prem BOM purchased before a signed LOI, no Track 2 selling before a referenceable Track 1 customer.
• Platform and self-serve framing. We would lose that fight to an Enverus checkbox. We sell outcomes with a person attached.

Labor plan #

This is a labor company augmented by machine intelligence, and GAAP makes that visible: delivery labor is the largest line in cost of services in every scenario, every year. The plan treats payroll as the business.

• Founder, FDE #1. Time allocated 50% delivery, 25% sales, 25% G&A, and expensed to those lines in the statements. No "draw" hiding in the footnotes. Salary $120K in the early years, market rate ($160–185K) once gross profit covers it. The bear case's first sacrifice is this line, and the plan says so out loud.
• Fractional ops admin from month 6, roughly $3K/mo as a contractor. Contracts, AP, onboarding paperwork. The cheapest burnout insurance available.
• FDE hires at $190–195K base times a 1.15 payroll burden, hired on customer load and never on ambition. Bear hires at months 22, 40, 54. Base at 16, 28, 38, 48, 56. Bull adds seven. Profile: cross-vertical field engineer, industrial controls or agency background, comfortable on a customer site.
• Capacity truth. Customers per FDE plateaus around 8–12 wedge accounts or 3–4 engagement accounts blended. Revenue per FDE lands at $250–270K bear, $440–470K base, $540–570K bull. Services shape. Palantir, not Snowflake. We do not apologize for it.
• Learn and teach as a labor multiplier. Every engagement trains the operator's own staff on the tooling. This caps support hours per account, builds the trust that drives referrals, and is the honest version of land-and-expand.

5-year GAAP financials #

Assumption discipline first. Bear assumes a five month sales cycle, roughly 18% annual wedge churn, price pressure from an incumbent down-market tier, and AFK attach capped at 25%. Base assumes the plan executes with one channel assist (BETA-style referrals from Year 2). Bull, the fair-weather case, assumes the channel fires and pricing holds, but still no government contract and no supermajor anchor. All cases: $150K founder capital, no outside money, 21% federal tax with NOL carryforward, subscriptions collected 30 days after billing. Texas has no state income tax; the franchise margin tax is de minimis at these revenues and is noted, not modeled.

Full statements with all five years, D&A, taxes, and deferred revenue print when you run gaap_model_5yr.py.

The bear case, honestly #

Under GAAP, the bear case fails as drawn. Cumulative cash bottoms near −$440K against $150K of capital. An earlier quick model in the MoE audit showed a solo bootstrap cash-positive almost immediately; that finding was an artifact of $12K/mo pricing and fat onboarding fees on day one. At wedge pricing, with a market-rate founder salary on the books, slow sales burn real money for three straight years. Both models are right about their own assumptions. The bear case uses the assumptions you do not get to choose.

The bear case survives only with explicit sacrifice, and the plan prices each option:

• Founder salary cut to roughly $60K through Year 3: saves about $210K and closes half the gap.
• No FDE hire #2 until gross profit covers the loaded cost: saves about $180K, at the price of capped growth.
• Or $450–600K of capital, which is no longer a bootstrap and should be named as such.

Enterprise value #

The question the fair-weather case answers: if this works, does it build anything worth owning beyond the founder's billable hours? Yes. Two methods, deliberately conservative: a DCF on unlevered free cash flow with a high private-company discount rate (18% bear, 16% base, 14% bull, 3% terminal growth), and an exit multiple on Year 5 EBITDA (5x bear, 7x base, 9x bull, the services to tech-enabled range).

What a buyer actually acquires in base or bull, with value independent of the founder: the AFK subscription book (20–36 monitors at roughly 90% incremental margin), the operator relationships, and the eval-set and workflow library. Plausible buyers: an Enverus tuck-in, Quorum, a BETA-class land services firm digitizing, or a PE platform rolling up tech-enabled compliance. The bear case builds nothing sellable, which is the strongest argument for the kill gates in chapter 09.

Market, bearishly restated #

The addressable claim survives at the floor and only at the floor. The v1 build-up ($50K/yr across roughly 2,000 mid-size operators, plus enterprise and land-firm layers) sums to about $142M/yr in Texas alone; the $500M top of the v1 range is asserted, not derived. We need 0.7–2% of the Texas floor to hit Year 5 base or bull revenue. We claim nothing beyond that. Oklahoma and New Mexico are Year 4 options, not assumptions. Production and well-count context: EIA and the RRC statistics pages are the primary sources; treat any private restatement of them, including v1's, as indicative.

Demand evidence is real and someone else paid for it. Collide's funded existence and its Winn Resources case study, and the defensive language in the Enverus ONE launch, both confirm operators are being educated at someone else's expense. Arriving second with a cheaper, in-person offer is a fine place to stand.

Competition, worst case: Enverus ships a sub-$1K filing tier; Collide raises an A and moves down-market; open source eats the pipeline layer. Our answer is never price. It is the FDE relationship, the operator's own staff trained on our tooling, and monitors wired into their Slack that are annoying to rip out. If that answer fails, the result is the bear case, and the bear case has kill criteria.

Risk register, ranked by expected damage #

Operating calendar and gates #

Months 1–3

Build the filing pipeline against the public RRC archive. Eval set of 25 wells with known answers. E&O bound, LLC and engagement letter templates done. Selling starts week 4, not after the product is pretty: friend-of-friend, AAPL Houston chapter, Wildcatters community presence.

Months 4–12

Four to six paying wedge customers at full price. A runbook per customer. SOC 2 readiness tooling starts month 13 and only if an engagement-tier prospect demands it. Month-12 gate: fewer than 5 customers means survival mode or honorable shutdown; 5 or more means hire the ops admin fully and begin engagement-tier sales.

Year 2

First engagement-tier customers. First AFK monitors ship after the first workflow proves itself, never before. FDE #2 on the load trigger. One channel conversation (BETA-class land services) pursued seriously, as upside and never as plan.

Years 3–5

Compound. Target 60% of new revenue from existing-customer expansion and referrals. AFK attach pushed toward 35–50%. Track 2 confidential-workflow deployments only against signed LOIs. Revisit Oklahoma when Texas revenue per FDE exceeds $450K.

What works #

Stripped of v1's ambition, five things in these documents are genuinely good, and they are the whole plan.

1. The wedge is real. Recurring, quantifiable pain with public data, public EDI specs, and a competitor-funded proof of demand.
2. The FDE motion matches the buyer. Conservative operators buy a person, not a platform. The founder has shipped this shape of work before, under regulator scrutiny.
3. Houston converts geography into margin and trust. Drive distance is a COGS line and a sales weapon at once.
4. The two-track gate prevents the capex trap. No appliance before an LOI. No Track 2 before a Track 1 reference. v1's most disciplined idea, kept verbatim.
5. The AFK overlay is the quiet path from consultancy economics to enterprise value. Promoted from v1 footnote to the core thesis: engagements are CAC for monitor subscriptions.

The pragmatic mantra, applied: run the bear plan's discipline, spend to the base plan's triggers, and let the fair-weather case remain what it is. Proof that the ceiling justifies the grind, at $11–21M of enterprise value, without ever needing a miracle.

Adversarial comb log #

The final pass checked every number in this wiki against model output and hunted for claims that flatter the thesis. Findings, including the ones that cut against us:

• Verified against gaap_model_5yr.py output: all income statement figures in chapter 08 (bear 49/503/1,053; base 99/1,393/2,835; bull 150/2,163/4,501 revenue $K), all margin percentages, cash balances, headcounts, and the EV table in chapter 10 (DCF −0.27/4.79/11.35 $M; multiples 0.24/8.96/21.41 $M). All match the printed run of 2026-06-09.
• Verified against financial_model.py: the chapter 03 truth table (contribution $4,500/$4,350/$6,175; true GM 45.0%/32.2%/51.5%) and the $142M Texas MAM arithmetic.
• Discrepancy found and resolved: the MoE audit's quick cash model showed near-immediate cash positivity; the GAAP bear case shows a −$440K hole. Cause: the quick model assumed $12K/mo pricing with $25K onboarding from customer one, and treated the founder draw as $8K/mo without payroll burden. The GAAP model uses wedge pricing and market-rate burdened salary. The bear case stands as the planning number; the quick model survives only as an upper bound on capital efficiency. Documented in chapter 09 rather than smoothed over.
• Claim weakened on review: an earlier draft said the bear trough was "exactly −$442K." That is the Year 4 year-end print; the monthly trough sits within a few thousand dollars of it. The wiki now says "near −$440K." Precision should not exceed the model's resolution.
• Known model simplifications, disclosed: AFK attach is modeled as a rate on the whole base rather than per-cohort tenure; working capital is approximated (deferred revenue roughly offsets receivables); the DCF proxies free cash flow as net income, defensible because capex approximately equals D&A in all scenarios; the Texas franchise margin tax is noted but not modeled. None of these flips a conclusion; each is one line to change in the script if you disagree.
• Bias check: the bull case was constructed to exclude the two upside mechanisms the red team rated least credible (government contracts, supermajor anchors). If you believe in those, the bull numbers are conservative. That is the correct direction for a fair-weather case to err.

Sources and files #

Primary sources

• Texas Railroad Commission: filing requirements, statistics, and the public data sets the wedge product is built against.
• EIA petroleum data: production context for the Texas market claims.
• FASB ASC 606: revenue recognition treatment for onboarding fees.
• Society of Petroleum Engineers: the certification benchmark referenced for model fluency claims.
• Enverus and Collide: the incumbent and the newcomer this plan positions against.

Project files

• gaap_model_5yr.py: the 5-year GAAP model. Run it to print full statements and EV.
• financial_model.py: the unit economics audit and tornado analysis.
• BUSINESS-PLAN.md: the plan in prose form.
• REPORT.md: the mixture-of-experts audit of v1.
• lean-informatics-feasibility-wiki.html: the v1 sector research this plan descends from.
• charts/: all eight figures, generated by the two scripts.

This is not legal, financial, or regulatory advice. RRC filing is a regulated activity; a Texas-licensed compliance professional reviews the workflow before any submission is automated.