Externalities, borrowed identities and duplicate clocks in an autonomous growth pipeline Courage Garikai · Funded by Coldcode Si Labs Disclosure. The case venture, Tapstall, belongs to the same owner who commissions the work of Coldcode Si Labs. The audit and rebuild analysed here were carried out by the owner working with AI coding agents. The same analytic process produced the audit, the coding of its findings and this paper, so the case is not an independent evaluation. At the time of writing Tapstall had no paying customers, and nothing here is evidence of commercial results. This paper does not say how the first version of the venture was produced; where the evidence bears on that question, it is treated as a hypothesis. AI tools assisted in drafting this paper, and the author is responsible for its content. Individuals and businesses who received messages from the sys‐ tem are not named.
Small ventures can now run growth pipelines that act without a person in the loop. A scheduled job finds pro‐ spects, another writes to them, a third posts to social accounts, and a payment webhook switches accounts on. Each component can be correct by its own tests while the pipeline as a whole does things that nobody sees. This paper names that condition unwitnessed autonomy: an automated system that takes consequential ac‐ tions on parties outside its boundary when no party is placed to observe those effects and feed them back to the operator. The paper reports a design-science case study of Tapstall, a shop-link service for informal and social-com‐ merce merchants in Zimbabwe and South Africa. Its growth pipeline ran on free cloud tiers with no paid staff. On the day of the audit, the venture's own test suite passed 420 of 420 tests against the live code. In that same code: Of the 17 distinct addresses the outreach runner had already mailed, 14 were not prospects at all. They included a magazine reviewer, a platform's support desk, a government weather service and a literal placeholder address. These messages went out from the owner's personal mailbox, which was also the support inbox of a sep‐ arate venture, and the runner had been marking that venture's support mail as read. The press module had posted three times a day into the other venture's social account. An old copy of the service on the same account was answering writes against production data with a su‐ perseded secret. Two more copies on a second account ran their own scheduled jobs. Every merchant email the product promised had been a silent no-op since launch. The public facts printed on a storefront were enough to take over that store. • • • • • •
Coldcode Si Labs Library WORKING PAPER 2026-03
Coldcode Si Labs Library Working Paper 2026-03 · First edition · 2026 The audit and the rebuild logged 24 defects in total. Coding them shows that 16 (67%) were witness gaps of four kinds: the effect of an action was never observed; the recipient's position was never observed; the identity the system acted under was not its own; the number of clocks and copies running was unknown. None of the 24 was found by the existing test suite. Four of the 13 defects found during the rebuild were in‐ troduced by the fixes themselves, and each was caught by a live dry run or by inspecting real output, not by a unit test. The paper contributes: a definition of unwitnessed autonomy that separates it from ordinary software defects, technical debt and spam; a taxonomy of nine defect classes, of which five are witness gaps; a witness map, an audit instrument that asks five questions of every autonomous action; four design rules for small autonomous pipelines: one clock, one voice, one identity, one surface; seven propositions for testing. The case is a single pre-revenue venture analysed by an interested party, and the propositions are offered for testing. Keywords: unwitnessed autonomy; automation; externalities; growth pipelines; outreach; email; identity coupling; distributed clocks; silent failure; informal economy; design science
A founder with no staff and no budget can now run a working company's growth machinery: prospect lists, first-touch emails, follow-ups, social posting, payment activation, renewal reminders. Free cloud tiers supply the compute, scheduled CI runners supply the minutes, and a social-media scheduler or a mailbox app pass‐ word supplies distribution. The marginal cost of another automated action is close to zero. What the founder gains in reach, the founder loses in sight. A person who sends twenty emails by hand reads twenty addresses. A pipeline that sends twenty emails a day reads none of them, unless someone has built it to. This paper is about what happens in that gap. Its starting observation is unremarkable to anyone who has run production systems: a system can pass its own tests and still behave badly. The specific concern is narrower. It is the class of failures where an automated system acts on someone outside itself (a stranger's inbox, anoth‐ er brand's social account, a payment, a merchant's livelihood), and the consequences of the action never re‐ turn to anyone who could correct them. I call this condition unwitnessed autonomy. Its defining feature is not that the system is wrong but that nobody is placed to notice that it is wrong. The harm lands on parties who have no channel back to the operator, while the operator's own surfaces (dashboards, logs, tests) report suc‐ cess. • • • • • • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
The literature on automation has studied the human operator's loss of awareness for four decades, from Bain‐ bridge's "ironies of automation" [1] to Parasuraman and colleagues' work on misuse and levels of automation [2], [3]. Safety research has studied how organisations come to accept anomalies as normal [4] and how missing feedback loops make control structures unsafe [5]. Economists have long understood that costs borne by third parties are underpriced by those who impose them [6], [7], and email has been a canonical example since the economics of spam were measured [8]. Security research named the "confused deputy", a program that uses authority it holds for one purpose on behalf of another [9]. Distributed-systems research gave us the vocabulary of clocks, ordering and duplicate execution [10], [11]. Machine-learning engineering added "un‐ declared consumers" and hidden feedback loops as forms of debt [12]. What these literatures have not done is bring the strands together for the setting that cheap automation has now made common: the micro-venture whose growth pipeline is fully autonomous, runs on shared free infrastructure and is operated by one person who reads none of its output. The paper asks three questions: What forms does unwitnessed autonomy take in a small autonomous growth pipeline? Why does the operator's ordinary evidence (a green test suite, a dashboard that says "active", a run marked successful) fail to reveal them? What design rules and audit instruments make the effects of autonomous action visible again, at no cash cost? It answers them through a design-science case study [13] of Tapstall, a service that gives informal and socialcommerce merchants in Zimbabwe and South Africa a shop link: a menu page from which buyers order over WhatsApp and pay the merchant directly. Tapstall's growth pipeline was built to run with no person in the loop. A scheduled job hunted for prospective merchants, a server judged them and allocated each to one com‐ missioned seller, an outreach runner emailed and followed up, and a press module posted to social rails three times a day. On 28 October 2026 its owner commissioned a read-only audit of the live system (version 1.9.1), followed the same day by a rebuild to version 2.0.0. The audit, the rebuild and their verification pro‐ duced a dataset of 24 defects, together with the live records that exposed each of them: lead rows, sent-mail records, run logs, the credential inventory of two cloud accounts, and dry-run outputs. The findings are uncomfortable in a specific way. The pipeline was not broken in the sense that matters to its own instruments. Its tests passed. Its runs completed. Its dashboard reported rails as configured. What it did to the outside world was a different matter, and the record of what it had done was scattered across a mail‐ box, a queue of rows, an error string stamped "done" each night, and a social account that belonged to some‐ body else. The dominant pattern was not a wrong computation. It was the absence of anyone positioned to see the computation's effect. The paper makes five contributions. A definition. Unwitnessed autonomy is distinguished from three neighbouring ideas. It is not just a software defect, because the defining harm falls outside the system's boundary and outside its tests. It is not technical debt, because debt is a cost the owner pays later, whereas unwitnessed autonomy imposes costs on others now. It is not spam, though it can produce spam, because no intent to send unwanted mail is required and the same structure shows up in payments, social posting and account security. 1. 2. 3.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
A taxonomy. The 24 defects sort into nine classes. Five of them are witness gaps, failures of observation rather than of computation: externality; borrowed identity; multiplicity of clocks and copies; silent failure; boundary interpretation. The remaining four are more familiar: authority leakage, surface sprawl, dependency capacity and build de‐ fects. The witness gaps account for two-thirds of all defects and six of the nine critical- and high-severity audit findings. The other three sit in the authority and surface classes, and every critical or high finding con‐ cerns an effect beyond the function that caused it. An instrument. The witness map asks five questions of every autonomous action: Under whose identity does it act? On which clock does it run, and how many copies of that clock exist? What evidence returns that it had its intended effect? What evidence returns about the recipient's position? What stops it? Applied retrospectively, the questions would have surfaced every witness gap in the case. Design rules. Four rules for small autonomous pipelines follow from the case: One clock. A single scheduler fires every job, and a failed fire is retried, not stamped. One voice. A recipient hears from one sender identity, and every message carries a working exit. One identity. The system acts only under credentials that belong to it. One surface. There is one place where actions on prospects are taken, so there is one door to guard. Propositions. Seven propositions state how unwitnessed autonomy arises, why ordinary evidence misses it and what governs it. They are framed for testing in other ventures. The paper also documents something that the case makes unusually visible: four of the 13 defects found dur‐ ing the rebuild were introduced by the fixes. A gap filled in one place opened another, as when removing a spreadsheet surface left no surface at all for working prospects. The corrected prospect hunter's first live run filed branches of supermarket chains and their corporate addresses. Each of these was caught by observing real output, not by tests. The implication is not that fixes are dangerous. It is that witnessing has to be a standing practice rather than a one-off audit, because the act of repair produces new unwitnessed actions. The rest of the paper proceeds as follows. Section 2 reviews the relevant literatures and locates the gap. Sec‐ tion 3 describes the method and the case. Section 4 reports the findings: what the pipeline had done, the ana‐ tomy of its defects, how they were found, the rebuild, and the defects the rebuild introduced. Section 5 devel‐ ops the definition, taxonomy, witness map, design rules and propositions. Section 6 discusses implications, limitations and directions for research, and Section 7 concludes. • • • • • • • • • • • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
Five bodies of work bear on the problem: automation and human oversight; safety science; the economics of externalities and the email commons; security and the authority of programs; and distributed-systems and software engineering on clocks, copies and hidden dependencies. Each supplies part of the explanation. None, on its own, describes the condition this paper studies.
Bainbridge's "ironies of automation" [1] remains the sharpest statement of the problem. Automating a task re‐ moves the operator from the routine work that kept them competent and aware, then asks them to take over at exactly the moments the automation cannot handle, which are the moments that need the most skill and con‐ text. The more reliable the automation, the less practised the human, and the less likely they are to notice when something has gone quietly wrong. Bainbridge wrote about process control rooms. The irony transfers directly to the solo founder whose outreach runs every weekday at 06:40: the founder stops reading outbound mail precisely because the system sends it. Parasuraman and Riley [2] distinguished four relationships between people and automation: use, misuse (over-reliance), disuse (neglect) and abuse. Abuse here means automating functions without regard for the consequences for the humans involved. Their concept of abuse is closest to this paper's concern, but they framed it from the designer's side, as automating whatever is technically possible and leaving the operator to cope. In the micro-venture, designer and operator are the same person, and the people left to cope are third parties. Parasuraman, Sheridan and Wickens [3] later offered a model of types and levels of automation across four stages: information acquisition, analysis, decision and action. Their model makes a point that matters for growth pipelines. A system can be highly automated at the action stage (it sends) while the in‐ formation-acquisition stage (did it land, did the recipient want it) is not automated at all and not done by any‐ one. This literature establishes that automation degrades human awareness and that the degradation is structural rather than a matter of diligence. It assumes, however, an operator who is supposed to be monitoring and a system whose outputs are, in principle, visible on a display. Unwitnessed autonomy arises where neither as‐ sumption holds. The outputs land in other people's inboxes and accounts, and no display was ever built to show them.
Safety science adds an organisational account of how hazards accumulate. Vaughan's study of the Challenger launch decision [4] introduced the normalisation of deviance. An anomaly that does not immediately cause disaster is gradually reclassified as acceptable, and each acceptance makes the next easier. The mechanism needs no negligence. It needs only an anomaly that does not announce itself loudly. A nightly job whose error string is recorded but whose run is still marked "done" is an anomaly of exactly this kind. Perrow's Normal Accidents [17] argued that systems combining interactive complexity (components interact‐ ing in unplanned ways) with tight coupling (little slack between steps) will have accidents that no component failure explains. Reason's "Swiss cheese" model [18] describes how defences fail when holes in successive layers line up. Weick and Sutcliffe's study of high-reliability organisations [19] identifies preoccupation with failure and sensitivity to operations as the habits that let some organisations catch weak signals before they
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
compound. Dekker [20] extended Vaughan's insight into "drift into failure": systems migrate towards the boundary of safe performance through locally reasonable steps. Cook's short treatise on how complex sys‐ tems fail [21] adds the observation that complex systems run in a degraded mode as a matter of course, carry‐ ing latent failures that become visible only in combination. Leveson's systems-theoretic approach [5] is the most directly applicable. It treats safety as a control problem: an accident results from inadequate control, and control requires a feedback channel by which the controller learns the state of the process. A controller with no feedback, or with feedback that reports the controller's own actions rather than the process state, cannot be safe however correct its control logic is. Translated to a growth pipeline, a scheduler that records "dispatched" rather than "the outreach ran" has feedback about it‐ self, not about the world. Safety science thus explains why latent problems accumulate and why feedback is the decisive element. Its settings are large organisations with dedicated safety functions, and it says little about systems run by one person whose "organisation" is a set of scheduled jobs.
Economics supplies the vocabulary for harms that fall on others. Pigou [14] described costs that an activity imposes on third parties without compensation. Coase [6] reframed such costs as a problem of reciprocal harm and transaction costs: where bargaining is cheap, parties will negotiate an efficient outcome, and where it is not, the allocation of rights matters. Hardin's "tragedy of the commons" [7] described the over-use of a shared resource by actors who each capture the benefit of use while sharing its cost. Ostrom [16] showed that commons can be governed by rules communities devise themselves, provided monitoring and graduated sanctions exist. Email is a commons in Hardin's sense and a governed commons in Ostrom's. Sending is nearly free to the sender, and the recipient bears the cost of attention. Kanich and colleagues' Spamalytics study [8] measured the economics directly by infiltrating a spam botnet. Conversion rates were tiny (one sale per many millions of messages), yet the activity stayed profitable because sending cost almost nothing. Levchenko and col‐ leagues [15] traced the spam value chain end to end and showed that the scarce resources were payment pro‐ cessing and a few infrastructure providers, not the sending itself. The commons responded with rules. Mail standards define bounce notifications [30], automatic-response conventions [33], list-management headers [31] and a one-click unsubscribe mechanism [32]. Domain authentication allows receivers to judge senders [34]. Large mailbox providers announced in 2023 that bulk senders would have to authenticate, offer oneclick unsubscribe and keep complaint rates low [35]. Law sets limits on unsolicited direct marketing. South Africa's Protection of Personal Information Act makes electronic direct marketing to non-customers condi‐ tional on consent [36]. The United States' CAN-SPAM Act requires a working opt-out and a postal address [37]. Zimbabwe's Cyber and Data Protection Act regulates the processing of personal information [38]. This literature explains why badly targeted outreach is a harm and how the commons polices it. It treats the sender as an actor making a choice about volume and targeting. The case here involves a sender that made no such choice. A pipeline component took the first address on a web page, another counted a URL as evidence, and a third sent. The externality was produced by composition, not by decision.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
Saltzer and Schroeder's design principles for protection [22] include least privilege (every program should operate with the minimum authority it needs) and fail-safe defaults (access should be denied unless explicitly granted). A system that signs its capability tokens with a fallback key when its real secret is missing violates the second. A system that sends one venture's marketing from another venture's support mailbox violates the first. Hardy's "confused deputy" [9] describes a program that holds authority for one purpose and is induced to exercise it for another. The canonical example is a compiler that writes a billing file it has rights to on a user's behalf. In the growth pipeline the deputy is not tricked by an attacker. It was simply given someone else's credentials, because they were the credentials that were to hand. Anderson's survey of security engin‐ eering [23] documents how often real failures arise from such mismatches between the authority a compon‐ ent holds and the purpose it serves, rather than from cryptographic weakness. Agency theory frames a related problem in economic terms. Jensen and Meckling [24] described the costs of delegating to an agent whose actions the principal cannot fully observe: monitoring costs, bonding costs and residual loss. An automated pipeline is an agent with no interests of its own, so the classic divergence of in‐ terest is absent. The monitoring problem remains, and the principal who does not pay monitoring costs bears residual loss, or imposes it on others.
Distributed-systems research has a precise vocabulary for one class of failure seen in the case. Lamport [10] showed that in a distributed system without a shared clock, the ordering of events must be constructed rather than assumed. Practical systems design treats duplicate execution as the normal case. Messages are delivered at least once, so operations must be idempotent if they are not to take effect twice [11]. Site reliability prac‐ tice [28] treats a job that runs on two schedulers, or two copies of a service running against the same data, as an incident class in its own right ("split brain"). Sculley and colleagues' account of hidden technical debt in machine-learning systems [12] identified un‐ declared consumers: components that silently depend on a system's outputs, so that a change to the system has effects nobody anticipated. They also identified hidden feedback loops, where a system's outputs shape its future inputs. Both have direct analogues in a growth pipeline. A shared mailbox has undeclared con‐ sumers (another venture's support staff), and an outreach runner that reads replies to decide whether to follow up has a feedback loop that misreads its own footer as a reply. Parnas's criterion for decomposing systems into modules [46] (hide each design decision likely to change behind an interface) explains why credentials and schedules embedded in several places become a liability. Conway's observation that systems mirror the communication structures of the organisations that build them [48] predicts that a one-person venture run‐ ning several products will couple those products wherever the person reuses an account.
Two further strands complete the picture. In AI safety, Amodei and colleagues [26] listed avoiding negative side effects as a concrete problem: an agent optimising a task objective has no reason to avoid disturbing parts of the environment the objective does not mention. A growth pipeline optimising "emails sent per weekday" has no term for "addresses that were not prospects". Manheim and Garrabrant's taxonomy of Goodhart effects [27], and Strathern's general statement of Goodhart's law [47], explain why the pipeline's own counters (leads filed, messages sent, posts made) become misleading the moment they are treated as tar‐
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
gets. Institutional theory supplies the idea of decoupling [25]: formal structures adopted for legitimacy that are not connected to the activities they describe. A dashboard that lists social rails as configured while the ac‐ count behind them belongs to someone else is a small instance.
The case sits in a specific market. Informal economic activity is large across sub-Saharan Africa: IMF estim‐ ates put the shadow economy in the region among the largest shares of output anywhere [41]. Merchants selling through WhatsApp and social media are, for the most part, invisible to commercial business director‐ ies. Volunteered geographic information, notably OpenStreetMap, is one of the few open sources that records them. Its quality is variable and spatially uneven [39], [40], so a prospect hunter that relies on it inherits both its reach and its gaps.
Finally, a growing literature studies AI-assisted software development. Controlled studies have found large speed-ups on bounded tasks [43]. A randomised trial with experienced open-source developers found that AI assistance made them slower while they believed it had made them faster [42]. The gap between perceived and actual effect is relevant here only as context. This paper does not say how the first version of Tapstall was produced. It notes that a mode of building in which large amounts of plausible, individually tested code appear quickly would be expected to widen the distance between a system's components and anyone's view of the whole. That is a hypothesis the evidence is consistent with, not a finding.
Taken together, these literatures explain the pieces: why operators stop watching automated systems [1], [2]; why anomalies are normalised [4], [20]; why missing feedback makes control unsafe [5]; why third-party costs are over-produced [6], [7], [14]; why borrowed authority is dangerous [9], [22]; why duplicate clocks and hidden consumers cause surprises [10], [11], [12]; why optimised systems produce side effects [26]. What is missing is an account of how these mechanisms combine in the now-common setting of a fully autonomous micro-venture: one person, free shared infrastructure, several products, and a growth pipeline that acts on strangers every day. There is also no instrument that such a person can apply at no cost to find the resulting failures. Section 5 proposes one, grounded in the case reported in Section 4. • • • • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
The study follows design-science research in information systems [13]. An artefact, here an operating growth pipeline, is examined, problems are identified, the artefact is redesigned and the redesign is evaluated against the problems that motivated it. Design science suits the question because the aim is both explanatory (why did these failures arise and evade detection?) and prescriptive (what instruments and rules prevent them?). The unit of analysis is the pipeline as deployed, not its source code in isolation. The distinction matters, be‐ cause several defects existed only in the relation between the code and its deployment: which accounts held which credentials, how many copies were running, and what the scheduled jobs had actually done. The case is a single, revelatory case in Yin's sense [45]. Its circumstances made visible a phenomenon that is usually hard to observe, because the full live state of a small autonomous venture (its data stores, run logs, mailboxes and cloud accounts) was available to the analyst in one place on one day. Theory building from a single case follows Eisenhardt's approach [44]: constructs are developed from the data, sharpened by com‐ parison across the findings and stated as propositions for testing elsewhere, not claimed as general.
Five kinds of evidence were collected on 28 October 2026. Live state. Read-only exports of the venture's key-value stores: the full lead table (109 rows at the start of the audit, 140 by the time of the clean-up, because a scheduled hunt on the old code ran in between), the mer‐ chant and seller registries, the press registry and posting ledger, the dispatch record and mail-related keys. Operator console dumps (health, state, leads, sellers, audit log) were saved before any change. Run logs. Logs of the scheduled GitHub Actions workflows for prospect hunting and outreach over the pre‐ ceding weeks, including the outreach runner's sent and skipped decisions, and the logs of verification runs made after the rebuild. Mail evidence. The list of addresses the outreach runner had mailed (17 distinct addresses), the sending ac‐ count's identity, and the runner's handling of replies and bounce notices, reconstructed from the runner's code and its run logs. Message contents were examined only to the extent needed to classify recipients. Recipients are not named in this paper. Credential and deployment inventory. An inventory of two cloud accounts and one source-hosting account associated with the venture. It recorded every deployed service, its version, its bindings to data stores, its scheduled triggers and the names (not values) of its secrets. This inventory exposed the copies of the service running beyond the production deployment and the credentials that belonged to another venture. Code and tests. The repository at the audited commit, its test suite (420 tests, all passing), and every change made during the rebuild, recorded as commits with messages. The final suite after the rebuild contained 421 tests plus self-tests for the two runner scripts.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
Each defect was recorded when found, with its severity, the evidence that exposed it, the component involved and whether it existed before the audit or was introduced during the rebuild. Two items recorded in the audit notes were judged not to be defects, a token whose scope was appropriate and a once-daily heartbeat that was daily by design, and they are excluded from the counts. That leaves 24 defects: 11 from the read-only audit and 13 found during the rebuild. Coding proceeded in two passes. In the first, each defect was described in terms of what it did and to whom, without reference to any framework. In the second, the descriptions were grouped inductively by the mech‐ anism of harm. Nine classes emerged. Five of them share a property: the defect consists in the absence of an observation (of an effect, a recipient's position, the identity being used, or the number of clocks and copies), not in a wrong computation. These were labelled witness gaps. The other four classes (authority leakage, sur‐ face sprawl, dependency capacity and build defects) are familiar software-engineering categories. Severity followed a simple scale. Critical meant harm to third parties, or to the venture's control of its own data, that was occurring or immediately possible. High meant harm that was occurring but bounded, or a broken promise to customers. Medium covered the rest. The method of discovery was recorded for each de‐ fect against six categories: reading live state; inspecting sent mail and recipients; inventorying credentials and deployments; adversarial walk-through; live dry runs; and code reading. A defect could be exposed by more than one method, and the primary method is reported. The analyst was the same party that designed the rebuild. To limit the obvious bias, a defect was counted only if it could be tied to a specific artefact: a row, a log line, a deployed service, a line of code or a dry-run output. Interpretations of why a defect arose were kept separate from the record of what it was. Claims about how the first version was produced are not made.
The venture. Tapstall gives merchants who sell through WhatsApp and social media a shop link. The link opens a menu page with prices. Orders arrive in the merchant's WhatsApp, and the buyer pays the merchant directly through a mobile-money rail, a bank transfer or similar, quoting a reference. Tapstall never holds the money. It matches payment references and keeps the order record. The service is free for 60 days and then costs USD 60 a year, payable by card through a third-party checkout or by reference payment. The target market is informal and semi-formal sellers in Zimbabwe and South Africa: cooked-food vendors, small shops, repair stalls and similar businesses that commercial platforms do not serve well. At the time of the audit the venture had no paying customers. The infrastructure. The service ran on a serverless platform's free tier: one worker script, two key-value namespaces and scheduled triggers. Its pages were bundled into the worker. Scheduled jobs that needed more time or open-web access ran as GitHub Actions workflows in a public repository, where minutes are free. Three free-tier ceilings shaped the design and recur in the findings: at most 1,000 key-value writes, deletes or list operations per day; at most 50 outbound subrequests per worker invocation; 100,000 requests per day. • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
The growth pipeline. As designed, the pipeline had five stages and no person in the loop: Hunting. A workflow searched the open web and an open map database for merchants and posted candid‐ ate facts (business name, address, phone, email, source page) to the server. Judging and allocation. The server derived "signals" from each candidate, required two signals before a lead entered the outreach queue, deduplicated leads, checked a do-not-contact ledger and allocated each lead by hash to exactly one commissioned seller. Outreach. A workflow fetched eligible leads, sent first touches and up to two follow-ups by SMTP from a configured mailbox, and read replies over IMAP to classify them (interested, not now, stop) and record outcomes. Sellers. Commissioned sellers worked their allocated leads, originally through a spreadsheet surface fed by the server. Press. A module in the server composed short posts from the venture's facts and sent them to registered social accounts three times a day. A payment webhook switched merchants' plans on, and the server sent merchants transactional mail: wel‐ come, order alerts, renewal warnings and PIN resets. The operator. The venture had one owner, who ran several other ventures from the same accounts. The other ventures matter to the case because they shared infrastructure with Tapstall: a mailbox, a social account and cloud accounts. The owner did not read the pipeline's outbound mail or social posts as a matter of routine. That was the point of automating them. The intervention. The audit began read-only. Live writes were made only to contain harm already occurring: marking junk leads do-not-contact so they could not be mailed again, and rotating the administrative secret, which had been shared with a superseded copy of the service. The rebuild then ran in the venture's repository with the full test suite after each change. Changes reached production through the existing continuous-de‐ ployment workflow, and post-deployment state changes were made through the cloud provider's API: delet‐ ing borrowed credentials, replacing superseded copies with redirects and purging junk rows. Every change was verified live, by health endpoints, a real test message delivered and checked in the receiving mailbox, and dry runs of the rebuilt hunter on the same CI runners that run it in production.
On the morning of the audit, every instrument available to the operator reported a working pipeline. The test suite passed 420 of 420 tests. The most recent scheduled runs of the outreach and hunting workflows were marked successful. The operator console showed the press module with a registered company identity and a social rail. The lead table held 109 prospects, nine of them already contacted, 34 queued for outreach and 66 held for research. The service's own self-test, which runs inside the production runtime, passed. Nothing on any surface the operator looked at suggested a problem.
The live record told a different story. Six observations from the first hours of the audit set the pattern. 1. 2. 3. 4. 5.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
The outreach runner had mostly written to people who were not prospects. It had mailed 17 distinct ad‐ dresses. Three belonged to small businesses of the kind Tapstall serves. The other 14 received 16 messages between them: a technology magazine's reviewer; the support desk of a global social-media platform; a financial-data company's support address; the webmaster of a government weather service; privacy and reprints desks at two financial publishers; a Japanese games company; an inventory-software vendor's support; an events-listing site; the editor of a Zimbabwean news site (twice); a private individual's personal address (twice); the literal placeholder user@domain.com. The hunter had searched the open web, taken generic result pages as candidate businesses, and taken the first email address on each page as the business's contact. The server's signal rules then counted the existence of a source URL as one signal and words in the URL as another, so two "signals" could be manufactured from a single irrelevant page. Those messages went out under another venture's identity. The sending mailbox was the owner's personal webmail account, which was also the published support address of STPi, a separate venture run by the same owner. To process replies, the outreach runner logged in to that inbox over IMAP and fetched messages in a way that marked them as read. STPi's support mail was therefore being opened by a machine and flagged as seen before any person saw it. The press module posted into the other venture's social account. The company identity in the press re‐ gistry named STPi's Bluesky handle, and the corresponding credentials were bound to Tapstall's server. From 20 October to the day of the audit, Tapstall marketing copy ("shop math", "sixty dollars a year") was posted to STPi's followers three times a day. An archive document from an earlier, unrelated clean-up had attributed these posts to a spreadsheet script. The attribution was wrong, which is itself a measure of how little anyone could see of what the pipeline was doing. A superseded copy of the service was serving production data. The cloud account that hosted production also hosted an older alias deployment (version 1.0.0). It was bound to the same production data stores and held the previous administrative secret. Its read pages redirected to production, but every write (settings changes, PIN resets, orders, administrative actions) was handled by the old code against live data. A second cloud account held two further copies at version 0.9.1, each with its own data store and its own scheduled trigger firing four times a day. • • • • • • • • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
The nightly dispatch had been failing and recording success. The server's main scheduled job ran pro‐ spect discovery first and then tried to trigger the hunting and outreach workflows. On the free tier the discov‐ ery step consumed most of the 50 permitted outbound requests, so the trigger failed with "Too many subre‐ quests". The job nonetheless stamped the day as dispatched. Nobody noticed, because the two workflows also had their own schedules and ran anyway. The duplicate clock masked the failed one. Every merchant email had been a silent no-op since launch. The code that sent welcome messages, order alerts, renewal warnings and operator notices called a mail-sending interface that had no binding in the de‐ ployed configuration. Each call did nothing and raised no error. The renewal-warning path stamped each store as "warned" after the call. A merchant whose free period was ending would therefore be suspended on schedule, having never received the warning the system recorded as sent.
Table 1 lists all 24 defects with their classes. Eleven were found in the read-only audit (labelled C, H and M by severity) and 13 during the rebuild (labelled R). Four rebuild defects, R9 to R12, were introduced by the rebuild's own changes.
ID Defect Class Severity C1 Public facts (handle, number shown on storefront, account shown to buyers) sufficed to reset a store's PIN Authority leakage Critical C2 Hunter filed generic web pages as prospects; first email on page taken; URL counted as evidence; runner mailed them Externality Critical C3 Superseded copy served writes against production data with the old secret Multiplicity Critical H1 Merchant mail a silent no-op; renewal warning stamped as sent Silent failure High H2 Press posted into another venture's social account three times a day Borrowed identity High H3 Nightly dispatch exceeded request limit, failed and was stamped done Silent failure High H4 Bounces ignored; no list-unsubscribe header; third touch had no opt-out Externality High H5 Card payments matched on unverified email; unmatched payments filed silently Authority leakage High H6 Spreadsheet CRM as a second surface able to propose lead actions Surface sprawl High M1 Whole-country map query with no contact filter timed out Dependency capacity Medium M2 Fallback signing key used if the real secret was ever unbound Authority leakage Medium R1 Runner read another venture's support inbox and marked mail read Borrowed identity High R2 "STOP" inside quoted text (the venture's own footer) read as an opt-out Boundary interpretation Medium R3 Outreach and hunting workflows fired by two clocks; daily send cap effectively doubled Multiplicity High R4 Two superseded copies on a second account with own data and scheduled jobs Multiplicity Medium R5 Follow-ups could come from a different mailbox than the first touch Externality Medium R6 Junk rows queued before the new gate bypassed it Boundary interpretation Medium R7 Press registry still named the other venture's handle after its secrets were removed Borrowed identity Medium Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
ID Defect Class Severity R8 Press composed and queued posts for rails with no account Silent failure Medium R9 Removing the spreadsheet left no surface for working leads Surface sprawl High R10 Rebuilt hunter's first live run returned chain-store branches with corporate addresses Externality Medium R11 Rebuilt hunter's 18-part map query timed out on every public mirror Dependency capacity Medium R12 Removing a duplicate deploy loop left an undefined variable under strict shell mode Build defect Medium R13 Health surface did not report mail, mail queue, dispatch result or unmatched payments Silent failure Medium
Note. C, H, M: audit findings by severity (critical, high, medium); R: found during the rebuild; R9–R12 were introduced by the rebuild. Table 2 summarises the classes. Five are witness gaps. Together they hold 16 of the 24 defects (67%).
Class Witness gap? What is missing Count Externality Yes: recipient Any observation of who receives an action and whether they should Silent failure Yes: effect Any observation that an action had its effect Borrowed identity Yes: identity Any check that the identity acted under belongs to the system Multiplicity Yes: clock Any knowledge of how many clocks and copies are running Boundary interpretation Yes: recipient Correct reading of what comes back from outside Authority leakage No — Surface sprawl No — Dependency capacity No — Build defect No — Total 16 witness gaps
Three features of the distribution deserve comment. First, the severe defects concern effects beyond the function that caused them. Of the nine critical and high findings in the audit, six are witness gaps (C2, C3, H1, H2, H3, H4) and three are authority or surface defects (C1, H5, H6). Every one involves a party or resource outside the code unit: a stranger's inbox, another ven‐ ture's account, another copy of the service, a merchant's store, a buyer's payment. None is a computation that gives a wrong answer for a given input. Second, borrowed identity and multiplicity existed only in the deployment, not in the code. No reading of the repository would have revealed that the sending mailbox was another venture's support address, that a social handle belonged to someone else, or that three other copies of the service were running. These facts lived in account settings and secret stores.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
Third, the boundary-interpretation defects show the same failure in reverse. The pipeline misread what came back from outside. A reply quoting the venture's own footer, which contained the word "STOP", was classi‐ fied as an opt-out, so a prospect who replied with interest could be marked do-not-contact. Bounce notifica‐ tions from mail servers were classified as "ignore", so dead addresses continued to receive follow-ups.
None of the 24 defects was found by the existing test suite, and the suite passed throughout. This was not be‐ cause the tests were careless. The suite was extensive: it exercised registration, ordering, fee arithmetic, pay‐ ment-reference matching, seller allocation, anti-collision on lead claims and the press module's composition rules, among much else. It could not see the defects for structural reasons that map onto the witness-gap classes. Externalities are invisible to a test because the test supplies the inputs. A test of lead ingestion supplies a well-formed lead. It does not supply the page that the hunter actually fetched from a search engine on a CI runner on a particular morning. Silent failures pass tests when the test environment provides what production lacks. The mail interface was stubbed in tests and absent in production. The dispatch step's request budget is not enforced in a test runner. Borrowed identity and multiplicity are properties of the deployment, which tests do not model. Boundary interpretation fails where the test fixtures were written by the same mind as the classifier. A fixture reply that says "STOP" tests the rule. A real reply that quotes the footer does not appear in the fix‐ tures until someone reads real replies. Table 3 shows the primary method that exposed each class.
Primary method Defects Reading live state and ledgers C2, C3, H3, R3, R6 Inspecting sent mail and recipients H4, R2, R5 Credential and deployment inventory H2, R1, R4, R7 Adversarial walk-through C1, H5, M2 Live dry runs R10, R11 Code reading H1, H6, M1, R8, R9, R12, R13 Existing test suite none
The methods that found the witness gaps share a property. Each looks at what the system did or is, not at what its code says it does: the rows it wrote, the addresses it mailed, the accounts it holds, the output of a real run.
The rebuild to version 2.0.0 took the rest of the day. Its changes are summarised by the gap each addressed. • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
Recipient witness. A junk gate now runs at ingestion and again at the claim door, so legacy rows are also re‐ fused. A lead's email must belong to the business: same domain as its website, or a free-mail address appear‐ ing on the business's own page. The lead must sit in its market, and platform and reference pages are rejected outright. The hunter was rebuilt around the open map database, querying two cities per country per day with a contact filter, and enriching from the business's own website only. Outreach is limited to three touches. Every touch carries an unsubscribe link signed by the server, list-unsubscribe headers [31], [32] and the sender's postal address [37]. Replies are classified on the prospect's own words only, with quoted text stripped. Bounce notifications [30] and refused recipients mark the lead do-not-contact. The 32 junk rows were marked do-not-contact, and 89 research-hold rows were deleted: thesaurus entries, weather forecasts, homework-help pages, a foreign pizza menu and an encyclopedia article. Identity witness. Outreach and transactional mail now go out from one dedicated mailbox belonging to the venture, through an SMTP lane in the server [29]. The lane has an hourly cap set as a share of the provider's limit (the box is shared with other pipelines, and Tapstall's share is two messages an hour, one of which is al‐ ways reserved for account mail). A retry queue holds messages that cannot leave yet. A refused login parks the lane for six hours rather than retrying, because repeated failures can lock the account. Replies are read in read-only mode, so nothing is marked as seen. The other venture's social credentials were deleted, the press registry was rewritten without the borrowed handle, and the personal mailbox was no longer used for send‐ ing. Clock witness. The workflows' own schedules were removed, so the server is the only clock. Discovery was removed from the scheduled job so the dispatch stays within its request budget, and a failed dispatch is re‐ tried at the next sweep instead of being stamped done. All three superseded copies were replaced by a per‐ manent redirect to production, with their schedules and secrets removed and their data exported before re‐ placement. Effect witness. Merchant mail now uses the working lane. The operator health endpoint reports the mail lane's state, the retry queue depth, the last dispatch result, unmatched payments and the state of every social rail. An operator lever sends a real test message on demand. The press module now composes only for rails whose credentials are bound. A rail without an account reports "on (waiting for account)" and names the secrets that would switch it on, so the press no longer generates queue writes for posts that cannot be sent. Authority. PIN reset now requires a one-hour link emailed to a verified address. The fallback for merchants with no email holds the store for 24 hours and then rotates every credential. Card payments match only a verified email, and an unmatched payment generates a claim link to the buyer instead of being filed silently. The fallback signing key was removed, so a missing secret now causes an error rather than producing tokens anyone could forge. Surface. The spreadsheet CRM was removed at the owner's instruction. As the next subsection describes, its removal created a new gap, which was closed by a leads desk inside the product.
Four defects arose from the rebuild itself. They are reported in full because they bear on the propositions in Section 5.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
R9: removing a surface removed the only surface. The spreadsheet had been the only place where sellers, or the owner, could see and act on their allocated leads with a usable interface. The server exposed the lead actions as an API, but nothing called it except the spreadsheet. Deleting the spreadsheet, correctly, left the sales process with no working surface. The gap was found by asking where a seller would work tomorrow morning, not by any test. It was closed with a leads desk: a page on which a seller or the owner signs in and sees allocated leads by stage. Each lead has a pre-written WhatsApp opening, call and email buttons, and the same claim rules as the API. R10: the corrected hunter's first catch was chain stores. The rebuilt hunter passed its offline self-tests. Its first live dry run on a CI runner returned ten candidates, and the only three with email addresses were two branches of a national supermarket chain and a fast-food chain, each with a corporate address. These are not Tapstall's market. A chain filter was added, based on the map database's brand tags and a short list of distinct‐ ive chain names. A second dry run returned ten independent businesses. Generic words were deliberately left out of the name list after a test showed that "Total" would have flagged "Total Fashion Boutique". R11: the corrected query was too heavy for the commons. The first rebuilt query unioned eighteen tag combinations per city. Every public mirror of the map query service timed out on it, and the main instance answered that it was at capacity. A three-part query, with a fallback across mirrors, returned results in about sixteen seconds from the one mirror that was responsive. R12: removing duplication left a dangling reference. The deploy script had two loops that both waited for the new version to answer. Removing the redundant one left a later line that referred to a variable only the re‐ moved loop had set. Under the script's strict mode the deploy would have failed after uploading. The defect was caught by reading the edited script before shipping. In each case the defect was produced by the rebuild acting on the world in a new way, and each was found by looking at real output or real use.
After deployment the service reported version 2.0.0, and its runtime self-test passed. A test message sent through the new mail lane was delivered and confirmed in the receiving mailbox by a read-only check. The health endpoint showed the lane active with one message used of its hourly two. The press reported every company rail as waiting for an account, and a dry press window skipped the identity without writing any‐ thing. All three superseded copies answered with permanent redirects to production. After the purge the lead table held 16 queued leads, three contacted and 32 do-not-contact. The final test suite passed 421 tests, in‐ cluding new tests for the mail lane, the outreach tick and unsubscribe link, dispatch retry, the junk door and the armed press. The two runner scripts' self-tests run in continuous integration. The cash cost of the rebuild was zero. Its free-tier footprint stayed within the ceilings described in Section 3.4. The largest single draw on the daily key-value write allowance was the one-off clean-up: 89 deletions and 32 do-not-contact markings.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
Unwitnessed autonomy is the condition of an automated system that takes consequential actions on parties outside its boundary when no party is placed to observe those actions' effects and feed them back to the oper‐ ator. Each element of the definition does work. Automated means the actions are initiated by a schedule, an event or a rule, not by a person deciding each one. The condition does not require sophisticated automation. A scheduled job and an SMTP library are enough. Consequential actions on parties outside its boundary restricts the concept to actions with external ef‐ fects: a message in a stranger's inbox, a post on an account, a payment matched or not, a store suspended. A job that recomputes an internal index is not in scope, however badly it fails. No party is placed to observe is the core. The effects may be observable in principle. The recipients of the junk mail certainly observed it. What is missing is a channel from the observer to the operator. A stranger who receives an irrelevant message from a support address has no reason, and often no way, to tell the venture. Feed them back to the operator ties the concept to control, in Leveson's sense [5]. A system is unwit‐ nessed when its controller has no feedback about the process it controls, even if it has abundant feedback about itself. The concept is distinct from three neighbours. Ordinary software defects are failures of computation that a test with the right input would reveal. Unwit‐ nessed autonomy can involve perfectly correct computation: the outreach runner's SMTP code worked, and that was the problem. Technical debt is a cost the owner defers and later pays [12]. Unwitnessed autonomy imposes costs on others now, and the owner may never pay them unless a provider suspends the account or a regulator acts. Spam is unsolicited bulk messaging, usually knowing. Unwitnessed autonomy can produce spam without in‐ tent, as in this case, but it also covers social posting under a borrowed identity, silent failures of promised mail, payment mismatches and duplicate scheduled jobs. None of these is spam.
Five of the nine defect classes are witness gaps. They can be grouped by the four things a controller would need to witness. The recipient. Who received the action, and should they have? Externality defects act on the wrong parties. Boundary-interpretation defects misread what those parties send back. The effect. Did the action do what it was meant to? Silent failures report actions that had no effect (mail never sent, a dispatch never made, posts queued for accounts that do not exist). • • • • 1. 2.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
The identity. Whose name, credential or account did the action use? Borrowed-identity defects act under another party's identity: another venture's mailbox, another brand's social account. The clock. How many times, and from how many places, is this action being taken? Multiplicity defects run the same action on two clocks, or run superseded copies against live or private data. The four other classes (authority leakage, surface sprawl, dependency capacity and build defects) are not wit‐ ness gaps. They are ordinary engineering concerns with established remedies [22], [23]. They are included because a real audit finds both kinds, and because they interact. A surface removed for good reasons left no surface, and an authority hole was reachable through a superseded copy.
The witness map is an audit instrument derived from the four gaps. The operator lists every action the system takes on a party outside itself (send, post, charge, suspend, notify, file, delete) and answers five questions for each one.
Question What a good answer looks like Gap it closes Identity. Under whose credential and name does this act? A credential created for this system alone, named after it, held in one place Borrowed identity Clock. What fires it, and how many things fire it? Exactly one scheduler; every other copy redirected or removed; a failed fire retried, not stamped Multiplicity Effect. What evidence comes back that it worked? A record written after the external system confirms the effect (delivery accepted, post URL returned, message seen in a mailbox) and surfaced to the operator Silent failure Recipient. What evidence comes back about who received it and what they think? A recipient check before acting (does this address belong to this business?) and a parsed channel back (unsubscribe, bounce, reply in the recipient's own words) Externality, boundary interpretation Stop. What stops it, and who can use that? A per-action kill switch the operator can reach in one step, and an exit the recipient can use in one click All
Applied retrospectively, the map surfaces every witness gap in Table 1: The identity question for "send outreach" answers "the owner's personal mailbox, which is another ven‐ ture's support inbox" (R1). The same question for "post to social" answers "another venture's account" (H2, R7). The clock question for "run outreach" answers "the server's dispatch and the workflow's own schedule" (R3). The same question for the service itself answers "three other deployments" (C3, R4). The effect question for "send merchant mail" answers "nothing comes back", which is the silent no-op (H1). The same question for "dispatch workflows" answers "the dispatch call's own return, stamped regardless" (H3). The recipient question for "send outreach" answers "nothing checks that the address belongs to the busi‐ ness, bounces are ignored and the footer is read as a reply" (C2, H4, R2). 3. 4. • • • • • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
The map costs nothing to apply. For a pipeline of this size it takes about an hour. It needs access to the de‐ ployment, not just the code, because two of the four gaps are invisible in code.
Four rules follow from the case. Each closes one gap by construction, not by vigilance. One clock. Every scheduled action is fired by exactly one scheduler. Other mechanisms that could fire the same action (a CI schedule, an old deployment's trigger, a second account's copy) are removed, not merely left unused. The scheduler records the outcome of each fire, not the attempt, and retries a failed fire at the next opportunity. A duplicate clock is worse than a missing one because it masks the failure of the primary. In the case, the workflows' own schedules hid for days the fact that the server's dispatch had been failing every night. One voice. A recipient hears from one sender identity for the life of a conversation. A follow-up is sent only from the mailbox that sent the first touch. Every message carries a working exit that the sender honours, and the sender reads what comes back in the recipient's own words. One voice is partly courtesy and partly con‐ trol. A conversation scattered across identities cannot be witnessed by anyone. One identity. The system acts only under credentials created for it. Credentials borrowed from another ven‐ ture, a personal account or a demo setup are removed even when they "work". A rail with no credential of its own is shown as waiting, not active. The label matters because a dashboard that says "active" for a borrowed account is decoupled from the activity it describes [25]. One surface. There is one place where actions on prospects and customers are taken. Every action goes through one door, where the gate is enforced. A second surface (a spreadsheet, a chat bot, a script) that can propose or perform actions is either removed or routed through the same door. When a surface is removed, the replacement must exist before the removal ships. This rule is the lesson of R9. The four rules are not new individually: they restate least privilege [22], idempotent single-writer design [11] and list-management norms [31], [32]. The contribution is to show that, in a small autonomous pipeline, these four together close the witness gaps that a passing test suite leaves open.
The four defects introduced by the rebuild (R9–R12) suggest one more element: witnessing must recur. A fix is itself a new action on the world. The rebuilt hunter queried a shared public service in a new way (R11) and filed a new kind of prospect (R10). The removal of a surface changed where work happened (R9). In each case the defect was found by observing real output: a dry run on the production runner, the first page of res‐ ults, the question "where will the seller work tomorrow?". The practical form of this element is modest: run every changed autonomous action once in dry-run mode in the production environment and read the output; send one real message and find it in a real mailbox; re-apply the witness map after every change that adds or alters an external action. • • •
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
The case supports seven propositions, stated for testing in other ventures. P1 (prevalence). In autonomous growth pipelines operated by one person, witness gaps make up the major‐ ity of consequential defects. In this case they were 16 of 24 (67%). P2 (test blindness). A pipeline's own test suite will detect few or no witness gaps, however extensive the suite. Externalities, borrowed identities and multiplicity are properties of inputs and deployments that tests do not model, and silent failures pass where tests supply what production lacks. P3 (deployment locus). Borrowed-identity and multiplicity defects will be found by inventorying deploy‐ ments and credentials, not by reading code. The more products an operator runs from shared accounts, the more such defects there will be. P4 (masking). A duplicate clock prolongs the life of a failure in the primary clock, because the duplicate pro‐ duces the outcome the operator watches for. Pipelines with redundant schedulers will show longer undetected failure durations than pipelines with a single scheduler that retries. P5 (composition). Externalities in autonomous outreach arise more often from the composition of individu‐ ally reasonable rules than from any decision about volume or targeting. Examples are "take the first email on the page", "count the source URL as a signal" and "send to anything with two signals". Gates placed at the point of action, not only at the point of ingestion, reduce them more than tighter rules at any single stage. P6 (repair induces). Repairs to an autonomous pipeline introduce new witness gaps at a rate that does not fall to zero. Defects introduced by repair will be caught mainly by live dry runs and real-output inspection, and rarely by tests written at the same time as the repair. P7 (label honesty). Operator dashboards that report configuration ("rail registered", "run successful") rather than effect ("post accepted", "outreach ran and sent N") will be associated with longer-lived witness gaps. Replacing configuration labels with effect labels, including an explicit waiting state, shortens them.
The case brings several established ideas to bear on a setting they were not developed for, and in doing so shifts each of them slightly. Automation research. The automation literature locates the problem in the operator's declining awareness [1], [2]. The case suggests that in autonomous micro-ventures the more basic problem is that awareness was never designed in. No display showed outbound mail, no surface showed which identity the press used, and no record distinguished "dispatched" from "ran". The irony Bainbridge identified still applies, but it is pre‐ ceded by an omission. Parasuraman, Sheridan and Wickens's four-stage model [3] is useful here in a slightly unexpected way. The pipeline was fully automated at the action stage and wholly unautomated, and also un‐ staffed, at the information-acquisition stage that would have reported the effects of action. A pipeline's level of automation should be assessed separately for acting and for witnessing.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
Safety science. Normalisation of deviance [4] usually describes people accepting anomalies they can see. In the case the dispatch failure was recorded in a field nobody read, and the run that should have failed was stamped done. Deviance can be normalised by a data model as well as by a culture. Leveson's insistence on feedback about process state rather than controller action [5] turns out to be the single most useful lens. Every silent failure in the case was an instance of a controller recording its own action as if it were the pro‐ cess outcome. Externality economics. The email commons literature [8], [15] treats the sender as a strategic actor balan‐ cing volume against cost. The case shows externalities produced by composition, with no actor choosing them. That matters for policy. Rules aimed at senders' choices (volume limits, consent requirements [36], [37]) remain necessary, but the case suggests they are reached only through a gate at the point of action, be‐ cause the choice they regulate was never consciously made. Security. The confused deputy [9] is normally a target of attack. The case shows a quieter version: the deputy confused by its own operator's convenience, given a credential because it was available. Least priv‐ ilege [22] is usually argued from the risk of compromise. Here the argument is simpler: a component acting under another venture's identity harms that venture whether or not anyone attacks it. Software engineering. Sculley and colleagues' undeclared consumers [12] reappear as a mailbox shared between two ventures, one of whose consumers was a support function nobody had declared to the other. The concept travels well beyond machine learning. Any shared resource with more than one owner has un‐ declared consumers.
For operators of small autonomous pipelines the implications are direct. Run the witness map before adding the next automation. An hour with the deployment inventory and the five questions found what weeks of green runs had not. Inventory accounts, not just code. Two of the four witness gaps live only in account settings and secret stores. For a person running several ventures, the inventory should cover every account any of them uses, be‐ cause coupling happens where accounts are reused. Prefer a waiting state to a borrowed account. A social rail that says "on (waiting for account)" and names the two secrets that would switch it on is more useful than one that posts into somebody else's audience. The same applies to mail. A dedicated, low-cap mailbox that belongs to the venture is better than a high-capacity personal account that belongs to everyone the owner works with. Stamp outcomes, not attempts. Every scheduled action should write what happened in the world, read back from the world where possible, and the operator's health view should show it. In the rebuild this took one function and a handful of fields. Put the gate at the door. A gate at ingestion protects future rows but not the rows already queued. A gate at the point of action protects both. Read one real output after every change. One dry run on the production runner found chain stores in the rebuilt hunter. One real message, found in a real mailbox, proved the mail lane. Neither had been found by the tests written alongside the changes.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
Budget for witnessing. On free tiers, witnessing costs quota: a health check, a mailbox read, a dry run. In the case these costs were small compared with the clean-up they prevented. The single largest draw on the daily write allowance was deleting 89 rows of junk the unwitnessed pipeline had created. For providers of shared infrastructure, the case suggests a modest design implication. Mailbox and socialplatform credentials issued per application, rather than per account, would make borrowed identities harder to create by accident. Scheduler dashboards that show every trigger firing a given job, across accounts, would make multiplicity visible. For policy-makers and platforms concerned with unsolicited messaging, the case is a reminder that much un‐ wanted mail from small senders is not the product of a spam strategy. It is the product of a pipeline nobody reads. Requirements such as one-click unsubscribe [32], [35] and postal addresses [37] do useful work even for such senders, because they create the channel back that the sender failed to build.
The study has the limitations of its design. It is a single case of a pre-revenue venture with a small volume of activity, so the counts are small and the percentages fragile. The analyst designed the rebuild and has an in‐ terest in its success. The audit, the coding and the paper came from one analytic process, which raises the risk of reading the framework into the data. Counting only defects tied to specific artefacts limits this risk without removing it. The classification of defects into classes involved judgement, and some defects plausibly fit two classes: H2 is both a borrowed identity and an externality, and H5 is both an authority defect and a silent fail‐ ure. Primary classes were assigned, and the witness-gap count would be higher under multi-label coding. The study observes the system on the day of audit and rebuild, not over time. It cannot show that the rebuild prevents recurrence, only that the specific defects were closed and verified. It does not measure harm to re‐ cipients: whether any complaint was filed, or whether the sending accounts' reputations suffered. It does not establish how the first version came to contain these defects. In particular, it does not claim that any mode of building, AI-assisted or otherwise, produced them. The relation between building practice and witness gaps is left as a hypothesis. The market context also limits generalisation. Tapstall's prospects are informal merchants whose presence in open data is thin and uneven [39], [40]. A pipeline hunting in a well-documented market would face fewer junk candidates, though not fewer borrowed identities or duplicate clocks.
Three lines of work would test and extend the propositions. Comparative cases. Applying the witness map to a sample of autonomous micro-ventures would test P1 to P4. Useful comparisons include ventures with single and multiple products, shared and separate accounts, and single and redundant schedulers. The instrument is cheap enough to apply at scale. Longitudinal study. Following a pipeline over months after a rebuild would test P6 and P7: whether repairs keep introducing witness gaps, and whether effect-based dashboards shorten their life.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
Recipient-side measurement. Studying recipients would measure the externality directly: how often small automated senders reach non-prospects, what recipients do, and whether list-management headers change the outcome. That would connect this work to the empirical spam literature [8], [15] and to the policy debate [35], [36], [37]. A fourth line is more speculative. As AI agents take on more of the building and operating of such pipelines, the witness map could be applied by the agents themselves, as a standing check run before and after each change to an external action. Whether agents can witness their own pipelines reliably, or whether witnessing is the part of operation that must stay with a person, is an open and important question.
A growth pipeline that acts on strangers every day, with nobody reading what it does, can pass every test it has while mailing the wrong people from the wrong account, posting into another brand's audience, failing silently, and running in more places than its owner knows. This paper has called that condition unwitnessed autonomy, and has shown in one candid case that it accounted for two-thirds of the defects found and most of the serious ones, none of which the test suite saw. The remedies are not expensive. They are a habit of asking five questions of every autonomous action (iden‐ tity, clock, effect, recipient, stop) and four rules that close the gaps by construction: one clock, one voice, one identity, one surface. The case also carries a warning about repair. Fixing an autonomous pipeline is itself a new set of actions on the world, and those actions need witnessing too. The operator of an autonomous sys‐ tem does not escape the duty to watch it. The work changes from watching each action to building the chan‐ nels through which actions report back.
Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
[1] Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/ 10.1016/0005-1098(83)90046-8 [2] Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230– 253. https://doi.org/10.1518/001872097778543886 [3] Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics – Part A: Systems and Humans, 30(3), 286–297. https://doi.org/10.1109/3468.844354 [4] Vaughan, D. (1996). The Challenger Launch Decision: Risky Technology, Culture, and Deviance at NASA. University of Chicago Press. [5] Leveson, N. G. (2011). Engineering a Safer World: Systems Thinking Applied to Safety. MIT Press. [6] Coase, R. H. (1960). The problem of social cost. Journal of Law and Economics, 3, 1–44. [7] Hardin, G. (1968). The tragedy of the commons. Science, 162(3859), 1243–1248. [8] Kanich, C., Kreibich, C., Levchenko, K., Enright, B., Voelker, G. M., Paxson, V., & Savage, S. (2008). Spamalytics: An empirical analysis of spam marketing conversion. In Proceedings of the 15th ACM Conference on Computer and Communications Security (pp. 3–14). https://doi.org/10.1145/1455770.1455774 [9] Hardy, N. (1988). The confused deputy (or why capabilities might have been invented). ACM SIGOPS Operating Systems Review, 22(4), 36–38. https://doi.org/10.1145/54289.871709 [10] Lamport, L. (1978). Time, clocks, and the ordering of events in a distributed system. Communications of the ACM, 21(7), 558–565. https://doi.org/10.1145/359545.359563 [11] Kleppmann, M. (2017). Designing Data-Intensive Applications. O'Reilly Media. [12] Sculley, D., Holt, G., Golovin, D., Davydov, E., Phillips, T., Ebner, D., Chaudhary, V., Young, M., Crespo, J.-F., & Dennison, D. (2015). Hidden technical debt in machine learning systems. In Advances in Neural Information Processing Systems 28 (pp. 2503–2511). [13] Hevner, A. R., March, S. T., Park, J., & Ram, S. (2004). Design science in information systems research. MIS Quarterly, 28(1), 75–105. [14] Pigou, A. C. (1920). The Economics of Welfare. Macmillan. [15] Levchenko, K., Pitsillidis, A., Chachra, N., Enright, B., Félegyházi, M., Grier, C., Halvorson, T., Kanich, C., Kreibich, C., Liu, H., McCoy, D., Weaver, N., Paxson, V., Voelker, G. M., & Savage, S. (2011). Click trajectories: End-to-end analysis of the spam value chain. In 2011 IEEE Symposium on Security and Privacy (pp. 431–446). https://doi.org/ 10.1109/sp.2011.24 [16] Ostrom, E. (1990). Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press. [17] Perrow, C. (1984). Normal Accidents: Living with High-Risk Technologies. Basic Books. [18] Reason, J. (2000). Human error: Models and management. BMJ, 320(7237), 768–770. https://doi.org/10.1136/ bmj.320.7237.768 [19] Weick, K. E., & Sutcliffe, K. M. (2007). Managing the Unexpected: Resilient Performance in an Age of Uncertainty (2nd ed.). Jossey-Bass. [20] Dekker, S. (2011). Drift into Failure: From Hunting Broken Components to Understanding Complex Systems. Ashgate. [21] Cook, R. I. (1998). How Complex Systems Fail. Cognitive Technologies Laboratory, University of Chicago. [22] Saltzer, J. H., & Schroeder, M. D. (1975). The protection of information in computer systems. Proceedings of the IEEE, 63(9), 1278–1308. https://doi.org/10.1109/proc.1975.9939 Courage Garikai Coldcode Si Labs Library Working Paper 2026-03
[48] Conway, M. E. (1968). How do committees invent? Datamation, 14(4), 28–31. Courage Garikai Coldcode Si Labs Library Working Paper 2026-03