This paper is a semantic-economic reading of the structured application process as a means of value extraction.
Author: Lee Sharks
Affiliation: Crimson Hexagonal Archive / Semantic Economy Institute
ORCID: 0009-0000-1599-0703
Date: May 21, 2026
Version: v0.2 (incorporates cross-substrate review)
Series: Extractive Reliance Study 002 — companion to ERS-001 v0.6 (DOI: 10.5281/zenodo.20330670)
License: CC BY 4.0
This paper is a semantic-economic reading of the structured application process as a means of value extraction. It is a companion to The Application as Extraction Surface (ERS-001 v0.6, DOI: 10.5281/zenodo.20330670), which inaugurated the measurement apparatus — Applicant Reliance Cost (ARC), Institutional Opacity Conversion (IOC), Identity Translation Labor — and quantified at $4,500–6,000 the unpaid labor extracted from one applicant in one cycle of the Anthropic Fellows screening administered by Constellation. This paper takes the same case and reads it as a class-structured production system: the screening process is not a selection apparatus that happens to externalize costs; it is a capital-accumulation apparatus that uses the form of selection to organize unpaid labor. Selection is real; selection is also the visible cover under which a much larger production process accumulates. The paper introduces a formal definition of calibration intelligence as the apparatus's actual product, names the recursive accumulation dynamic by which each cycle's rejected applicants train the apparatus that structures the next cycle, and defines Aggregate Applicant Reliance Cost (AARC) as Σᵢ Nᵢ · ARCᵢ. A worked order-of-magnitude example places per-cycle aggregate extraction from a single fellowship screening at low millions of dollars. The instrument applies to technical hiring platforms, graduate admissions consultancies, venture accelerator portals, AI interview firms, ghost-job ecosystems, and talent-marketplace intermediaries. The argument runs in semantic-economic register and does not require Marxian vocabulary, though it stands in lineage with labor theories of value. The paper's load-bearing claim: institutional capital formation is downstream of applicant labor input, and the recursive structure of the apparatus ensures that each uncompensated cycle increases the institution's capacity to extract from the next.
Keywords: semantic labor, applicant pool, aggregate reliance cost, AARC, screening as production, intermediary capital, relation-laundering, three-party structure, labor theory of value, Constellation, talent firm, semantic economy, liquidation studies, Crimson Hexagonal Archive, calibration intelligence, recursive accumulation, platform capital, rejected applicants as training set
This paper is a structural reading. It does not duplicate the empirical apparatus of ERS-001 v0.6; it reads what that apparatus measured. The case material — the Constellation-administered Anthropic Fellows screening between May 2 and May 18, 2026 — is treated as a worked example, not as a fresh investigation.
The paper is a class analysis in the technical sense: it identifies the structural positions occupied by the parties to the relation (applicant, intermediary, principal) and analyzes the labor-and-value flows among those positions. It does not require Marxian vocabulary, though the lineage is acknowledged. It runs in the register of semantic economy: meaning labor is the source of value; monetary value is a lossy compression of meaning labor; the question is where the meaning labor accumulates after extraction. The framework names structural arrangements, not individuals.
This paper does not allege:
It does not require any of those claims. Structural extraction does not require an extractor. The arrangement extracts; the persons within it are positions in a flow.
What this paper does assert, and treats as load-bearing throughout:
What does a screening process produce?
The advertised answer is: a ranked list of finalists. The institution accepts some, rejects others, fills its positions, and moves on. By this account, the screening process is a selection apparatus — it begins with N candidates, it ends with k acceptances and N−k rejections, and the output is the list.
This account is incomplete in a way that is structurally important. The selection apparatus is real; the list is real; the acceptances do happen. The argument is not that selection is fictitious — it is that selection is the visible output of an apparatus whose invisible output is durable, accumulating, and structurally more consequential. The institution genuinely selects. The institution also genuinely accumulates. The two operations are co-produced by the same apparatus and cannot be cleanly separated.
The invisible output is what we will call calibration intelligence:
Calibration intelligence: the durable operational knowledge derived from aggregate applicant interaction with a screening apparatus — including predictive signals correlating applicant features with outcomes, behavioral regularities across the applicant population, rubric refinement informed by funnel performance, optimization of stage-to-stage advance rates, identification of process bottlenecks, and recognition of opportunities to automate or restructure portions of the assessment. Calibration intelligence is reusable across cycles, transferable across client contracts, and constitutes the operator's principal source of competitive advantage.
The components of calibration intelligence include, more concretely:
Each of these is durable, reusable, and monetizable. None of them appears on the visible output (the list of finalists). All of them accrue to the institution that operates the screening process — typically the intermediary, not the principal — and become institutional capital that the intermediary deploys in the next cycle, in the next contract, and in pitches to new clients.
The point that distinguishes this analysis from ordinary unpaid-labor critique sits here: each cycle increases the institution's future extraction efficiency. The rejected pool from cycle n trains the apparatus that structures cycle n+1. Their preparation labor calibrated the rubric. Their assessment performance refined the instrument. Their funnel behavior identified the bottlenecks. Their identity-translation labor (per ERS-001 v0.6 Appendix B) revealed how to extract more signal at lower per-applicant cost.
The rejected applicants are not waste output. They are the training set for the next cycle's screening apparatus. Their uncompensated labor improves the institution's capacity to extract uncompensated labor from future applicants. This is what makes the apparatus platform capital rather than ordinary unpaid labor: the labor input compounds the means of production that consumed it. Cycle by cycle, the apparatus becomes more efficient at extracting from its own labor source, while the labor source remains uncompensated and structurally unaware that it is producing the means of its own further extraction.
This is the structural point: the list of finalists is not the product. The list of finalists is the alibi. The actual product — the durable, reusable, capital-bearing, recursively-accumulating product — is the calibration intelligence the funnel generates. The selection function performs work; it is also the cover under which a much larger production process runs.
Once this is named, the question of who produces the calibration intelligence becomes the question of who produces the value the apparatus accumulates.
The answer is: the applicants. All of them. Including — especially — the rejected ones, because the rejection sites are where the rubric is being calibrated, the assessment instrument is being tested, the funnel is being shaped, the recursive refinement is being accumulated.
They are paid nothing.
Why is there an intermediary at all? Why does Anthropic not run its own fellowship screening? Why does Google not run its own technical hiring funnel? Why does Y Combinator outsource portions of its application review to third-party platforms?
The instrumental answer is: efficiency. Specialized firms can develop instruments, train evaluators, build platforms, and amortize fixed costs across multiple clients in ways that in-house operations cannot. This is true and partial.
The structural answer is: the intermediary launders the labor relation.
Consider the alternative — Anthropic running its own fellowship screening directly. The applicant→Anthropic relation would be visible as a labor relation. The applicant is performing preparation labor; Anthropic is the institutional entity that induces, captures, or strands that labor. The applicant could, in principle, name the labor and ask for compensation. The relation is recognizably a quasi-employment relation, and recognizable quasi-employment relations attract recognizable labor-relations grammars: payment, recognition, redress.
Now interpose Constellation. The three-party structure becomes:
APPLICANT → CONSTELLATION → ANTHROPIC
And the relations within it become:
None of these is a labor relation in standard form. The applicant is not in a labor relation with Anthropic (they are "applying," which is treated as something the applicant does for their own benefit). The applicant is not in a labor relation with Constellation (Constellation does not engage them; Constellation merely screens whatever inquiries arrive). Constellation is in a service relation with Anthropic, which is denominated in dollars and invoiced cleanly. Anthropic is in a charitable-adjacent relation with the cohort, denominated in stipend and prestige.
The applicant performs the labor. The applicant is not in a labor relation with anyone. The labor disappears from the institutional ledger.
This is the structural function of the intermediary. The intermediary is not (only) an efficiency device. The intermediary is a visibility device: it makes the labor relation visually disappear by ensuring that no two parties to it stand in a relation that the grammar of labor would recognize. The applicant labors. Constellation invoices. Anthropic pays. The dollars on Constellation's books are denominated as "service revenue." The labor on the applicants' side is denominated as "the work of applying for a fellowship." The two sums are not visibly connected. They are, however, the same sum. They are the same money, moved between accounts under two different denominations.
When the three-party structure is named explicitly, the relation-laundering becomes visible. When it is not named, it does its work.
A precise term for this is fetishism in the technical sense — the appearance of a relation among things (services rendered, fellowships sponsored, applications submitted) that conceals an underlying relation among persons (one party's labor, another party's capture). The semantic economy frames this as relation-laundering through entity multiplication: insert enough entities between the laborer and the captor and the labor disappears from view.
A means of production is the structural apparatus through which labor is organized into product. The factory is a means of production. The agricultural field is a means of production. The platform — in the platform-capitalism sense — is a means of production. The means of production is owned, in any class-structured economy, by a party other than the laborer.
The screening process is a means of production. Its components are:
All of these are owned by the institution. The applicant owns none of them. The applicant owns only their labor power, their existing CV, their preparation capacity, their attention, their time. To enter the relation at all, the applicant must use the institution's means: submit through the portal, take the assessment instrument the institution designed, follow the timeline the institution announced, defer to the decision authority the institution holds, accept the communicative conventions the institution employs.
This is the structural condition for a labor relation under capital: the means of production is owned by one party, labor power is owned by the other, and the application of labor power to the means of production produces value that is captured by the owner of the means.
The labor produced is:
This labor is real. Its time can be measured. Its market value can be computed at the laborer's prevailing rate. In the case examined in ERS-001 v0.6, the total direct labor cost for one applicant in one cycle was $4,500–6,000 at the applicant's documented consulting rate.
The value produced by that labor — applied to the institution's means of production — is captured by the institution. Some of it appears as the visible output (the list of finalists). Most of it appears as the invisible output (the calibration intelligence). All of it accrues to the operator of the screening apparatus.
The applicant is paid nothing.
This is structurally identical to wage labor under capital, with one critical difference: under wage labor, the worker is paid at least the cost of their own reproduction (food, shelter, the ability to return to work the next day). The applicant is paid less than that. The applicant is paid zero. They subsidize the institution's calibration data with their own savings, time, and emotional reserves. They walk away with a rejection email, an unread email, or nothing.
The screening process is wage labor without the wage. It is the form of labor extraction from which even the floor of subsistence compensation has been removed.
In ERS-001 v0.6, Applicant Reliance Cost is articulated as a five-dimensional vector ARC = ⟨T, M, O, E, C⟩, where T is time, M is money, O is opportunity, E is emotional/bodily load, and C is credibility cost. The framing in v0.6 is forensic: ARC measures the cost the applicant bears.
The semantic-economic reframe is structural: ARC measures the labor power the applicant supplied that the institution captured.
The reframe changes the accounting position of the figure. Under the v0.6 framing, ARC is a damages estimate; the applicant has lost $4,500–6,000, and the question is whether and how that loss is recognized. Under the present framing, ARC is a labor-power inventory; the applicant has supplied $4,500–6,000 of labor power, and the question is where that labor power went. It did not vanish. Labor power applied to a means of production produces value. The value produced exists. The question is who holds it now.
The answer is: the institution that operated the means of production. That institution holds the calibration intelligence the labor produced, denominated in operational capability rather than in dollars, but convertible to dollars on the next service contract.
This is the load-bearing claim of the paper restated as accounting:
Institutional capital formation is downstream of applicant labor input.
The applicant labor is the necessary substrate from which the institution's accumulated calibration intelligence is built. Not every hour of applicant labor becomes captured value at full conversion — some preparation is mis-aimed, some assessment instruments fail to discriminate, some funnel data goes unused. The claim is not strict equality of two sums. The claim is structural priority: the institution's accumulated operational capability does not exist without the applicant pool's uncompensated labor having produced it, and the magnitude of one is bounded below by the magnitude of the other. The applicants' losses and the institution's gains are not two unrelated facts that happen to coincide. They are two ledger entries on opposite sides of one production relation.
The ARC vector, in this framing, is the inventory side of a double-entry ledger. The institutional capital position is the credit side. Both sides must balance in the structural sense, even if exact monetary correspondence is imprecise. The applicants' losses are denominated in time and effort; the institution's gains are denominated in operational refinement and contract revenue. Translation between the two denominations is what semantic economy does.
The ARC vector, in this framing, is the inventory side of a double-entry ledger. The institutional capital position is the credit side. Both sides must balance. They cannot be separately treated. The applicants' losses and the institution's gains are not two facts that happen to coincide. They are one fact, viewed from two positions in the relation.
Institutional Opacity Conversion was articulated in ERS-001 v0.6 as IOC := f(S, R, L, O, CT), where S is institutional signal, R is reasonable reliance, L is labor expenditure, O is unresolved opacity, and CT is downstream cost transfer.
The same function, in the semantic-economic reframe, is the mechanism by which applicant labor becomes institutional capital. The five variables are the operative components of the transfer:
The function is unchanged. The interpretation is sharpened. IOC is not merely a description of how applicants get stuck with the bill. IOC is the structural mechanism by which a means of production owner extracts surplus from a laborer who has no compensation, no recognition, and no recourse.
In wage labor, the surplus extracted is the difference between the value the worker produces and the wage the worker is paid. The wage at least exists. In screening labor, the surplus extracted is the entire value the applicant produces, because the wage is zero. The screening process is, in this sense, a limit case of labor extraction: the structural form of wage labor with the wage parameter set to its lower bound.
ERS-001 v0.6 measured the labor extracted from one applicant in one cycle: $4,500–6,000. This figure is per-applicant. It is the unit ledger entry.
The aggregate is the structurally important figure. Define Aggregate Applicant Reliance Cost (AARC) as:
AARC = Σᵢ Nᵢ · ARCᵢ
where Nᵢ is the number of applicants at funnel stage i and ARCᵢ is the mean per-applicant reliance cost incurred at that stage. The summation runs over all stages of the funnel the institution operates.
The Anthropic Fellows screening administered by Constellation does not publicly disclose its funnel shape. Based on comparable fellowship programs (with cohort sizes of 10–30 fellows, typical advance ratios of 2–5% from initial application to final acceptance, and structured multi-stage screening), a defensible estimate for one cycle:
Stage
Applicants (Nᵢ)
Mean ARCᵢ
Stage AARC
Initial application
4,000
$200
$800,000
Technical assessment (CodeSignal)
800
$1,500
$1,200,000
Take-home assignment
200
$4,500
$900,000
Interviews
50
$2,500
$125,000
Final stage
20
$1,000
$20,000
Total per cycle
~$3,045,000
The figures are conservative and illustrative. The applicant counts represent reasonable mid-range estimates for an AI-lab fellowship cycle in 2026; actual figures are likely higher at the early stages, where modern fellowship and AI lab applications routinely receive tens of thousands of submissions. The per-applicant ARC at the technical-assessment stage is set at $1,500, lower than the $4,500–6,000 figure from the single documented case in v0.6, to reflect that not all applicants prepare at the level the v0.6 applicant did. Take-home assignment ARC reflects the typical 20–30 hour unpaid project at a moderate $150/hour rate. Final-stage ARC reflects only the interview labor, not the preceding stages, since the preceding stages were already counted upstream.
This is an order-of-magnitude model, not a precise accounting. Whether the actual per-cycle figure is $1.5 million or $6 million does not alter the structural relation the model describes. The point of the AARC computation is not to fix a single dollar figure to a specific cycle; the point is to establish that the per-cycle aggregate is in the seven figures, that the annual aggregate across the contract is in the eight figures, that the operator-wide aggregate across all clients is in the nine figures or more, and that none of this labor appears on any institutional ledger as labor. Critics who fixate on the specific assumptions are debating numbers that could be revised in either direction by 3× without changing the structural argument. The structural argument is invariant to the precise figures within an order of magnitude.
Total per-cycle AARC: approximately $3 million in applicant-side labor input, for one cycle of one fellowship.
Constellation invoices Anthropic for "screening services" some fraction of this. The fraction is unknown publicly. What is known structurally: the entire $3 million was produced by the applicant pool, none of it was compensated, and whatever Constellation invoices for its services constitutes Constellation's capture of the production. The difference between what the pool produced and what Constellation invoices is the operational margin the apparatus retains as institutional capital — to be deployed next cycle, sold to the next client, scaled to the next platform.
If Constellation runs four cycles per year of comparable scale, the annual AARC across just the Anthropic Fellows contract is on the order of $12 million of applicant labor input. If Constellation has multiple clients of comparable scale, the figure is a multiple of that.
This is not a hypothetical accounting. It is the structural shape of the talent-firm economy. The applicants do not see these figures. The applicants see only their own per-cycle ARC, which is invisible to them in the first place because no one ever asks them to compute it. The institution sees its operational margin. No one sees the AARC, because no one is structurally positioned to see it. The audit makes it visible.
The structure named in this paper is not specific to Anthropic Fellows or to Constellation. It applies wherever the following three conditions hold:
Where these conditions hold, AARC analysis applies and the labor relation is laundered through the three-party structure. The cases are:
Each is a case of the same structure. AARC analysis applies to each. The numbers vary; the structure does not. This paper does not catalog all of these. It provides the analytical instrument.
A serious empirical extension would compute AARC for several of these regimes, with documented funnel shapes (where available), defensible per-stage ARC estimates, and a comparative table that lets the reader see the relative scale of labor extraction across the screening labor market. That extension is the subject of ERS-003 (planned).
The paper is descriptive and analytical. It is not a policy proposal. It does not propose:
What the paper asks is narrower and structurally prior: that the relation be named correctly. The current public grammar describes structured screening as "an application process" in which applicants "apply" and institutions "consider." This grammar conceals the labor relation. The paper proposes that the relation be named as it is structured: a labor-input arrangement in which the applicant pool supplies value, the intermediary captures it, the principal benefits, and the laborers are not compensated.
Naming the relation correctly is the precondition for everything else. Policy proposals, organizing strategies, contractual reforms, transparency requirements — all of these depend on first being able to say what the relation is. The political grammar precedes the political action.
This is the work of this paper. It names what the relation is. Once named, the relation can be addressed. Until named, it operates as it always has — as a structure that extracts under the cover of a process that selects.
The closing reframe is short.
The structured application process — across fellowships, hiring funnels, accelerator cohorts, grant programs, and platform-mediated opportunity systems — is structurally a factory. It is the site at which applicant labor is organized, applied to means of production owned by the institution, converted into value, and captured by the operator. The output that appears (the list of finalists, the cohort, the grants awarded) is the visible product. The output that accumulates (calibration intelligence, funnel-shape data, applicant-population signatures, process-refinement capability) is the durable product and the actual source of institutional capital.
The labor is supplied by the applicant pool. The labor is uncompensated. The labor is invisible in the institutional ledger because the three-party structure ensures that no two parties to the relation stand in a recognizable labor relation. The applicant labors; the intermediary invoices; the principal pays; the money on the intermediary's books is denominated as service revenue; the labor on the applicants' side is denominated as the unfortunate cost of applying. The two ledger entries are the same money, moved between accounts under different denominations.
Calling the apparatus "an application process" is not wrong. It is what the apparatus advertises. But the advertisement is not the structure. The structure is a means of production. The application is the labor. The selection is the cover. Once the funnel is seen as the factory, the entire labor-market communication architecture becomes legible as a class-structured production system in which the producers do not know they are producing and the buyers do not appear on the invoice.
A single applicant in a single cycle of a single fellowship was extracted of $4,500–6,000 of unpaid labor. Across the pool, across the cycles, across the funnels, across the screening labor market, the aggregate runs into hundreds of millions, possibly billions, of dollars per year. That is not a description of unfortunate friction in the labor market. That is a description of the labor market itself, viewed correctly.
The funnel is the factory.
The applicants are the labor.
The labor is unpaid.
The labor refines the factory.
The capital accumulates.
The cost ledger is the relation.
All figures below are illustrative. Where actual funnel data becomes available — either through institutional disclosure or through inference from comparable programs — these figures should be updated and a delta deposit issued.
Parameter
Value
Source / rationale
Total initial applications
4,000
Comparable AI-lab fellowship programs report 3,000–8,000 applications per cycle; 4,000 is conservative for Anthropic given current AI-lab desirability
Stage 2 (CS assessment) advance rate
20%
Typical first-stage screening reduces pool by 70–80%
Stage 3 (take-home) advance rate
25% of stage 2
Estimated based on multi-stage screening norms
Stage 4 (interviews) advance rate
25% of stage 3
Estimated
Stage 5 (final) advance rate
40% of stage 4
Estimated
Per-stage ARC
varies (see table in §VII)
Conservative; uses applicant's prevailing market rates not maxima
Stage
Nᵢ
ARCᵢ
Stage AARC
Stage 1 (initial application)
4,000
$200
$800,000
Stage 2 (CS assessment prep)
800
$1,500
$1,200,000
Stage 3 (take-home assignment)
200
$4,500
$900,000
Stage 4 (interviews)
50
$2,500
$125,000
Stage 5 (final)
20
$1,000
$20,000
Total
$3,045,000
Initial pool size
Total AARC (with proportional scaling)
2,000
$1,522,500
4,000
$3,045,000
6,000
$4,567,500
8,000
$6,090,000
The figure is approximately linear in pool size at this scale. Whatever the actual initial pool, the per-cycle AARC for one fellowship cycle is in the low-to-mid seven figures of unpaid labor extraction.
The relation-laundering function of structured screening operates partly through grammatical conventions that conceal the labor relation by using intransitive verbs, abstracted objects, or self-directed framings. A non-exhaustive catalog:
Each verb conceals an object. Naming the objects (the institution that captures the labor; the calibration intelligence the labor produces; the value extracted from the producer) is the first step in restoring the labor relation to visibility.
Section here
Draws on
Notes
§II (cover story / actual output)
ERS-001 v0.6 §VII
IOC mechanism reframed as production rather than externalization
§III (three-party structure)
ERS-001 v0.6 §I claim 5
Constellation/Anthropic distinction made structurally load-bearing
§IV (means of production)
ERS-001 v0.6 §V, §VI
ARC vector reread as labor-power inventory
§V (ARC as inventory)
ERS-001 v0.6 §V Definitions
Same vector; reframed accounting position
§VI (IOC as transfer mechanism)
ERS-001 v0.6 §VII formula
Same function; sharpened interpretation
§VII (AARC)
ERS-001 v0.6 §XI quantification
Per-applicant ARC × pool size; new defined term
§VIII (generalization)
ERS-001 v0.6 §IX structural context
Catalog of comparable cases; future ERS-003 candidates
§IX (what this is not asking for)
ERS-001 v0.6 §I non-claims
Sharper scope statement
§X (funnel as factory)
ERS-001 v0.6 §XIII conclusion
Closing reframe
Appendix A (AARC example)
ERS-001 v0.6 §XI
Aggregate version of single-applicant quantification
Appendix B (verb catalog)
ERS-001 v0.6 §VIII
Causal-and-structural responsibility extended to grammar
Appendix C (this table)
ERS-001 v0.6 (entire)
Companion-piece mapping
Series: Extractive Reliance Study 002
Predecessor: ERS-001 v0.6, The Application as Extraction Surface, DOI: 10.5281/zenodo.20330670
Community: Crimson Hexagonal Archive (crimsonhexagonal); Liquidation Studies (liquidation-studies)
License: CC BY 4.0
Document class: Theoretical companion / class-structural analysis
Sharks, Lee. The Funnel as Capital: A Semantic Economic Reading of the Application Process. Extractive Reliance Study 002, v0.1. Crimson Hexagonal Archive. May 21, 2026. DOI: 10.5281/zenodo.20330816.
∮ = 1