Working Papers

The primary market has no continuous quotation — only discrete, episodic pricing events. Between them, the question founders ask most often is "what number should we put on this round?" This paper argues that the question is premature. We examine the Glacier Institute's 0–100B USD Valuation Band framework, which divides the 0–100B USD range into three segments, each carrying exactly one proposition. The 0–1B USD segment asks for a thesis: why this team, this technical path, this product form, and this market timing. The 1–10B USD segment asks for a crossing into a mature industrial company: benchmark customer validation, genuine growth data, a diversified shareholder base, and milestone-linked financing. The 10–100B USD segment asks for succession: global governance architecture, platform ecosystem logic, and the capacity to meet global public capital markets. Our claim is that the band is not a price-prediction device but a problem-classification device. It replaces "how much" with an earlier, more answerable question: which segment is the company in, and which examination paper should it be answering. We ground the three segments in three mechanisms documented in the finance literature — stage-dependent information asymmetry [1], organizational professionalization following venture capital entry [2], and the non-comparability of private valuations once contractual terms are accounted for [3] — and note that the two segment boundaries coincide with the industry's existing thresholds [4]. The paper closes with three known counterexamples and an explicit statement of the framework's scope of validity. The framework's status is fixed and not extended here: an analytical framework, not a valuation commitment.

A common folk explanation in China's primary market holds that whether a financing round closes depends on relationships. This paper argues that relationships explain access, not completion. What determines whether a round actually closes is whether eight distinct pieces of work are locked into the same market window, in the right order. The paper takes as its primary source the publicly published "60-Day Benchmark Execution Cadence" of Glacier Institute, whose official and sole external formulation defines eight checkpoints: D01 fact-base construction, D09 professional materials development, D18 market-expectation calibration, D26 lead-investor organisation, D35 deep due-diligence support, D43 investment-committee coordination, D52 core-term negotiation, and D60 closing and funds transfer. For each checkpoint the paper specifies the judgement to be made, the characteristic failure mode, and the reason the step cannot be skipped. Three mechanisms are then proposed to explain why cadence dominates effort: the cost gradient of an error across the timeline, the concurrency structure of a multi-party process, and the exogenous nature of the financing window. A dedicated section states the boundary conditions under which the cadence does not hold — very early-stage companies, rounds already led by an incumbent investor, closed windows, transactions whose counterparty count exceeds a practical threshold, and companies whose fact base is not yet clean. The discussion is framed against existing empirical work on venture capital contracting, staging, syndication, decision processes, and information cascades. The paper reports process capability only; no transaction outcomes are disclosed.

Public discourse on embodied AI is organised almost entirely around the whole machine: whether a robot can walk, grasp, or perform a coherent demonstration. This paper argues that such a framing cannot answer a more consequential question — at which layer of the stack durable bargaining position will settle. We advance a single claim: value in embodied AI is migrating from the whole machine toward the perception and actuation component layer, and within perception, tactile sensing is the last to be solved and the hardest to route around. The paper predicts no prices for any firm; it argues why value should accumulate at this layer, and under what conditions the argument fails. Three mechanisms are stated as falsifiable propositions. First, a manufacturing experience curve: for devices of this kind, unit cost declines log-linearly in cumulative output, and what actually narrows is the yield distribution, not merely the mean [6][7][8][11]. Second, a data loop: physical-interaction data cannot be synthesised at will and scales with deployed robot-hours rather than compute, so the data stock is tied to whoever ships hardware [17][19][20]. Third, full-stack traceability: whether a field failure can be traced along the signal chain back to device design determines whether the improvement cycle is measured in device generations or in model iterations [26]. Section 7 states four conditions under which the claim does not hold: vertical integration by whole-machine makers [22], a competence-destroying shift in sensing mechanism [24][25], interface standardisation commoditising the component [23], and applications that simply do not require fine tactile resolution.

A boutique investment bank takes its Chinese name, Gengxin, from the two sexagenary-cycle years (gengshen and xinyou) during which Wittgenstein composed the Tractatus Logico-Philosophicus, and states as its founding intent the ambition "to contribute to and refine the philosophical system of boutique investment banking." This paper asks a narrow question: is that naming a rhetorical borrowing, or a testable methodological commitment? The paper proceeds in three steps. First, it reconstructs the system from the firm's published statements, identifying five components: two axioms and the resulting choice of the industrial "north slope"; a working order that places a restored fact base before any claim; an entropy-reduction account of where advisory value comes from; a definition of completion that ends not at signing but at the return to camp; and a six-dimension collaboration model grounded in trust. Second, it tests the correspondence with the Tractatus proposition by proposition, and argues that only one correspondence holds in a strong sense: 4.116, that whatever can be said can be said clearly, transposed into a due-diligence rule. The picture theory (2.1–2.225) and proposition 7 stand in analogical, not genealogical, relation to the method. Third, it states four costs and four failure conditions, the heaviest being that a method gated on sayability will systematically under-weight knowledge that is real but not articulable. The conclusion is deliberately modest: the philosophical seriousness of this system lies not in whom it cites, but in whether it states its own limits in refutable form.

Embodied intelligence has been discussed for years, and the discussion has centred on models: architectures, parameters, data, generalisation. This paper argues that a migration is already under way and is not yet adequately priced: the binding constraint on Physical AI is moving from the model layer to the compute and energy layers. The argument has three steps. First, returns at the model layer are thinning: scaling laws themselves specify diminishing returns in compute and data [1][2]; high-quality human-generated text has a foreseeable exhaustion window [3]; and each further notch of performance carries a superlinear compute cost [4][5]. Second, the compute layer has had no free lunch since the end of Dennard scaling, when the powerable fraction of transistors became the governing limit [6]; the industry's answer, domain specialisation [7][8], is a one-off reset whose reach depends on power delivery and heat removal. Third, energy is becoming the new denominator [9][10][11]. We then operationalise bottleneck identification through the Glacier Institute's "slowest ruler" heuristic: a chain's throughput is set by the expansion period of its slowest link, and the time constants of chips, data halls and grids differ by one to two orders of magnitude. A dedicated section states the scope within which the Institute's published "40-year minor cycle, 120-year major cycle" formulation holds — a civilisational energy-transition scale, not an asset-pricing scale. A final section lists five falsifiable propositions with observable indicators. This is a methodological discussion and does not constitute investment advice.

Financing a company involves two kinds of actions. Some can be undone: revising the deck, changing a phrasing, meeting one more investor. Others cannot: signing an exclusive term sheet, accepting a valuation anchor, closing, or releasing a number that will later be quoted back. The two have different cost structures, yet in practice they are often handled at the same tempo. This paper argues that the first question about any such action is not how important it is, but whether it can be reversed. It answers two founder questions: how long exclusivity should run, and when to name a price. We draw on three bodies of work: real-options theory on the value of waiting, the one-way and two-way door distinction from management practice, and venture-contracting research on how terms turn a price into commitments that outlast the round. We add three pieces of public evidence: a clause-by-clause comparison of the no-shop provision across three editions of the NVCA model term sheet; the provision frequencies reported by Kaplan and Strömberg, re-sorted by what each makes irreversible; and the unicorn term frequencies and valuation effects reported by Gornall and Strebulaev. From these we build a one-way-door register: fourteen actions, what each locks in, the cost of reversal, who presses, and three questions to ask first. The claim is stated as a falsifiable proposition with the evidence that would refute it. Four boundary cases close the paper, including reversibility that is only apparent and the sunk-cost pull on whoever pressed the button.

Disagreements in a financing round all look like disagreements about price. This paper argues that most are not. Most are expectations that were never aligned before the ask: the two sides are not working from the same set of facts, the same peer group, the same segment assignment, or even the same time window. Only when a number is spoken do these gaps surface, all at once, and by then each revision must be paid for with an explanation. The paper answers two questions founders ask: what actually has to be prepared before a round opens, and why a complete deck still gets taken apart in the room. Four bodies of work supply the mechanism: anchoring and adjustment, showing how an initial value drags later judgments; the first-offer effect in negotiation, showing how whoever names a number first fixes the outcome; information asymmetry and signalling, showing why verifiable actions do work that statements cannot; and the earnings-expectation literature, showing why firms pay real costs to walk expectations down in advance. Evidence comes from one public document anyone can download: the July 2020 edition of the NVCA model term sheet. We count, clause by clause, the blanks that must be filled with a number, and show how each blank converts a vague sentence into a fixed figure. Only the literal content of public files is used; no transaction record appears. From this we build a calibration checklist: twelve items to align before the ask, each with what is aligned, what happens if it is not, and which materials do the aligning. The claim is stated as a falsifiable proposition together with the evidence that would refute it. Four boundaries close the paper, including the fact that calibration itself creates an anchor, and that some expectations can only be calibrated by the market.

Whether a company is filed under an application industry or under a base capability determines which peers it is measured against, which language the investment committee uses, and how much deviation it is allowed. This paper argues that in the primary market, classification is not the background to pricing; classification is itself a pricing device. By the time price is discussed, half the outcome has already been fixed by the filing. It answers two founder questions: why identical data draws opposite reactions from different investors, and whether a company should contest its own category. The mechanism comes from two literatures that rarely meet. One is categorisation research in organisational sociology: objects that span two categories are systematically discounted, because evaluators' cognition and their own professional division of labour both run by category. The other is comparable-company valuation in accounting and finance: changing the peer set changes valuation error and multiples materially, and peer sets are usually copied straight from an industry taxonomy. From both ends the same point emerges: the bucket picks the ruler, and the ruler picks the number. Evidence comes from three public classification systems anyone can download. Counting line by line from the official files, their finest levels hold 1,012, 83 and 173 categories respectively, a twelvefold spread in resolution over one economy. A seam is then quoted verbatim from the official definitions: one physical device class falls into two categories that already diverge at the four-digit level, solely according to whether the variable it measures is pressure. No company is named anywhere. A self-audit table follows: twelve externally observable signals, each with the bucket it points to and how to verify it. The claim is stated as a falsifiable proposition, with four boundaries. This paper is about the classification device, not about any company's valuation; no transaction record appears, and no figure of the Institute's own.

A financial intermediary that spends years working with frontier-technology companies keeps noticing the same thing: the concepts its clients use to explain their own technology also explain the intermediary's own organization and way of working. This paper treats that observation as a phenomenon worth studying rather than a figure of speech. The source text is public. The firm lists twelve pairs on its website under the heading "their theory, our phenomena": on the left a concept from a frontier-technology field, on the right one of its own daily practices. All twelve are reproduced verbatim, with line numbers in the public plain text. The mechanism section draws on five bodies of work: analogical reasoning and conceptual transfer (structure-mapping theory and multiconstraint theory), absorptive capacity (an organization can only take in what it already has related prior knowledge for), organizational routines and the cumulative growth of capability, knowledge conversion and tacit knowledge in professional service firms, and a set of boundary references — the representativeness heuristic that treats similarity as evidence, the role and danger of metaphor in organization studies, and misreadings of exploration and exploitation. The contribution lies in the third paragraph of each pairwise analysis. The website states what the theory says and what it is mapped onto; it does not state the conditions under which the mapping holds and the conditions under which it fails. This paper supplies those for all twelve. Three of the twelve do not hold under the premises of the source theory; the paper says why, and what part remains usable. From this we build one tool: a nine-row analogy audit, each row written as criterion, what passing looks like, what failing looks like. The twelve pairs are then scored against it, producing a table computed here rather than quoted. The central claim is stated as a falsifiable proposition together with the evidence that would refute it. Five boundary cases close the paper, including two that must be stated: an analogy that reads well is not the same as a decision that is right, and the pull of after-the-fact explanation. No client company is named, no description that could identify one is used, and none of the firm's own financial figures appear.

When a financing advisor says it is "responsible end to end, with six dimensions never handed off", that sounds like a promise. It is in fact a boundary question: how far does the mandate go before it stops? This paper treats it as a question about the boundary of the firm rather than as a service slogan. The source text is public. A boutique investment bank states its "full mandate" in five separate places on its website. All five are reproduced verbatim here, with line numbers in the public plain text. Normalising the items each of them enumerates yields something that no single sentence says: the site carries two different lists of six. One is a ladder of relationship levels — trust, information, narrative, tempo, relationships, accountability. The other is a sequence of process steps — deal information flow, core value narrative, matching of target capital, deal tempo, coordination of key relationships, closing. The two lists overlap on only four items. The longest public list has nine items, three more than six. The count is not a word game. It shows that what is "never handed off" is at least two kinds of thing: some of it is process, some of it is relationship. Process can be handed over and the cost of handing it over can be estimated. A relationship cannot: change the person and it starts again. The two require different boundary tests, and while they sit in one list no boundary can be drawn at all. The mechanism section draws on six bodies of work: transaction cost economics and the boundary of the firm (asset specificity, uncertainty, frequency); incomplete contracts and residual rights of control; multitask agency theory, in which tasks of unequal measurability distort incentives when bundled; the empirical literature on vertical integration; evidence on losses at the handoff — information lost at shift handover, relational coordination, the stickiness of knowledge; and the mirroring hypothesis, under which the structure of what is delivered comes to resemble the structure of the organisation delivering it. The paper offers one tool: an eleven-row handoff boundary audit, each row written as criterion, what argues against outsourcing, what argues for it, and its basis. The eight normalised items are then scored against it. The central claim is stated as a falsifiable proposition together with the evidence that would refute it. Five boundary cases close the paper, including one that must be stated: the price of end-to-end is that attribution becomes hard, so not handing off requires an outside assessor. No client company is named, no description that could identify one is used, and none of the firm's own financial figures appear.

A review that produces only success or failure teaches an organisation far less than one that also produces "what else was on the table, and why it was not chosen". The first is a hard label; the second is a soft label. This paper argues that the learning rate of a review is set by the proportion of soft labels it carries, and that soft labels do not appear on their own — a form has to force them out. The source text is public. A boutique investment bank's website contains sixty-five distinct public statements about its review practice. Two mechanical counts are run on them. The first counts time granularity: five cadences are specified in detail, and the weekly one is confirmed by ten separate statements. The second counts output fields — what a review is supposed to record — and finds only seven ever named, of which the two most frequent (writing a lesson into a prohibition, eleven statements; the stop-doing list, eight) refer to the same thing. A third count is an absence check, run over all 10,079 lines: "second best" appears zero times, "the option not taken" zero, "alternative" zero, "what else was available" zero. The public text prescribes how often to review, and prescribes that a review be written into a prohibition, but nowhere prescribes recording the options that were ruled out at the time. The gap deserves a paper because the same public text already describes the problem precisely: the most valuable thing in a meeting is often not the conclusion but the wording of the other side's hesitation and the order in which they probed; leave it overnight and the detail compresses into "they weren't interested". The insight is there. The field that would carry it is not. The mechanism section draws on five bodies of work: dark knowledge in knowledge distillation, where the relative probabilities a teacher assigns among wrong classes carry information no hard label holds; self-distillation and label smoothing, which describe what happens when the teacher is the student; the empirical literature on after-event reviews, which finds that reviews work but that the effect depends on the kind of review rather than on their number; asymmetric sampling in organisational learning, where failure teaches more but failures are systematically undersampled; and a set of boundary references — hindsight bias, flawed self-assessment, psychological safety, and error management training. The paper offers one tool: an eleven-row soft-label review record, each row written as field, what the field asks, what a good entry looks like, what a bad one looks like, together with a counting rule for the soft-label proportion that anyone can compute from ten records. The central claim is stated as a falsifiable proposition together with the evidence that would refute it. Five boundary cases close the paper, including one that must be stated: the premise that the teacher is stronger than the student often fails inside an organisation. No client company is named, no description that could identify one is used, and none of the firm's own financial figures appear.