Quota-bounded autonomy, proof-first acquisition and verification debt in a zero-cash, AIbuilt venture Courage Garikai Funded by Coldcode Si Labs Disclosure. The case venture, TightBucket, was commissioned by the same owner who commissions the work of Coldcode Si Labs. The author had access to the venture's source code, documentation and operating telemetry. The case is therefore not an independent evaluation. At the time of writing TightBucket had not launched commercially: it had taken no payments and has no customer outcomes to report. Nothing in this paper should be read as evidence of commercial success. The frameworks proposed here are design propositions offered for testing. The venture was built with AI coding agents, and AI tools assisted in drafting this paper; the author is responsible for its content.
Generative AI coding tools and generous cloud free tiers have pushed the cash cost of building a working software business towards zero. When building stops being the scarce input, the design problem for a new venture changes. This paper asks what a business model looks like when it is designed around hard external constraints rather than despite them, and what new scarcities appear when a small team can build almost any‐ thing quickly. The paper reports a design-science case study of TightBucket, a website "revenue-leak" auditing service built in three days in October 2026 by its owner working with AI coding agents. The owner set explicit con‐ straints. The budget was zero. Checkout had to run entirely on a third-party merchant platform. Email had to go through a free mailbox. Four customer audiences had to be served under one brand. The system had to op‐ erate autonomously, with a human only for important decisions. And it had to stay at or below 20% of every free cloud quota while running at full capacity. Drawing on the venture's code, its documentation, its commit history, a log of 43 defects found in two review passes, and three days of usage telemetry compared against the cloud provider's own analytics, the paper makes four contributions. Constraint-native business model design. It conceptualises this as a distinct design mode in which con‐ straints are first-class design inputs, recorded in a constraint ledger, rather than environmental conditions to be overcome. Quota-bounded autonomy. It specifies this as a governance mechanism for autonomous operations, built from four steps: measure, cross-check, pace and degrade. It reports that the venture's self-meter agreed with provider analytics on AI use and, after the first day, on reads, but under-counted writes by a factor of 2.7 to 11. That gap justifies the rule "govern by the higher of the two numbers". 1. 2.
Coldcode Si Labs Library WORKING PAPER 2026-01
Coldcode Si Labs Library Working Paper 2026-01 · First edition · 2026 The autonomy boundary. It derives this from two properties of an action, its reversibility and its extern‐ ality to third parties, and explains why the venture automated first-touch email but kept social-network outreach and every reply in human hands. Verification debt. It introduces this as the liability created when AI produces systems faster than people can check them. In the case, 40 of the 43 defects were silent rather than visible, about half would have af‐ fected people outside the firm, and a fifth were residues of decisions the owner had changed. The paper closes with seven testable propositions and a Constraint-Native Business Model Canvas for practi‐ tioners. Keywords: business model design; constraints; bootstrapping; freemium; design science; AI-assisted soft‐ ware development; autonomy; platform dependence; serverless computing; verification debt.
For most of the history of the software start-up, the cost of building was the first obstacle a founder faced. Code had to be written by scarce engineers, servers had to be rented before the first customer arrived, and a payment stack had to be integrated, secured and maintained. Business model research developed under those conditions. Its classic questions concerned how a firm creates, delivers and captures value [1][2]. Its practical tools, such as the Business Model Canvas [3], were designed to help teams decide what to build before com‐ mitting expensive resources to building it. Two developments have since loosened that first constraint considerably. The first is the arrival of AI systems that write working software. Controlled studies have found large speed gains on bounded programming tasks [4], and field and experimental studies of generative AI in knowledge work report productivity effects that are substantial but uneven [5][6][7]. The second is the maturing of serverless cloud platforms that bundle compute, storage, model inference and scheduling into free allowances. Those allowances are large enough to run a real product at low volume. The serverless model was anticipated as a simplification of cloud pro‐ gramming in which developers stop provisioning servers and pay, if at all, per unit of use [8]. Its free tiers turn that into a zero-cash proposition for a venture that stays small enough. When the cost of building falls, the binding constraints on a venture move elsewhere. Herbert Simon's obser‐ vation that "a wealth of information creates a poverty of attention" [9] applies with new force to markets flooded with cheaply built software: the buyer's attention becomes scarcer as supply grows. Other constraints that were always present become decisive. These include the daily sending limits and reputation rules of email providers, the terms under which social platforms allow automation, the policies of payment platforms, the per-day quotas of free infrastructure, and the time and judgement of the one person who is legally and reputationally responsible for everything the system does. None of these is a cost that can be paid down with more engineering. They are limits within which the business model must fit. This paper starts from that shift and asks two research questions: RQ1. How can a business model be designed around hard external constraints, so that the con‐ straints become inputs to the design rather than obstacles to it? 3. 4.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
RQ2. When AI makes building cheap and fast, which new scarcities and liabilities emerge, and how can a venture govern them? We address the questions through a design-science case study [10][11] of a single venture, TightBucket, ob‐ served in fine detail during its construction and pre-launch operation. TightBucket audits websites for "reven‐ ue leaks", meaning problems with speed, pricing clarity, forms, social proof and email deliverability that cost a business customers. It scores each site out of 100 and ranks the fixes by impact. It serves four audiences: ecommerce merchants, software-as-a-service (SaaS) founders, sales directors and agencies. It offers free in‐ stant audits and ten free tools, a ladder of paid plans, a done-for-you "Fix Sprint" service, and integrations for three third-party platforms. It runs on a single serverless worker and database on a free cloud plan. What makes the case useful is not its scale, which is still small, but its specification. The owner wrote down the constraints in advance and held the build to them: Budget: zero cash. Payments: checkout handled entirely by a third-party merchant platform. Email: a free personal mailbox, with no scripting add-ons. Audiences: four, each with its own product and free tools, under one roof. Autonomy: autonomous operation, with a human reserved for important decisions. Social networks: no automation of social-network outreach. Audience at launch: none to start from. Growth model: customers should come looking for the business, like "bees to the flower", rather than only being chased. Quotas: at most 20% of each free cloud quota and of the free monthly allowance of continuous-integra‐ tion minutes, "so that I never need to worry about quotas", while running at full capacity. A specification this explicit allows each design decision to be traced back to the constraint that produced it. The venture's source repository, its operating documents, a review log of every defect found after the first re‐ lease, and its usage telemetry together give an unusually complete record of how a constraint-bound design came together. The paper makes four contributions. First, it proposes constraint-native business model design as a distinct mode of design. Prior work treats con‐ straints in three main ways: as resource scarcity that entrepreneurs overcome through bricolage or bootstrapping [12][13]; as the means-driven starting point of effectuation [14]; as a stimulus to creativity whose effects depend on its type and degree [15]. The constraint-native view goes further, and treats external constraints as specifications to be recorded, quan‐ tified and given budgets in a constraint ledger before any value proposition is chosen. The case shows how the ledger shaped the activity system [16], the pricing architecture, the distribution design and the operating model. • • • • • • • • • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Second, it specifies quota-bounded autonomy, a governance mechanism that lets an autonomous system run "flat out" without the risk of exhausting a hard external limit. The mechanism has four steps: Measure every unit of resource use where it is incurred. Cross-check that meter against the provider's own analytics. Pace background work evenly across the quota window. Degrade gracefully, rather than fail, when a budget is reached. The case data show why the cross-check is necessary. Over three days the venture's own meter matched the provider exactly on AI consumption and, after the first day, to within 2% on database reads. But it recorded between 2.7 and 11 times fewer database writes than the provider counted. The difference came from activity outside the application's instrumentation: schema migrations, deployments and administrative queries. Third, it develops the concept of an autonomy boundary: the line between what the system may do alone and what it may only prepare for a human. The boundary was drawn using two properties of each action, how re‐ versible it is and how much it affects third parties. This explains choices that otherwise look inconsistent. The system sends first-touch emails automatically but never posts to a professional network on the owner's be‐ half. It audits any public website without asking, but never publishes a negative audit of a named company. It prepares every reply but sends none. Fourth, it introduces verification debt: the liability created when systems are produced faster than they can be checked. After its first complete release the venture went through two line-by-line reviews that docu‐ mented 43 defects. Among them: a deployment check that blocked all releases; a login lockout that any stranger could trigger against the owner; a scheduled email that would have reached real people signed with a literal placeholder name. Classified by type, the findings suggest that the dominant risk of AI-built ventures lies less in code that fails to run than in code that runs and does the wrong thing, plausibly and quietly. This is consistent with security research on AI-assisted programming [17][18]. The remainder of the paper proceeds as follows. Section 2 reviews the relevant literatures on business model design, resource-constrained entrepreneur‐ ship, free and freemium pricing, signalling, platforms, automation and AI-assisted work. Section 3 describes the design-science method and the data. Section 4 describes the case. Section 5 reports the findings as a set of design principles, each traced to evidence. Section 6 proposes frameworks and testable propositions. Section 7 discusses contributions, limitations and further research. Section 8 concludes. 1. 2. 3. 4. • • • • • • • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Bracketed numbers refer to the reference list at the end. Following the disclosure above, we separate three kinds of statement throughout: findings observed in the case record, propositions inferred from it, and recom‐ mendations for practitioners.
The paper draws on seven literatures. For each, we state the core idea we use and the gap the TightBucket case helps to address.
Business model research describes how a firm creates value for customers and captures part of it. Teece defines a business model as the logic, data and other evidence that support a value proposition for the cus‐ tomer, together with a viable structure of revenues and costs for the firm delivering it. He separates the busi‐ ness model from strategy, which is the more specific analysis needed to protect a model from imitation [1]. Zott, Amit and Massa's review found the concept used in inconsistent ways. They identified a shared core that treats the business model as a new unit of analysis, emphasises a system-level and holistic view, and centres on activities as the means by which value is created and captured [2]. Zott and Amit then specified business model design as the design of an activity system, described by its content (which activities), its structure (how they are linked) and its governance (who performs them). Its value is driven by novelty, lockin, complementarities and efficiency [16], themes first identified in the study of e-business value creation [19]. Casadesus-Masanell and Ricart draw a related distinction between strategy, business model and tactics. In their account a business model is the realised logic of a firm, made of concrete choices and their con‐ sequences. Strategy is the contingent plan for choosing the business model, and tactics are the residual choices available once a model is in place [20]. This framing matters for the present case because the owner's constraints removed many choices before any strategy was formed. With checkout fixed on a merchant plat‐ form, email fixed on a free mailbox, and cash fixed at zero, a large part of the "choice space" was predeter‐ mined by external rules. Massa, Tucci and Afuah observe that the term "business model" is used to mean three different things: an at‐ tribute of real firms, a cognitive schema in managers' minds, and a formal conceptual representation such as a canvas [21]. Foss and Saebi's review of business model innovation likewise notes that weak construct clar‐ ity limits cumulative theory [22]. Chesbrough argues that the main barriers to business model innovation are organisational. Established firms struggle to experiment with models that conflict with existing assets, so ex‐ perimentation and effectuation-like processes are needed [23]. Two recent studies bring the business model literature closer to nascent ventures. McDonald and Eisenhardt studied ventures in a new market and found that the more effective designers engaged in "parallel play". They learned from peers without engaging them as rivals, tested their core assumptions, and postponed com‐ mitting to a single business model until they had learned enough [24]. Tidhar and Eisenhardt combined ma‐ chine learning with case studies to show that the choice of revenue model is consequential. Its effectiveness depends on fit with product and user attributes, and successful firms often combine more than one revenue model [25].
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Porter's earlier argument that strategy consists in a distinctive fit among activities, involving trade-offs that rivals cannot easily copy [26], also explains why a business model can be more than the sum of its parts. Gap. This literature treats the environment as the source of opportunities and threats. It largely does not treat environmental rules, such as quotas, platform terms, sending limits and payment policies, as design inputs that fix part of the activity system before strategy begins. Nor does it consider ventures in which an AI sys‐ tem performs most of the building activities.
A large literature studies how entrepreneurs act when resources are scarce. Bhide's account of "bootstrap fin‐ ance" observed that most new businesses start without outside capital. He advised founders to become opera‐ tional quickly, favour quick cash-generating projects, and grow within their means [27]. Winborg and Land‐ ström catalogued the financial bootstrapping methods small firms use to reduce their need for capital, from delaying payments to sharing resources with other firms [13]. Baker and Nelson's study of entrepreneurial bricolage showed firms "making do" by recombining resources at hand, including resources others regard as worthless, and thereby constructing their own resource environments [12]. Sarasvathy's theory of effectuation contrasts causal reasoning, which selects means to reach a given goal, with effectual reasoning. Effectual reasoning starts from given means: who I am, what I know and whom I know. It accepts affordable losses, forms partnerships and leverages contingencies [14]. The resource-based view [28] and work on resource orchestration [29] explain how firms build advantage from resources that are valuable, rare and difficult to imitate. They also show how such resources must be structured, bundled and leveraged. Research on constraints and creativity complicates the assumption that scarcity is purely harmful. Acar, Tarakci and van Knippenberg reviewed the cross-disciplinary evidence. They distinguish constraints on in‐ puts, processes and outputs, and report that the relationship between constraints and creativity is often an in‐ verted U: moderate constraints stimulate creative search, while severe ones stifle it [15]. Gap. In this literature constraints are things the entrepreneur suffers or works around. The free-tier economy introduces constraints of a different kind. They are generous, since a whole product can run within them. They are hard, since exceeding them stops the service or triggers a bill. And they are quantified, measured in rows written, requests served or minutes billed. Such constraints invite engineering against an explicit budget rather than improvisation. The case also shows a founder imposing a stricter constraint than the provider's: the 20% headroom rule. Self-imposed constraints of this kind have received little attention.
Information goods have high fixed costs and near-zero marginal costs. This makes versioning, price discrim‐ ination and lock-in central to their economics [34]. When marginal cost is near zero, a free tier becomes a cheap way to acquire customers. Behavioural research shows that "free" is not merely a low price. Shampani‐ er, Mazar and Ariely found that people respond to a zero price disproportionately, treating free products as more valuable than a simple cost-benefit model predicts [30].
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Studies of the move from free to fee show that the transition can succeed. It requires marketing actions tar‐ geted differently at free and paying users [31]. How much to give away free is a strategic variable whose op‐ timum depends on demand conditions and user heterogeneity [32]. In freemium software specifically, Gu, Kannan and Ma show that the design of the free version's limits, notably its capacity, shapes whether and when users upgrade to the premium version [33]. Gap. These studies concern the price of a product that already exists. They pay less attention to the free offer as an acquisition and qualification instrument, where the free product is itself the proof that the paid product works. The case's free instant audit plays this role.
Akerlof showed that when buyers cannot observe quality, markets can unravel as low-quality offers drive out high-quality ones [36]. Spence showed that a costly signal can separate types if the signal costs low-quality types more to send than high-quality types [35]. A new venture without an audience, reputation or referrals faces an acute version of Akerlof's problem. A cold message from an unknown sender carries little informa‐ tion about quality, and in crowded inboxes it carries negative information. A diagnostic that performs real, specific work on the prospect's own asset before any request is made can act as a signal. It is cheap for a competent provider to produce because it is automated. It is expensive for an in‐ competent one to fake convincingly, because it must be specific and correct. And it can be checked by the re‐ cipient. This logic underlies the "proof-first" stance of the case. Gap. Signalling theory is rarely connected to the operational design of acquisition systems, for example to how a diagnostic must be built so that it remains a credible signal at scale.
The case venture depends on several platforms. These are a cloud provider for compute and storage, a mer‐ chant platform for checkout and memberships, an email provider, three integration marketplaces and several social networks. Platform economics explains how platforms subsidise one side of a market to attract another [37][38], and defines multi-sided platforms as organisations that enable direct interactions between distinct sides affiliated with them [39]. Platform strategy research shows that platform markets involve strategic trade-offs [40]. It also shows that complementors face real risks: a platform owner may enter their product space, and does so more often in successful categories [41]. Gap. For a small venture these relationships are seldom studied as a deliberate offloading of business func‐ tions such as checkout, tax handling, payouts and identity, where the venture accepts dependence in ex‐ change for avoiding fixed obligations. The case makes such offloading an explicit design choice and also shows its costs.
Human-factors research offers a precise vocabulary for automation. Parasuraman, Sheridan and Wickens dis‐ tinguish four functions that can be automated to different levels: information acquisition, information analys‐ is, decision selection and action implementation [42]. Lee and See argue that the goal is appropriate reliance, meaning trust calibrated to the system's actual capabilities, rather than maximum trust [43]. Bainbridge's
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
"ironies of automation" warn that automating the routine parts of a job can leave humans with the hardest parts, while also making them less practised at those parts and responsible for monitoring a system they rarely touch [44]. In management research, Raisch and Krakowski describe an automation–augmentation paradox. Automation and augmentation are interdependent across time and space, so attempts to pursue one alone produce prob‐ lems [45]. Economists frame AI as a fall in the cost of prediction that raises the value of complementary hu‐ man judgement [46]. Gap. Little of this work addresses the externalities of autonomous business actions: what happens to third parties, to the operator's legal exposure, and to a shared reputation when a system sends, posts or publishes on someone's behalf. The case shows that externality, more than technical difficulty, determined where autonomy stopped.
Evidence on generative AI's productivity effects is now substantial. In a controlled experiment, programmers with an AI pair-programmer completed a bounded task 55.8% faster than a control group [4]. In professional writing tasks, access to a chat assistant reduced time taken and raised assessed quality [6]. In a large custom‐ er-support deployment, an AI assistant raised average productivity, with the largest gains for less experienced workers [5]. A field experiment with management consultants found strong gains on tasks inside the AI's cap‐ ability "frontier", but a significant drop in correctness on a task outside it. The authors describe this as a "jagged" frontier [7]. Security research tempers these findings. Pearce and colleagues found that roughly 40% of AI-generated pro‐ grams in security-relevant scenarios contained vulnerabilities [17]. Perry and colleagues found that parti‐ cipants using an AI assistant wrote less secure code than a control group, yet were more likely to believe their code was secure [18]. Coase's question of why firms exist, given that markets can coordinate activity, was answered by transaction costs [47]. If AI and borrowed infrastructure reduce the cost of producing a firm's internal activities, the firm's boundary and minimum efficient size may change. A single owner can now operate an activity system that once needed a team. Gap. The literature measures speed and quality at the level of individual tasks. It has not yet described the organisational liability that accumulates when a whole venture is built faster than anyone can review it, nor how such a liability is discovered and paid down. We call this liability verification debt.
A final strand links strategy to learning. A randomised controlled trial found that start-ups taught to treat their ideas as hypotheses to be tested made better decisions, including more timely pivots and terminations [48]. Firms that adopted A/B testing saw performance improvements and made more substantial product changes [49]. Gans, Stern and Wu argue that entrepreneurs cannot test every alternative. Strategy therefore involves choosing among mutually exclusive paths and committing to one, along four dimensions: customer, techno‐ logy, organisational identity and competition [50].
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
The case's growth plan embeds explicit "gates": thresholds at which a channel, message or page is killed or scaled. It deliberately avoids revenue forecasts. Both choices place it within this experimental tradition, but they also show the difficulty of experimentation before launch, when there is no data to test against.
Taken together, these literatures suggest that a business model designed under the new conditions faces four questions that existing frameworks treat only partly: How should hard, quantified external rules be incorporated into the design of the activity system? (Sec‐ tions 2.1 and 2.2) How can a free diagnostic be engineered to act as a credible signal of quality at scale? (Sections 2.3 and 2.4) Where should autonomy stop when actions have externalities? (Sections 2.5 and 2.6) What liabilities accumulate when building is fast, and how are they governed? (Sections 2.7 and 2.8) Sections 4 and 5 examine how the case answered each question.
The study follows design-science research. Design science produces and evaluates artefacts (constructs, models, methods and instantiations) that address a practical problem, and derives knowledge from their design and use [10]. We follow the process model of Peffers and colleagues: identify the problem; define the objectives of a solution; design and develop; demonstrate; evaluate; communicate [11]. In Gregor and Hevner's terms, the contribution combines exaptation, extending known solutions to a new problem, with improvement, offering new solutions to a known problem [51]. Rate limiting and metering are well understood in computing. Their use as a business model mechanism, governing an autonomous venture's activity system against free-tier quotas, is new. Business model canvases are well established. A canvas that records constraints, headroom and the autonomy boundary as design elements is an improvement. The empirical basis is a single, embedded case study [52][53]. A single case is appropriate when it is revelat‐ ory, giving access to a phenomenon previously unavailable for study [53], and when the aim is to build the‐ ory rather than test it [54]. The TightBucket case is revelatory for two reasons: its constraints were specified in writing before construction began, so each design decision can be traced to the constraint that produced it; 1. 2. 3. 4. 1. 2. 3. 4. 5. 6. •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
the complete construction record survives as version-controlled source code, operating documents, a de‐ fect log and telemetry, all at a level of detail rarely available for new ventures. The embedded units of analysis are design decisions, defects and days of operation.
Table 1 lists the data. All of it was produced in the course of building and operating the venture between 26 and 28 October 2026. None of it was produced for this study.
Source Content Volume Owner specification The constraints and requirements stated by the owner, as recorded in the project's planning notes 30+ stated constraints and corrections Build plan (v3) Strategy, channel design, architecture, gates, time budget ≈5,800 words Source repository Worker (TypeScript), cockpit and site, integrations, distribution engine (Python), database migrations, tests 15 commits; ≈7,800 lines of worker code, ≈6,450 lines of distribution code; 233 + 163 automated tests; 10 migrations Operating documents README, quota policy, set-up guide, runbooks, owner manual, distribution set-up checklist ≈10,900 words Review log Every defect found in two line-by-line reviews after the first complete release, with its fix 43 findings Usage telemetry The venture's own daily resource meter and, for the same days, the cloud provider's analytics as recorded by the venture 3 days Platform usage Continuous-integration minutes billed in October 2026 (provider billing interface) 1 month Database state Counts of audits, signals, queued items, outreach touches, members and payments at the end of the observation window 1 snapshot
No human subjects were studied and no personal data are reported. Database counts are aggregates. Creden‐ tials, addresses and private links are omitted.
The analysis had three parts. (a) Decision–constraint tracing. For each major design decision in the repository and the plan, we identified the stated constraint or constraints that required or motivated it. Decisions that no constraint explained were noted separately. The tracing produced the design principles reported in Section 5. (b) Defect coding. We coded the 43 review findings inductively into seven categories (Table 6 in Section 5.8) and recorded two attributes for each: Loud or silent: whether the defect produced a visible failure (an error, a blocked deployment, a missing page) or a silent misbehaviour (the code ran and did something plausible but wrong). • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Third-party exposure: whether its effects would have reached people outside the firm, such as pro‐ spects, customers or recipients of email. (c) Meter–provider comparison. For each day of telemetry we compared the venture's self-measured re‐ source use with the cloud provider's own figures. We then computed observed utilisation against each free quota.
Four threats deserve explicit statement. Involvement. The study is commissioned by the venture's owner and the author had full access. This per‐ mits detail but risks favourable interpretation. We mitigate it by reporting defects and negative findings, by separating observation from inference, and by refusing claims about commercial outcomes. Pre-launch status. TightBucket had taken no payments by the end of the window, so nothing here bears on whether the business model works commercially. The study concerns how a constraint-native model is designed and governed. Short telemetry window. Three days of usage at low, mostly synthetic volume cannot establish beha‐ viour under load. We use the data only to demonstrate the governance mechanism and the meter–pro‐ vider discrepancy. AI-built construction. The venture was built by its owner working with AI coding agents. This is part of the phenomenon studied, not a confound, but it means that "design decisions" were sometimes proposed by an AI system and accepted by the owner. Where it matters, we note who decided.
The owner's starting point was a conversation in which a simpler pipeline had been proposed: a free leadmagnet tool, a "no-brainer" upgrade offer and direct messages on a professional network. The owner asked for a "far superior" operation designed around a set of constraints. In the order the owner stated or later cor‐ rected them, the most consequential were: Zero budget. No cash outlay; free tiers only. Checkout entirely on a third-party merchant platform (Whop), including memberships and pay‐ ments. Email through a free personal mailbox (Gmail). An earlier design based on a spreadsheet and scripting add-on was to be removed, with Gmail sending preserved. Four audiences under one roof (sales directors, agencies, SaaS founders and e-commerce merchants), each with its own product and free tools. Autonomy. The pipeline should run by itself, with a human only for very important activities. No automation of the professional network. Third-party tools were preferred over brute-force automa‐ tion, which the network's terms prohibit. No existing audience or network to draw on. • • • • • 1. 2. 3. 4. 5. 6. 7.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Inbound pull. Customers should seek the business out ("bees to the flower"), not only be pursued. Quota headroom. The whole pipeline was to use at most 20% of the cloud provider's free quotas and of the free continuous-integration minutes while operating "at full tilt", "so that I never need to worry about quotas". Private repository and a provider subdomain for the live site, with no paid domain yet. Owner login built in, password only, rather than depending on an external access product. A separate interface for paying customers. This reversed an earlier decision to keep every customer in‐ teraction on the merchant platform. Line-by-line review. Every line of code reviewed for errors and points of breakage, without adding fea‐ tures during reviews.
Automatic client workspaces created and handed over at purchase, with no manual step. The specification is unusual in that it states budgets and boundaries rather than goals. It says almost nothing about revenue targets. The build plan likewise states that "no revenue forecasts are given here, on purpose", and replaces forecasts with gates (Section 5.7).
TightBucket's core product is an automated audit of a public website. Given a URL, the worker fetches the page through a guard against server-side request forgery, extracts its structure, and runs weighted checks in four modules: page speed, using a public performance API; pricing and offer clarity; forms and conversion paths; trust and email authentication, using domain-name records. It scores the site out of 100, ranks fixes by expected impact, and publishes a private report page. Any site owner can opt a domain out, which blocks audits of it and cancels any pending outreach to it. The public methodology page is generated from the check library itself, and the deployment pipeline fails if the pub‐ lished methodology drifts from the code that scores sites. Each audience has its own audit page and free tools (Table 2). There are ten free tools in all, ranging from a domain-authentication checker to a churn-risk scorer that runs in the browser so that the uploaded customer list never leaves the user's device.
Audience Audit Free tools E-commerce Store audit Store speed test; cart-abandonment calculator SaaS founders SaaS audit Pricing-page grader; churn-risk scorer; churn calculator Sales directors Pipeline audit Domain-authentication checker; cold-email grader; demo-form grader; lead-list prioritiser; pipeline-velocity calculator
8. 9. 10. 11. 12. 13. 14. 15. • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Audience Audit Free tools Agencies Agency audit Cold-email grader; lead-list prioritiser; white-label reports and a 30% recurring partner programme
The paid ladder has six rungs: Anonymous use: free; five instant audits a day and all browser tools. Free membership: $0; three saved audits a month, one monthly monitor and report history. Pro: $39 a month; 500 audits a month, five weekly monitors, export and an API key used by the browser extension and site plugin. Agency: $149 a month; 25 monitors, white-label reports and partner terms. Fix Sprint: done-for-you remediation of the top leaks, from $300 and scoped first. Store plan: $19 a month, sold inside the e-commerce platform's own app store through that platform's billing. Referrals earn one extra monitor per activated referral, up to ten. Every rung above anonymous use is sold through the merchant platform, which charges 2.7% plus $0.30 per successful domestic card transaction, with surcharges for international cards and currency conversion [58]. The exception is the store plan, which marketplace rules require to be billed by the e-commerce platform.
A single serverless worker on the provider's free plan serves: the static site; the public and authenticated APIs; report pages; the in-platform member view; the customer portal; the owner's cockpit; webhooks; two scheduled triggers, one running every minute and one daily, out of the five a free account allows. A serverless relational database holds all state. The provider's key-value store, whose free allowance permits only 1,000 writes a day, is not used at all. AI summaries use a small hosted model, chosen because it con‐ sumes about one-sixth of the inference units of the non-optimised variant of the same model. Email leaves through two lanes: the free mailbox's API, which carries transactional, nurture and reply traffic and a share of cold first touches; an authenticated mail connection to a second free webmail provider, which adds capacity for cold first touches on a separate reputation. That provider allows 15 messages an hour, and the cloud platform blocks the standard server-to-server mail port, so the lane uses an authenticated submission port. 1. 2. 3. 4. 5. 6. • • • • • • • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
A priority outbox orders traffic as transactional, then replies, then nurture, then cold, under a daily cap of 400. That cap sits below the free mailbox's own limit of about 500 a day [55]. Every message carries unsub‐ scribe and data-deletion links. One-click unsubscription follows the relevant standard [57] and current sender guidelines [56]. Suppression survives deletion, stored as hashes. The owner works from a phone-installable cockpit. Its tabs are Today, Outreach, Inbox, Community, Content, Gates, Seeds, Privacy, Health, Clients and Mesh. Login uses a password hashed with a standard key-deriva‐ tion function, a second factor that is off by default, and a per-address lockout. Paying customers receive a team workspace, created automatically when the purchase webhook arrives. They are invited by an email claim link and a button inside the merchant platform. Seats and roles are set by plan, and team members are invited by email. No per-client infrastructure is created. A separate distribution engine runs four times a day on the continuous-integration platform. It drafts and pub‐ lishes posts across more than thirty social "rails", using a scheduling service and direct platform interfaces. It is armed but dormant: each rail activates only when its own credentials exist, and the whole engine runs only when a repository variable switches it on. At the end of the observation window no social accounts existed, so no rail was active.
The venture was built in three days (Figure 1).
Date Milestone Oct First complete release ("1.0.0"); hardening pass; quota governor; CI optimisation; production-confirmed unit costs; owner-console login, product catalogue with covers and real screenshots; second email lane; password-only owner login; full review (32 findings) and customer portal; second review (11 findings); client team workspaces Oct Distribution engine ported from a sister venture, armed but dormant; fix to image serving found in a dry run Oct Owner console shows each distribution rail as active, "on (waiting for account)" or parked
At the end of the window the database held 19 audits, 29 buying signals, 31 community questions queued for the owner, 9 prepared outreach touches, no sent emails, no payments and one member: the owner's compli‐ mentary account. The venture was operational but had not launched.
This section reports the findings as eight design principles. Each principle states what the case did, the evid‐ ence for it, and what the evidence does and does not show. The first principle concerns the whole design. The next six concern the main parts of the activity system: capacity, acquisition, commerce, communication, autonomy and distribution. The last concerns the method of construction.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
What the case did. The owner stated budgets and boundaries before goals (Section 4.1). The build plan then treated each one as a specification. Its architecture section begins with the free-tier limits it must stay within, and only then describes the product. Evidence. Table 3 traces 24 major design decisions to the constraints that required or motivated them. Con‐ straint numbers refer to the list in Section 4.1. "External" marks a rule imposed by a provider or platform rather than stated by the owner.
# Design decision Traced to One serverless worker on the free plan serves every surface C1, C10 No key-value store at all C9; external (1,000 writes/day allowance) Audit findings compressed onto one row (≈120 → ≈11 row writes per audit) C9 Small, optimised AI model; deterministic templates as fallback C1, C9 Two of five scheduled triggers C9 (see note) Checkout, memberships, tax and payouts on a merchant platform C2 Store plan billed by the e-commerce platform External (marketplace billing rule) Email through the mailbox API; spreadsheet and scripting add-on removed C3 Daily email cap of 400 with four priority classes C3; external (≈500/day mailbox limit) Second cold lane over authenticated mail submission C14, C3; external (15/hour; blocked mail port) Instant audit with no sign-up C7, C8 Four audience audit pages and ten free tools C4, C8 Professional-network messages prepared for the owner, never sent automatically C6; external (network terms) Community questions queued for the owner, never auto-posted C5 Quota governor: measure, cross-check, pace, degrade C9 Built-in password login with per-address lockout C11 Customer portal separate from the console C12 Team workspaces created and handed over at purchase C15 Distribution rails armed but dormant until credentials exist C7 (no accounts) Churn scorer runs in the user's browser C1, C9 Line-by-line reviews, one test per fix, no features during reviews C13 Public methodology generated from the scoring code none 30% recurring partner programme for agencies none Kill-or-scale gates instead of revenue forecasts none
Note to row 5. The owner's 20% rule was applied to consumption quotas such as rows, requests, inference units and minutes. The count of scheduled triggers is a fixed allowance, and the case uses two of five. This is an interpretation of the rule, recorded here so it is not mistaken for compliance.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Twenty-one of the twenty-four decisions trace to at least one stated constraint or external rule. In five cases (rows 2, 7, 9, 10 and 13) the binding rule came from a provider or platform, not from the owner. The owner did not know these rules when stating the constraints; they were discovered during design. Only three de‐ cisions are unexplained by constraints: the generated methodology, the partner programme and the gates. All three are trust and learning devices rather than capacity devices. What it shows. In a constraint-native venture, most of the activity system is decided before strategy in the usual sense begins. What remains for "strategy" is mostly about trust and learning. A constraint list is also never complete at the outset, because external rules arrive during design. This is the case for a ledger that is maintained over time, rather than a list fixed at the start (Section 6.1).
What the case did. Because the pipeline runs without supervision, nothing human would notice a slow drift towards a free-tier limit. The case therefore built a governor in four steps. Measure. Every database call records the rows read and written as reported by the database, including index writes, which the provider bills as written rows [59]. Every AI call records the inference units it consumes. The totals are flushed to a single row per day. Cross-check. Every thirty minutes a read-only query to the provider's analytics retrieves the provider's own figures, and the governor uses the higher of the two. Pace. Background work (prospect audits, enrichment, AI classification) may use 70% of the budget, re‐ leased evenly across the UTC day with a one-hour head start. A busy morning therefore cannot consume the afternoon's allowance. Degrade, do not fail. When a budget is reached, anonymous visitors receive a cached report for the same site if one exists from the last 30 days, or a "try again after 00:00 UTC" message. AI summaries fall back to a second free provider and then to templates. Paying members, monitors and the paid API are never re‐ fused. The budgets are set at 20% of each free allowance: 20,000 row writes and 1,000,000 row reads a day, against provider allowances of 100,000 and 5,000,000 [59]; 2,000 inference units a day, against 10,000 [61]; 400 continuous-integration minutes a month, against 2,000 for a private repository [62]. The minutes budget is enforced by a guard that skips the slowest test job when 80% of the budget has been used. Evidence: unit costs. Unit costs were measured in production. An anonymous audit costs about 11 row writes and 10 row reads, a report view 1 write and 2 reads, and an idle scheduled minute no writes and about 2 reads. One AI summary costs 4.41 inference units by the governor's own calculation, against 4.4096 in the provider's analytics. 1. 2. 3. 4. • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Two engineering changes were driven by the meter: Before a "lean pass", an audit cost about 120 row writes, because each of 38 findings was stored as a row with two indexes. Compressing the findings onto the audit row cut this by about 91%. An early budget query for email made three range scans per message. Replacing them with one grouped count removed an estimated 2.6 million row reads a day at full sending volume, roughly half the pro‐ vider's daily free read allowance. Unmetered, this one query pattern would have broken the owner's 20% rule on its own. Evidence: meter versus provider. Table 4 compares the venture's meter with the provider's analytics for each day observed.
Day (2026) Rows read: meter / provider Rows written: meter / provider Provider ÷ meter (writes) AI units: meter / provider Requests (provider) 26 Oct 3,450 / 5,759 71 / 783 11.0× 4.41 / 4.41 27 Oct 9,599 / 9,480 402 / 1,839 4.6× 38.00 / 38.00 1,438 28 Oct (to 11:03 UTC) 5,615 / 5,433 396 / 1,051 2.7× 45.47 / 45.47
The two AI meters agree exactly. Reads agree to within 2% except on the first day, when schema migrations and deployments ran outside the application. Writes do not agree: the provider recorded between 2.7 and 11 times more writes than the application's meter. The cause is structural. Migrations, deployments, console queries and administrative scripts write to the database outside the code paths that the application instru‐ ments. The discrepancy was largest on the day of heaviest construction and fell as construction slowed. At the observed peak, the venture used 0.19% of the free read allowance, 1.84% of the write allowance, 0.45% of the inference allowance and 1.44% of the request allowance [60]. It also used 39 continuous-integ‐ ration minutes in October, 1.95% of the free allowance. What it shows. A self-meter is necessary but not sufficient. It undercounts precisely the activity that con‐ struction and operations generate, which is when an AI-built venture is most active. Governing by the higher of the two meters is a simple, conservative rule that removes this blind spot. The evidence does not show that the governor works under load. Observed usage never came near a budget, so pacing and degradation were exercised only in tests. The simple arithmetic of the budgets suggests the room available: about 1,800 an‐ onymous audits a day at 11 row writes each, within the 20,000-write budget. That figure is an estimate, not an observation.
What the case did. Every route into TightBucket begins with a piece of work done for the visitor or pro‐ spect: The instant audit needs no sign-up and returns a scored report with ranked fixes in seconds. The free tools produce an answer on the page. • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Outbound touches, whether email or messages prepared for the owner, attach an audit of the prospect's own site rather than a description of the service. The ladder continues in the same way (Section 4.2). A free account turns one-off proof into saved proof. A paid plan adds recurring proof through monitors. A team or white-label plan adds delegated proof, which an agency shows to its own clients. The top rung is done-for-you remediation. Evidence. By the end of the window the database held 19 audits and 9 outreach touches, each prepared with a prospect-specific audit. The deployment pipeline blocks any release in which the public methodology dif‐ fers from the scoring code (Section 4.2). The recipient of an audit can therefore check how the score was computed. What it shows. In signalling terms [35], the audit is a signal that is cheap for a competent sender because it is automated. It is costly to fake because it must be specific and correct about the recipient's own site. And it is verifiable because both the method and the evidence are visible. This answers the unknown sender's ver‐ sion of the lemons problem [36]. It also addresses attention scarcity [9]: the message carries information about the recipient rather than about the sender. The free audit is thus less a "price of zero" in the sense of Section 2.3 [30] than an acquisition and qualification instrument. The evidence shows how the signal was designed, not whether recipients find it credible. No conversion data existed at the time of writing.
What the case did. All paid rungs except the store plan are sold, collected and administered on the merchant platform. The venture receives verified webhooks and an identity the customer can use to sign in. It holds no card data, files no sales tax returns and runs no payout system. Evidence: cost. The platform's fee is 2.7% plus $0.30 per successful domestic card transaction, with 1.5% added for international cards and 1% for currency conversion [58]. On the venture's prices the effective fee rate falls as price rises (Table 5).
Plan price Domestic card International card International card with conversion $39 3.47% 4.97% 5.97% $149 2.90% 4.40% 5.40% $300 (lowest Fix Sprint) 2.80% 4.30% 5.30%
These are the only variable costs in the model, and they arise only when a customer pays. With no fixed costs, the venture's break-even point is its first sale. Evidence: dependence. Offloading did not remove work; it changed its form. Three of the 43 review find‐ ings (Section 5.8) were in entitlement logic that interprets the platform's events: a cancelled membership event could downgrade a complimentary account; a $0 payment could activate a referral reward; payments without an identifier could be stored twice. • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
The owner also reversed the founding decision to keep every customer interaction on the platform, and asked for a separate customer portal (constraint C12). The venture therefore runs its own sign-in, workspace and team logic alongside the platform's. What it shows. Commerce offloading trades margin and dependence for the absence of fixed obligations. That suits a zero-cash venture, but the trade has two hidden costs: the logic that translates platform events into entitlements, and the pressure to own the customer relationship. The platform-dependence literature [41] warns that the platform can change the terms. The case adds that the complementor must also interpret the platform correctly, and that interpretation errors are silent.
What the case did. Email capacity is modelled as a stock with a daily refill and a reputation that can be spent. Daily refill. The free mailbox permits about 500 messages a day [55]. The venture caps itself at 400, planned as 40 cold, 300 nurture and transactional, and 60 in reserve. Priority classes. Messages leave in priority order: transactional, then replies, then nurture, then cold. A second lane. The second provider's limit of 15 messages an hour, shared across all uses, makes it a narrow channel. Reputation. Sender guidelines require bulk senders to keep reported spam below 0.1%, to authenticate their domain and to offer one-click unsubscription [56][57]. The design follows from this. Every message carries unsubscribe and delete links, and suppression survives deletion by being stored as hashes. Two kill switches stop all email or only cold email. A cold-bounce gate trips automatically and pauses a lane when bounces, opt-outs, spam-folder placement or a provider lock are detected. Evidence. Five of the 43 review findings concerned the accounting of reputation rather than sending itself. For example: a recipient rejection at the mail server was not counted in the bounce rate, so the gate could not see it; a failure while resetting the connection after a rejection replaced the true classification, so a permanent "bad recipient" error was retried as a connection failure instead of being suppressed. None of the five would have prevented sending. Each would have blinded the mechanism that protects repu‐ tation. What it shows. When distribution depends on borrowed inboxes, the asset is the reputation, not the capacity. The binding risk is silent mis-measurement of the reputation signal, not exceeding a send limit. The case's sending volume was zero during the window, so the evidence is about design and defects, not about deliver‐ ability outcomes. • • • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
What the case did. The owner wanted the pipeline to run itself with a human "only for very important activ‐ ities" (C5). He also wanted first touches into real inboxes to be autonomous, while he personally worked the replies (C14). In practice the boundary was drawn by the consequences of an action rather than its difficulty: Automated without gates: audits, scoring, monitoring, buying-signal collection, report pages, and draft‐ ing of content and touches. Automated with gates: cold first-touch email. It runs under the daily cap, the priority order, the bounce gate, kill switches and hashed suppression. Prepared for the human: messages on the professional network, whose terms prohibit automation; replies to prospects; and answers to community questions. At the end of the window 31 community ques‐ tions and 9 touches were waiting for the owner. Human only: pricing, refunds, partner agreements, and scoping of done-for-you work. The plan sets a design target of about 25 minutes of owner time a day, spent mainly on replies and prepared messages. This is a target, not a measurement. Evidence. The clearest evidence is a defect. An AI classifier labelled an incoming reply as an "unsubscribe" message. The triage code then ran the data-erasure routine whenever the word "delete" appeared anywhere in that reply, for example in "how do I delete a monitor … unsubscribe". An irreversible action with con‐ sequences for a customer was triggered by a probabilistic classification combined with a keyword. The fix al‐ lowed erasure only on an explicit erasure request. A related defect allowed a double-submitted reply to be sent twice. The fix makes each send claim its mes‐ sage atomically. What it shows. Technical capability did not decide where the human was needed. Two properties did: whether the action can be reversed, and whether its effects fall on third parties or on the operator's legal and reputational position. Both properties already appear in the human-factors literature on levels of automation [42] and calibrated reliance [43]. The case makes them business model choices. They also reverse Bain‐ bridge's irony [44]: the human is kept on the actions that matter most, by design.
What the case did. The venture had no audience and no social accounts (C7). Its distribution engine was nevertheless built in full: more than thirty publishing rails across a scheduling service and direct platform in‐ terfaces, plus optional sets of rails for individual sellers. Each rail activates only when its own credentials ex‐ ist, and the whole engine runs only when one repository variable switches it on. On the final day the owner noticed that the console showed rails as "active" when they were merely switched on and waiting for an account. He asked for the label "on (waiting for account)". The console now distin‐ guishes active, on (waiting for account) and parked. Evidence. Two commits implemented the engine and a dry run. The dry run found one defect: images stored in the database were served as text. A third commit implemented the label. At the end of the window no rail was active. • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
What it shows. The cost of building capacity has fallen far enough that a venture can hold dormant capacity that it switches on when an account is created, rather than building a channel when it is needed. This changes the order of work. Accounts, not code, become the bottleneck, and a checklist of accounts to create becomes a strategic document. The label correction shows the risk. Dormant capacity reported as active gives an operator a false sense of reach. In an autonomous venture with few human eyes, the honesty of status labels is a governance issue, not a cosmetic one.
What the case did. The venture was built in three days by its owner working with AI coding agents. The owner required every line to be reviewed for "errors, breaks and points of breaking", with no new features added during reviews (C13). After the first complete release there was a hardening pass and then two docu‐ mented line-by-line reviews. Every finding was fixed with a test, and the automated test count rose from 187 after the first review to 206 after the second; it stands at 233 today. Evidence. Table 6 codes the 43 documented findings.
Category Findings Example Third-party-facing communication errors A rescan email would have reached real people signed with a raw template placeholder; triage rules fired on unsubscribe text quoted from earlier messages Resource, quota and capacity governance Blocked outbox rows occupied the selection window and starved sendable mail; outbound calls had no timeout Configuration and productsurface drift Deployment preflight still required a second-factor secret after the owner made it optional, blocking every deploy; copy still promised an authenticator Deliverability and reputation accounting Server-side recipient rejections were not counted in the bounce rate Data integrity and input validation A status was updated before the input was validated, so a rejected request still changed data Security and privacy Owner-login lockout was global, so anyone could lock the owner out for 24 hours; deletion scrubbed only the first of several matching records Money and entitlement logic Cancellation events downgraded complimentary accounts Total
Three patterns stand out. Most defects were silent. Only 3 of the 43 produced a visible failure: the blocked deployment, a set of portal routes that were not served, and a 404 for HEAD requests. The other 40 let the code run and do something plausible but wrong. They would not have been found by watching the system work, only by reading it. All three loud failures were themselves cases of decision drift (below). Half had third-party exposure. Twenty-two findings (the communication, deliverability, entitlement and privacy categories) would have affected people outside the firm: prospects, recipients or customers. • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
A fifth were "decision drift". Eight findings were residues of decisions the owner had changed. The op‐ tional second factor and the reversal of the platform-only customer interface left code, copy, comments and routes that still assumed the earlier decision. What it shows. Fast building creates a liability that grows with the speed of building, not with the size of the team. We call this liability verification debt: the stock of unreviewed artefacts whose behaviour has not been checked against the owner's intent. It extends the security findings of Pearce and colleagues [17] and Perry and colleagues [18] from code security to business behaviour. The pattern of silent, third-party-facing defects matches the "jagged frontier" [7]: the output looked equally competent where it was right and where it was wrong. Decision drift is a specific mechanism that the AI-productivity literature has not described. When an owner can change a decision in one sentence, and an agent implements the change in one place, the old decision survives everywhere else. The case also shows a way of paying the debt down: line-by-line reading, with each finding converted into a test that stops it returning. What the case does not show is how many defects remain. Two passes found 32 and then 11. That decline is consistent with diminishing returns, but it is not proof of completeness.
This section turns the principles into frameworks that practitioners can use and propositions that researchers can test. The frameworks are proposals drawn from one case. They are offered for use and testing, not as es‐ tablished results.
A Constraint Ledger is a maintained register of every constraint that bounds the business model. Each entry records six fields:
Field Meaning Example from the case Source Who or what imposes the constraint Cloud provider's free plan Type Hard (breach stops service or incurs cost), soft (breach degrades a result) or self-imposed Self-imposed (20% of a hard limit) Metric The unit in which the constraint is measured Database rows written per UTC day Budget The amount the business allows itself 20,000 (70% of it for background work) Owner The person or mechanism responsible for keeping within it Quota governor; owner reviews weekly Review date When the entry must be checked against the source Monthly, or on any provider pricing change
The ledger has three purposes, each grounded in the findings. It makes constraints design inputs. In the case, 21 of 24 major decisions trace to constraints (Section 5.1). It records constraints discovered during design, which in the case included five external rules the owner had not stated. • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
It gives a self-imposed constraint the same standing as an external one. Such a constraint is a strategic choice that buys protection against both provider changes and the owner's own attention lapses. A ledger differs from a risk register. A risk register lists uncertain events. A ledger lists certain rules and the budgets set against them.
Two quantities summarise a constraint-native venture's capacity position. Both are calculated separately for each metered resource: Headroom ratioH = budgeted use ÷ provider allowance. It is a policy choice; in the case H = 0.2 for every consumption quota. Observed utilisation U = max(self-measured use, provider-reported use) ÷ provider allowance. The governing rule is U ≤ H for every resource at all times, enforced by four steps: Measure every unit at the point of consumption. Cross-check against the provider and keep the higher figure. Pace background work evenly across the accounting period. Degrade gracefully, protecting paying customers first. The case's data show why step 2 matters. Self-measured writes understated provider-measured writes by a factor of 2.7 to 11 (Table 4). Under-measurement of this size would let a venture believe it is at 20% when it is at more than 50%. Choosing H is a strategic decision. A low H: protects against provider changes, sudden spikes and metering error; frees the owner from monitoring; costs unused capacity. A high H buys growth before any upgrade but leaves no margin. A ledger entry should also record the up‐ grade trigger: the observed condition at which the venture moves to a paid tier. Moving to a paid tier is not failure. It is a planned step in the zero-cash model (Section 6.6).
The Proof-First Value Ladder orders an offer so that each rung extends the same proof rather than adding un‐ related features:
Rung Proof Price in the case Economic role 1. Instant proof One diagnostic of the visitor's own asset, no sign-up Free Signal and qualification 2. Saved proof History of diagnostics; a free account $0 Identity and permission to follow up 3. Recurring proof Monitoring that re-runs the diagnostic on a schedule $39/month Retention through change detection
• • • 1. 2. 3. 4. • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Rung Proof Price in the case Economic role 4. Delegated proof Team seats; white-label proof shown to one's own clients $149/month Expansion; customers resell the proof 5. Done-for-you The provider fixes what the proof found From $300 High-margin service, scoped by the proof
The ladder joins the freemium literature [32][33] with signalling theory [35]. The free rung is a credible, checkable signal, and each paid rung raises the frequency, audience or consequence of the same proof. The ladder also limits feature creep. A proposed feature belongs on the ladder only if it extends the proof. In the case, the browser extension and the site plugin both qualified, as ways of running the same audit from other places. A generic analytics dashboard would not have qualified.
The Autonomy Boundary Matrix places each business action on two axes: the reversibility of the action: can its effects be undone cheaply and completely? its externality: do its effects fall on third parties, or on the operator's legal or reputational position? The resulting quadrants prescribe the level of automation:
Low externality High externality Reversible Automate. Audits, scoring, monitoring, drafting, signal collection Automate with gates. Cold first-touch email under caps, priority order, bounce gate, kill switches and suppression Irreversible Prepare for the human. Database migrations and bulk data changes; retiring a channel; scoping of done-for-you work Human only. Replies to prospects, messages on networks that prohibit automation, pricing, refunds, partner agreements
The case assigned replies and network messages to "human only" for sending, while preparing them auto‐ matically. The drafting sits in the top-left quadrant and the sending in the bottom-right. Data erasure shows that an action can be automated behind a narrow gate. After the defect described in Section 5.6, erasure runs automatically only on an explicit request from the data subject, which keeps the irreversible step tied to an unambiguous human instruction. The matrix differs from levels-of-automation taxonomies [42], which classify functions of a single task by how much the machine does. It classifies business actions by their consequences. It is grounded in the eras‐ ure defect (Section 5.6), where an irreversible action with customer impact had drifted into the "automate" quadrant through a classifier and a keyword. A practical test follows: list every action the system can take, place each in a quadrant, and check that the code's actual gates match the quadrant. • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Verification debt is the stock of artefacts produced by a venture, including code, copy, configuration and automated decisions, whose behaviour has not been checked against the owner's current intent. It differs from technical debt, which is a known compromise in design quality taken on deliberately. Verification debt is un‐ known behaviour taken on by speed. Its main components are silent defects and decision drift. We propose five measures, each computable from a review log of the kind kept in the case:
Measure Definition Case value Review yield Findings per review pass 32, then 11 Decay ratio Yield of pass n ÷ yield of pass n − 1 0.34 Silent share Silent findings ÷ all findings 40 / 43 = 93% Third-party share Findings with third-party exposure ÷ all findings 22 / 43 = 51% Drift share Findings caused by changed decisions ÷ all findings 8 / 43 = 19%
The silent share shows why monitoring a running system is not enough. The third-party share shows why the debt is a business risk and not only an engineering one. The drift share links the debt to how often the owner changes decisions. A decay ratio well below 1 across successive passes is a reasonable, though not conclus‐ ive, signal that the debt is being paid down. Two practices from the case reduce the debt: line-by-line reading by a reviewer who did not write the code, and one test per finding. A third, propagation of decisions, follows from the drift finding. It means searching the whole codebase, copy and documentation for every trace of a decision when that decision changes.
The case suggests five conditions under which a digital business can operate with no cash outlay: Every fixed cost sits on a free tier with headroom (U ≤ H for every metered resource). Every variable cost is linked to success. The only variable costs arise when a customer pays, as with payment-platform fees. Unit costs are measured at the level of the provider's meter, such as rows, inference units and minutes, not estimated in dollars. Owner attention is treated as the binding budget, with a design target for daily human time and an ex‐ plicit allocation of it, in the case mainly to replies. Upgrade triggers are explicit. The ledger records the observed conditions under which a paid tier is ad‐ opted, for example sustained CPU limits on the serverless plan. Under these conditions the venture's break-even point is its first sale, and the cost of the venture failing is the owner's time. This combination is new. It makes the choice of whether to start a venture closer to an afford‐ able-loss decision in Sarasvathy's sense [14] than to an investment decision. 1. 2. 3. 4. 5.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
The Constraint-Native Business Model Canvas (CN-BMC) extends the nine blocks of the Business Model Canvas [3] with three blocks. Each makes one of the findings a required element of a business model de‐ scription.
Added block Question it answers Case content (summary) Constraint Ledger What rules bound this model, in what units, with what budgets? Zero cash; merchant-platform checkout; free mailbox; 20% headroom; network terms; no audience Headroom and degradation plan What happens at the budget, and who is protected first? Pace background work; cached reports; template fallback; paying customers never refused Autonomy boundary Which actions run alone, which are gated, which wait for a human? Matrix in Section 6.4; owner time target ≈25 minutes/day (design target)
The added blocks connect to the original ones in three ways: The constraint ledger constrains the key resources and cost structure blocks. The degradation plan protects the customer relationships block. The autonomy boundary determines how much of the key activities block a human performs. The canvas therefore keeps the familiar language of value propositions and customer segments while making constraints, capacity and autonomy visible.
The findings suggest seven propositions. Each is stated so that it can be tested, and each is followed by a sug‐ gested test. P1 (Constraint-first design). Ventures that record external constraints as quantified budgets before choosing a value proposition will experience fewer unplanned service interruptions and fewer architectural reversals after launch than comparable ventures that discover constraints in operation. Test: a matched comparison of early-stage ventures on free tiers, coding planning documents for ledger-like content and incident logs for quota-related interruptions. P2 (Dual metering). In autonomous ventures, self-measured resource use will systematically understate pro‐ vider-measured use, and the gap will be larger during periods of construction than during steady operation. Governing by the higher of the two measures reduces quota breaches without materially reducing throughput. Test: paired daily series from multiple ventures, regressing the provider-to-meter ratio on construction activ‐ ity (commits, migrations, deployments). P3 (Proof-first acquisition). For a seller unknown to the recipient, outbound messages that carry a verifiable diagnostic of the recipient's own asset will achieve higher positive-reply rates than messages that describe the service. The advantage will be larger for recipients whose attention is scarcer. Test: a randomised field exper‐ iment varying message type, stratified by recipient seniority. This is ethically straightforward when the dia‐ gnostic relies only on public information and every message has an unsubscribe link. • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
P4 (Proof-extending ladders). Free-to-paid conversion and paid retention will be higher when paid rungs extend the same proof (saved, recurring, delegated) than when they add features unrelated to the free product. Test: cross-sectional analysis of freemium ladders, coding each paid feature as proof-extending or not, against conversion and churn. P5 (Externality-gated autonomy). At equal model accuracy, ventures that gate irreversible and third-partyfacing actions will suffer fewer severe incidents than ventures that gate actions by technical difficulty or model confidence. The severity of autonomous errors depends more on where the gates are placed than on how accurate the models are. Test: incident analysis across autonomous systems, coding each incident by ac‐ tion quadrant and by whether a gate existed. P6 (Verification debt). In AI-built ventures, most defects found by line-by-line review will be silent rather than loud. The share of defects caused by decision drift will rise with the frequency of the owner's decision changes during construction. Test: review logs from multiple AI-built ventures, coded with the scheme in Table 6, against counts of decision reversals in their change histories. P7 (The shifting bottleneck). Under the Zero-Cash Operating Model, the binding constraint on growth shifts from cash and engineering capacity to owner attention and borrowed reputation (inbox standing, plat‐ form accounts, community trust). Returns to additional automation therefore fall, while returns to trust assets rise. Test: longitudinal tracking of ventures' time allocation, automation investment and growth, examining whether growth responds more to trust-building activity than to automation once the model is operating. Together, P1 and P2 concern capacity, P3 and P4 acquisition, P5 autonomy, P6 construction and P7 strategy.
Business model design. Business model research treats the business model as an activity system that the de‐ signer chooses [16] and as the realised consequence of strategic choices [20]. The case shows a mode of design in which much of the activity system is fixed before strategic choice. Quantified external rules, to‐ gether with one self-imposed rule, determined 21 of 24 major design decisions. What remained for strategy was mainly about trust and learning. This suggests that in the free-tier economy the designer's first task is to map the rules that bound the choice space, not the opportunities within it. The Constraint Ledger and the Constraint-Native Business Model Can‐ vas make that task explicit. They also respond to concerns about construct clarity [21][22]. They add to the canvas representation without redefining the business model itself. Resource-constrained entrepreneurship. The bootstrapping, bricolage and effectuation literatures picture constraints as scarcities to be improvised around [12][13][14]. The constraints in the case were of a different kind: generous, hard and quantified. They invited engineering against a budget rather than improvisation. The case also shows a founder imposing a stricter constraint than the provider's. This is a deliberate selfbinding that trades capacity for freedom from monitoring. In terms of the constraints–creativity relationship [15], a self-imposed headroom rule is a way of choosing one's position on the inverted U. It is tight enough to force efficient design, as the reduction from 120 to 11 row writes per audit shows, but not so tight as to stifle the product.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Automation and autonomy. The human-factors literature classifies automation by function and level [42], and argues for calibrated reliance [43]. The case shows that in a business setting the boundary of autonomy is set by consequences. Reversibility and externality to third parties decided where the human was needed, more than technical difficulty or model accuracy. The Autonomy Boundary Matrix turns this into a design tool, and Proposition P5 makes it testable. AI-assisted building. The productivity literature measures gains at the level of individual tasks [4][5][6] and warns of a jagged frontier [7] and of insecure output [17][18]. The case adds an organisational construct, verification debt, along with a mechanism, decision drift, and a set of measures. The composition of the debt (93% silent, about half with third-party exposure, a fifth caused by changed decisions) indicates that the risk of AI-built ventures lies less in code that fails visibly than in code that behaves plausibly but wrongly to‐ wards customers and the public.
For founders and those who design business models, the findings suggest a sequence: Write the ledger first. List every rule that bounds the model, in the provider's units, with a budget, an owner and a review date. Expect to add external rules during design. Choose a headroom ratio and two meters. Set H deliberately, meter at the point of consumption, crosscheck with the provider and govern by the higher figure. Decide in advance who is protected when a budget is reached. Design the free offer as proof. Make the free rung a diagnostic of the prospect's own asset, verifiable and specific. Admit paid features only if they extend that proof. Place every action in the autonomy matrix. Automate the reversible and private, gate the reversible and public, prepare the irreversible for a human, and keep the irreversible and public in human hands. Then check that the code's gates match. Budget for verification. Schedule line-by-line reviews, convert every finding into a test, and, when a de‐ cision changes, search everything for its traces. Hold capacity armed but dormant, and label it honestly. Accounts and reputation, not code, are now the bottleneck. For directors and boards overseeing AI-built or AI-operated products, the measures of verification debt (re‐ view yield, decay ratio, silent share, third-party share and drift share) offer a governance view that test cover‐ age alone does not. A product can pass every automated test and still hold most of its defects.
The study has clear limitations. One case, before launch. All findings come from a single venture before commercial launch. Nothing here shows that the business model attracts customers, converts them or retains them. Propositions P3, P4 and P7 depend on outcome data that did not exist. Short telemetry window. Three days at low and partly synthetic volume establish only the meter–pro‐ vider discrepancy and the order of magnitude of headroom. They do not show the governor's behaviour under load. 1. 2. 3. 4. 5. 6. • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
Author's involvement. The case was commissioned by the same owner who commissions the think tank. We have mitigated this by reporting defects, separating design targets from measurements and mak‐ ing no commercial claims. Readers should still weigh the findings accordingly. Defect coding. The coding of the 43 findings was done by the research team and not checked by an inde‐ pendent coder. The categories and the silent/loud distinction are offered as a scheme for others to replic‐ ate. Changing provider terms. Free-tier allowances, platform fees and sender rules change. All such figures are as published in October 2026 [55]–[62] and will date.
Four lines of research follow. Proposition tests. Each proposition (Section 6.8) comes with a suggested test. P3, a randomised field test of proof-first outreach, is the most immediately feasible and the most consequential for practice. Comparative studies of verification debt. These would apply the coding scheme in Table 6 to review logs from many AI-built ventures and establish typical silent, third-party and drift shares. Longitudinal follow-up of the case. This would cover its launch, its first customers and whether its bot‐ tleneck shifts to attention and reputation as P7 predicts. We intend to report this in a later working paper, including negative results. Constraints as strategy. This line would ask when a self-imposed constraint such as a headroom rule creates advantage, by enforcing efficiency and resilience, and when it simply forgoes growth. That ques‐ tion connects the constraint-native view to the dynamic capabilities tradition [63].
When building is cheap, the design of a business changes. The case of TightBucket, built in three days with no cash and run within a fifth of its free quotas, shows four things: Constraints can be the starting material of a business model rather than obstacles to it. Autonomy can be governed by measuring every unit, cross-checking with the provider, pacing work and degrading gracefully. The proper limit of autonomy is set by reversibility and externality. The speed of AI-assisted building creates a new liability, verification debt, which is mostly silent and of‐ ten borne by third parties. These findings come from one venture before launch, and they are offered as frameworks and propositions to be tested, not as conclusions about commercial success. We hope they help those who design digital busi‐ nesses to begin where the case began: by writing down the rules, choosing their headroom, and deciding with care what the machine should do alone. • • • 1. 2. 3. 4. • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-01
[1] Teece, D. J. (2010). Business Models, Business Strategy and Innovation. Long Range Planning, 43(2-3), 172–194. https://doi.org/10.1016/j.lrp.2009.07.003 [2] Zott, C., Amit, R., & Massa, L. (2011). The Business Model: Recent Developments and Future Research. Journal of Management, 37(4), 1019–1042. https://doi.org/10.1177/0149206311406265 [3] Osterwalder, A., & Pigneur, Y. (2010). Business model generation: A handbook for visionaries, game changers, and challengers. Wiley. [4] Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot (arXiv:2302.06590). arXiv. https://doi.org/10.48550/arXiv.2302.06590 [5] Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044 [6] Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2586 [7] Dell’Acqua, F., McFowland, E., Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, 37(2), 403–423. https:// doi.org/10.1287/orsc.2025.21838 [8] Jonas, E., Schleier-Smith, J., Sreekanti, V., Tsai, C.-C., Khandelwal, A., Pu, Q., Shankar, V., Carreira, J., Krauth, K., Yadwadkar, N., Gonzalez, J. E., Popa, R. A., Stoica, I., & Patterson, D. A. (2019). Cloud programming simplified: A Berkeley view on serverless computing (arXiv:1902.03383). arXiv. https://doi.org/10.48550/arXiv.1902.03383 [9] Simon, H. A. (1971). Designing organizations for an information-rich world. In M. Greenberger (Ed.), Computers, communications, and the public interest (pp. 37–72). Johns Hopkins Press. [10] Hevner, A. R., March, S. T., Park, J., & Ram, S. (2004). Design Science in Information Systems Research. MIS Quarterly, 28(1), 75–106. https://doi.org/10.2307/25148625 [11] Peffers, K., Tuunanen, T., Rothenberger, M. A., & Chatterjee, S. (2007). A Design Science Research Methodology for Information Systems Research. Journal of Management Information Systems, 24(3), 45–77. https://doi.org/10.2753/ mis0742-1222240302 [12] Baker, T., & Nelson, R. E. (2005). Creating Something from Nothing: Resource Construction through Entrepreneurial Bricolage. Administrative Science Quarterly, 50(3), 329–366. https://doi.org/10.2189/asqu.2005.50.3.329 [13] Winborg, J., & Landström, H. (2001). Financial bootstrapping in small businesses. Journal of Business Venturing, 16(3), 235–254. https://doi.org/10.1016/s0883-9026(99)00055-5 [14] Sarasvathy, S. D. (2001). Causation and Effectuation: Toward a Theoretical Shift from Economic Inevitability to Entrepreneurial Contingency. Academy of Management Review, 26(2), 243–263. https://doi.org/10.2307/259121 [15] Acar, O. A., Tarakci, M., & van Knippenberg, D. (2019). Creativity and Innovation Under Constraints: A CrossDisciplinary Integrative Review. Journal of Management, 45(1), 96–121. https://doi.org/10.1177/0149206318805832 [16] Zott, C., & Amit, R. (2010). Business Model Design: An Activity System Perspective. Long Range Planning, 43(2-3), 216–226. https://doi.org/10.1016/j.lrp.2009.07.004 [17] Pearce, H., Ahmad, B., Tan, B., Dolan-Gavitt, B., & Karri, R. (2022). Asleep at the Keyboard? Assessing the Security of GitHub Copilot’s Code Contributions. In 2022 IEEE Symposium on Security and Privacy (SP) (pp. 754–768). https://doi.org/10.1109/sp46214.2022.9833571 [18] Perry, N., Srivastava, M., Kumar, D., & Boneh, D. (2023). Do Users Write More Insecure Code with AI Assistants? In Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security (pp. 2785–2799). https://doi.org/10.1145/3576915.3623157 Courage Garikai Coldcode Si Labs Library Working Paper 2026-01