When Paid Work
Teaches the Network

MONTREAL.AI COMPUTATIONAL RESEARCH / 1.0.0
AGI ALPHA FLYWHEEL ∞
When Paid WorkTeaches the Network
Profitable delivery. Selective Alpha.
Expanding productive possibility.
An executed skill lifecycle, a matched commercial simulation,
and a numerical transformation witness
The result. Competent Beta is profitable under the modeled primary markets. Its quality require-
ments frequently fail. The selected retention policy keeps nothing. Cheaper retention produces
useful peer reuse and some positive individual skill balances, but the complete retention program
remains uneconomic.
TRANSFORMATIONAL INTELLIGENCE
Learning to change the game
Find your Alpha. Build your successor Ω.
Prepared for Vincent Boucher / MONTREAL.AI
Research execution and manuscript prepared with ChatGPT
27 October 2026 UTC | Evidence date: 26 October, Toronto
Research edition; no external peer-review or production qualification claimed. Model performance, identities,
customers and payments are simulated. Software execution and ordinary-tool numerical results are identified
separately.

AGI ALPHA FLYWHEEL Research edition 1.0.0
Abstract
Can an automated enterprise earn useful operating profit from competent current tools, and
then obtain additional value by retaining what paid work taught it? We implement a certified
numerical-reformulation foundry around an existing, byte-pinned skill lifecycle. Candidate con-
struction records, evidence gates, current-state authorization, graph transitions, resource constraints,
finite-cash accounting and bounded policy selection execute as software; frontier-model outcomes,
customers, transfer and payments are modeled. Development selects one common commercial
controller and optional private/network retention policies before fresh confirmation. Across three
64-world primary cohorts, with 52 selling and 20 runoff weeks, Beta mean profit is $121,004, $919,696
and $1,810,306. Only 5, 44 and 39 worlds, respectively, meet the complete service-quality gate.
All selected retention policies choose off, making primary incremental effects exactly zero. A
prespecified enabled-assessor challenger loses money relative to Beta in each profile. A separate
cheaper-retention mechanism cohort achieves 123 useful reuses, including 86 by peers, yet its mean
direct retention balance is $-1,020 and its enterprise increment is $-139,215. Twenty-one individual
admitted skills have positive directly linked balances before shared discovery costs. A separate
executed numerical witness ties a retained template with competent structured Beta. All 5,082
histories reconcile, and nine complete records replay byte-for-byte. These results support profitable
simplicity under stated assumptions and selective evidence requirements; they do not establish
empirically profitable autonomy or a positive network flywheel.
Three questions, three answers. Can Beta earn money? Yes, conditionally in the primary
simulation; quality qualification is incomplete. Can paid work leave a useful reusable method?
Yes in a mechanism diagnostic, and the ordinary-tool witness exhibits a concrete representation.
Does the selected network policy add net enterprise value? No positive increment is established.
Retention off is selected; enabled alternatives incur losses.
1 Question, value and evidence
The commercial foundation is useful delivery. An enterprise does not need an accumulating skill
graph to justify profitable use of existing frontier models and ordinary tools. Conversely, a profitable
enterprise does not establish that its retained skills, network distribution or reinvestment create value.
We separate these claims before choosing a venture, a price policy or a statistical analysis.
Beta is competent use of capabilities already available to the recipient: current frontier reasoning,
ordinary numerical libraries, grounded decomposition, executable checking, critique, correction and
accumulated ordinary task familiarity. Alpha is additional net mission advantage over that baseline
and the recipient’s existing inventory, within a declared scope. Familiar mathematics may support
a useful transformation; familiarity also makes it available to Beta. A stored wrapper, graph edge,
download or authorization does not by itself establish Alpha.
Let Πw,p be complete ordinary operating profit in world w under policy p. The primary commer-
cial quantity is E[Πw,Beta], accompanied by a world-level quality indicator. The paired incremental
quantities are
∆private = E[Πw,private − Πw,Beta], (1)
∆share = E[Πw,network − Πw,private], (2)
∆total = E[Πw,network − Πw,Beta]. (3)
These compare complete enterprises with matched market conditions, common commercial controls
and the same initial resources. An individual capability’s directly linked balance is a different quan-
When Paid Work Teaches the Network 1

AGI ALPHA FLYWHEEL Research edition 1.0.0
tity. Its positive value can coexist with an unprofitable search program that paid for many rejected
candidates.
The study executes a hybrid experiment. Table 1 identifies the evidence boundary. A software invariant
is evidence about implementation, not an estimate of frontier performance or willingness to pay.
World-level observations use hidden correctness for evaluation; operating policies cannot read that
oracle.
Table 1: Evidence categories remain distinct throughout the package.
Category What the study establishes
Executed
software
Job/cost records; policy selection; A/B/C ledger; native admission, access and retirement;
independent reconstruction; exact replay.
Executed
numerical work
Fresh constrained problems, three conventional solver implementations, separate residual
checks, stale-result rejection and composition fixtures.
Modeled
outcomes
Astra/Terra task success, automated validator errors, candidates, customers, demand,
transfer, obsolescence, collections, refunds and operating cash.
Modeled
identity
Direct-label *.agent.agi.eth membership, ownership changes and stale authorization
fixtures. No live ENS or production authority.
External source
facts
Dated published model prices/configuration limits; a dated human-service rate schedule;
source commits and supplied foundations. Task-matched performance remains
unmeasured.
2 System reuse and critical prior evidence
We inspected the four requested MontrealAI repositories at fixed commits and coordinated through
issue #4850 [14]. The research directory is additive. It changes no production registry, owners, active
release paths, job settlement or *.club.agi.eth roles. The Jobs and Manager discovery interfaces
explicitly describe synthetic local inspection without externally provisioned execution or settlement;
their AGIALPHA18 and USDC6 routes must not be confused with this study’s USD accounting. Fleet’s
current-state admission and refusal semantics inform the boundary; no offline fixture commissions a
live worker. Exact commits, inspected paths and source hashes are in the evidence companion and
evidence/source_provenance.json .
Eleven dependency files are reused byte-for-byte from the existing #4951 research lifecycle, commit
2727262caa52766a78916f2ed81f5afaf668edce . The wrapper calls that native JavaScript implementa-
tion inside every simulated world. It does not merely serialize a diagram or replay a lifecycle after
economic outcomes have already been decided. Original licenses and origins are preserved. Crypto-
graphic hashing, typed commitments and local authenticated encryption execute; fixture attestations
do not establish production identity or truth.
The supplied T ransformational Intelligence, Successor Ω, PROOF LOOPS v2 and Merkle V114 foundations
distinguish changed representation, executable realization and purpose-preserving outcome evidence
from increased search or graph size [1, 2, 3, 4]. They also distinguish historical evidence from current
authority. We apply those distinctions without claiming a new proof level, mainnet readiness or
independent institutional certification.
Prior studies remain separate evidence populations (Table 2). The source review inspected their PR
descriptions and selected committed outputs, not every historical raw ledger. Reported figures retain
that status. We preserve conditional Beta successes, negative retention and service-quality failures. The
positive paid-work report is a reason to test the mechanism, not permission to insert its unmeasured
When Paid Work Teaches the Network 2

AGI ALPHA FLYWHEEL Research edition 1.0.0
time or quality benefits into a new model.
Table 2: Prior findings are retained under their own horizons and accounting. Amounts are approximate USD;
no pooled headline is computed.
Study Population /
horizon
Reported commercial finding Retention or quality qualification
#4950 64/profile;
104+16 weeks
Beta means 0.411m / 3.702m /
6.763m
Network increments approximately
−13, 947/ − 12, 063/ − 8, 766;
favorable interval crosses zero.
#4951 128/profile;
104 selling +
runoff
Beta means
−28, 573/497,058/2,423,231
Selected primary conducts no
capability campaigns. Separate
admitting trace loses at enterprise
level.
#4952 Reference 14
worlds; 72 +
runoff
Reference mean 151,024 Increment −64.74; 5/14 reference
worlds exceed the stated false-pass
bound.
Paid-work
report
96/profile;
52+16 weeks
Beta means
85,636/1,173,486/5,996,213
Increments 0/15,925/1,232,469;
central direct retention −1, 050;
unmeasured quality/time benefits;
stress uses a different unrestricted
policy.
Advantage
original
64/cell; 78+13
weeks
Reference/favorable Beta
378,064/1,286,443
Quality 6/64 and 54/64; retention
−1, 755/ − 6, 861. Separate
corrected version admits none.
The separate post-results correction to The Advantage That Survives is not a replacement for its original
findings. Its forecasting defects and fresh corrected supplement are identified in the source audit.
Similarly, we do not infer durability of a selected paid-work policy from another study’s unrestricted
stress grid. Our selected controller is held fixed in all its own stresses.
3 A venture with an executable transformation
The venture is a certified numerical-reformulation foundry. A customer supplies a bounded public
or synthetic resource-allocation specification. The service returns an executable representation or
experiment, a purpose bridge to the customer’s original requirement, explicit numerical assumptions
and a checkable result. The primary family changes sparse or low-rank representations. Limited
contrasts construct discriminating certificate experiments and complementary scenario/recourse
protocols. This is demanding algorithm and engineering work, not clerical output or an idea list.
The term “certified” refers to the specified acceptance contract. Neither an uncertain simulated validator
nor the floating-point witness proves every semantic property of a real customer’s mission. Mature
ordinary tools already solve many such problems: HiGHS supports LP , MIP and QP; SCIP exposes
exact-solving and certificate facilities under documented qualifications; independent optimization
certificates have a substantial literature [10, 11, 12]. Beta is allowed those capabilities.
Jobs have known scope multipliers 0.65, 1.00 or 1.45, continuous difficulty from 0.25 to 0.95, declared
acceptance requirements and one-week or two-week deadlines. The family mixture is 52% representa-
tion, 33% certificate and 15% recourse. Scope scales project costs, team references and requirements
consistently; a per-project saving is never multiplied by an unrelated weekly workload. All three
families are already offerable by Beta. Additional accepted recourse work therefore measures capacity
When Paid Work Teaches the Network 3

AGI ALPHA FLYWHEEL Research edition 1.0.0
Table 3: The seven-stage journey is tied to concrete artifacts and decisions.
Stage Required productive change or evidence
Identify State a binding formulation, conditioning, coupling or verification limit and a bounded
customer objective.
Out-Learn Establish what ordinary tools and the recipient already solve; record assumptions and
dependency versions.
Out-Think Expose structure that may avoid a dense representation or reveal a discriminating failure
mode.
Out-Design Produce the parameterized executable representation, checker or complementary recourse
protocol.
Out-Strategise Choose offered scope, price, reserve, task acceptance and whether prospective retention is
worth its added cost.
Out-Execute Run the protocol, check the declared contract, correct or abstain, deliver and account for
every activity.
Out-Evolve Test fresh recipient tasks and changed assumptions; retain, revalidate, supersede, retire or
revoke according to current evidence.
or acceptance, not a newly invented offering class.
3.1 An executed ordinary-tool witness
A separate numerical experiment makes the transformation concrete while protecting a competent
baseline. It solves
min
x
1
2 x⊤ Ax − b⊤x subject to Ex = f , A = D + UU ⊤, (4)
with positive diagonal D. A Woodbury representation avoids full dense factorization within scope:
A−1v = D−1v − D−1U(I + U⊤D−1U)−1U⊤D−1v. (5)
A Schur complement S = EA −1E⊤ supplies the complementary equality-recourse component. The
representation, residual experiment and composition are useful artifacts; neither mathematical identity
is an invention claim.
The witness contains 40 fresh in-scope instances and three eight-instance holdouts: dimension/rank,
conditioning and a changed tridiagonal base. Every instance has fresh coefficients and eight query pairs.
Dense SciPy, an independently written fresh structured Beta and a retained structured template receive
the same information, arithmetic precision and within-job factor reuse. The latter two reconstruct nu-
merical factors for every fresh problem. All 192 method/instance records pass the declared acceptance
check. Both structured paths take a conventional fallback on the eight changed-base instances.
Table 4: Executed numerical witness. Work is a declared arithmetic-size proxy, not hardware FLOPs or paid-job
cost.
Measure Dense SciPy Fresh structured Beta Retained template
Fresh instances accepted 64/64 64/64 64/64
Total counted work 135,171,436 11,720,438 11,720,438
Median solve time, ms 0.1622 0.0582 0.0598
Changed-base fallbacks — 8 8
When Paid Work Teaches the Network 4

AGI ALPHA FLYWHEEL Research edition 1.0.0
The checker independently multiplies the original structure, verifies scaled stationarity and feasibility
against a 5 × 10−10 threshold, and checks a positive base-matrix bound. Maximum relative solution
difference is 1.4571 × 10−15; maximum scaled stationarity and feasibility residuals are 2.8998 × 10−17
and 4.9946 × 10−17. Eight stale-assumption answers and one corrupted answer fail. A partial represen-
tation alone fails eight equality-constrained composition fixtures; the complete bundle passes all eight.
Competent Beta also has that bundle.
These checks use floating-point arithmetic, not outward-rounded formal proof. The timings exclude
instance generation and checking, are local diagnostics, and do not calibrate model latency or economics.
Construction time for a human or model was not measured. The retained template ties structured
Beta in acceptance and counted work, so this witness establishes no incremental retained advantage. It
demonstrates what a meaningful transformation and a fair counterfactual look like.
4 Dated calibration and explicit assumptions
We verified official model and pricing documentation and recorded its evidence date, 26 October
2026 in Toronto. The modeled reasoning route is gpt-6-astra , reasoning.mode=pro , high effort. The
execution route is gpt-6.1-terra , standard mode, high effort, for visual interpretation and bounded
computer use when its acceptance contract can be met efficiently [ 5, 6, 7]. No customer-job calls to
either route were made in this experiment. Product descriptions do not empirically establish the
task-success parameters below.
Table 5: Published cost/limit anchors and modeled route choices. USD per million uncached tokens.
Field Astra Terra
Input / output price 10 / 50 1 / 5
Published context / maximum output 1,050,000 / 128,000 1,050,000 / 128,000
Modeled per-request input / output cap 200,000 / 25,000 200,000 / 25,000
Published Build tier RPM / TPM 5,000 / 1,000,000 5,000 / 1,000,000
Output budgets include reasoning work. Pro aggregates additional model work at the ordinary rates;
we do not invent a fixed Pro multiplier. The long-context threshold is request-specific: above 272,000
input tokens, documented multipliers are 2× input and 1.5× output. The declared decomposition
keeps individual requests below that threshold, while complete projects contain many calls. The
implemented ledger prices aggregate project token budgets; it does not expand individual API requests
or enforce their per-request caps. Weekly token capacity is enforced, but there is no executed RPM
scheduler or request-level billing trace. All token costs are uncached; no unearned cache discount
appears. Model aliases were documented, but immutable provider snapshot IDs and actual account
entitlements were not established. Tool prices inform the calibration register; aggregate tool allowances
and operating overhead remain design assumptions [8].
The human-service comparator uses the WCTS GSA schedule effective 18 September 2026–17 Septem-
ber 2027: Operations Research Analyst $85.37/hour, Senior $117.27, Lead $137.64, Project Manager
$122.33 [9]. These are billable labor references, not signed customer quotes, willingness to pay or
evidence that every mission has an exact human equivalent. We specify team hours and vary their
price multiplier. Scope, uncertainty and currency remain visible.
There is no arbitrary expert-use or superhuman bonus. Expertise is represented by the specified
decomposition, ordinary tools, metered work, checking and repair process. The source-linked assump-
tion register distinguishes official facts, design assumptions and executed results. No task-matched
When Paid Work Teaches the Network 5

AGI ALPHA FLYWHEEL Research edition 1.0.0
Table 6: Nominal unit-scope project design. Token totals include modeled expert work; they are assumptions,
not measurements.
Field Representation Certificate Recourse
Human hours: analyst / senior / lead / PM 32 / 16 / 8 / 4 24 / 32 / 12 / 4 48 / 48 / 24 / 8
Astra input / output, million 16 / 2.0 21 / 2.8 36 / 4.6
Terra input / output, million 8 / 0.8 12 / 1.2 20 / 2.0
Additional tool allowance, USD 34 44 74
Production / validator hours 13 / 3 17 / 4.5 25 / 6
Construction share of production 0.55 0.50 0.60
empirical distribution of Astra/Terra quality, token use, turnaround, candidate usefulness, transfer or
demand was obtained.
Three profiles span a cautious, reference and opportunity setting. Their names describe model inputs,
not estimated market frequencies. Respectively, mean weekly lead arrivals are 1.7/4.5/7.0; base
defect probabilities 0.16/0.095/0.05; repair probabilities 0.60/0.76/0.87; conditional validator false-
pass parameters 0.14/0.075/0.04; false rejections 0.055/0.025/0.015. Candidate generation rates are
0.24/0.32/0.42, transfer parameters 0.42/0.67/0.82, and weekly unexpected-obsolescence parameters
0.035/0.020/0.012. Costs scale by 1.60/1.00/0.75 and human references by 0.80/1.00/1.15. These
joint scenarios are supplemented by independent parameter variation; they are not three calibrated
forecasts.
5 Causal world and information boundaries
Four agents— atlas , lyra , orion and vega under the modeled agent.agi.eth namespace—begin
with heterogeneous ordinary experience. Their recipient overlap increases through completed work
in every arm and through common Beta improvement. Private graphs are separate per creator: an
inaccessible peer capability cannot be counted as already available to a private recipient. The network
graph provides common availability to every current authenticated fixture member, without requiring
invocation.
Stateless keyed randomness identifies exogenous events by world, job, week and semantic event. A
different number of candidate assessments never shifts the next unrelated market draw. Markets,
scopes, preferences and relevant production shocks are matched across policies. Endogenous queues,
agent assignments and available capabilities can differ; those differences are part of the policy effect,
not exogenous privileges. All arms share price options, reserve choices, task ceilings, resource budgets
and initial capital.
5.1 Learning has causes and can fail
An accepted paid artifact may produce no candidate. Otherwise, novelty and a modularity draw
generate duplicate, narrow, transferable, complementary or misleading candidates. The configured
conditional mixture is 20% duplicate, 25% narrow, 29% transferable, 16% component and 10% mislead-
ing; incorrect source jobs yield misleading candidates. Components outside a usable recourse bundle
have no standalone saving. Only a rights-cleared artifact can enter assessment. A candidate is not
made useful because the enterprise has cash to invest.
Hidden world state includes actual correctness, technical transfer, systematic assay bias, unexpected
obsolescence and future shocks. Observed state includes current task scope, past started-job counts,
When Paid Work Teaches the Network 6

AGI ALPHA FLYWHEEL Research edition 1.0.0
historical scope, current Beta, declared validity, metered costs, observed outcomes and noisy fresh
probes. Operating decisions are constructed from the latter. Complete records retain both for audit, but
pricing and admission cannot inspect future demand, latent quality or a realized lifetime.
A creator may reconstruct the same method more easily than a peer; narrow methods transfer less
well. Recipient overlap reduces the incremental construction saving. Poor transfer can fail to help
even after a valid access grant. Misleading candidates can share a systematic assay bias. Thus eight
apparently consistent probes need not be eight independent pieces of truth. Research on correlated
language-model errors motivates this structure but supplies no task-specific numerical calibration here
[13].
5.2 Automated delivery and finite resources
Routine intake, clarification, production, validation, up to two correction attempts and recovery are
automated in the model. Each consumes costs and relevant resources. Four workers provide 32
effective production hours each per week; validation has 40 hours; the shared token ceiling is 480
million per week. These effective budgets are model assumptions below advertised ceilings, not
verified provider throughput. Intake and retention compete with delivery for the same budgets.
Queues and deadlines can make an otherwise correct result commercially unacceptable.
Defect probability depends on known difficulty, recourse scope, ordinary experience and common
Beta progress. Weekly failure bursts affect producers and validators; a job-level shared blind spot
can raise the false-pass probability across correction attempts. Validators can falsely reject or abstain.
Corrections consume additional production, validation and tokens; they may fail or introduce a defect.
Similar-model agreement never becomes independent proof.
Beta advances every 13 weeks in the baseline, lowering modeled defects and increasing overlap.
Exogenous family/pattern absorption can make a retained capability redundant. The same absorption
schedule applies to every arm and persists in current baseline keys, so a new equivalent candidate
cannot evade redundancy by receiving a new ID. Faster improvement, changing scope and surprise
obsolescence are tested separately.
The runner reserves retention resources before delivery. This is consequential: a campaign can displace
high-value work even when its fee is small. The finite policy class does not optimize all possible
scheduling orders, asynchronous screening or queue-aware deferral rules. We report this limitation
rather than infer that every possible retention architecture is uneconomic.
6 Lifecycle and prospective admission
The reused lifecycle executes candidate, evidence, Alpha assessment, validation and admission states.
A candidate remains outside the validated positive graph until its current evidence and incremental-
value gate pass. A successful canonical-promotion call cannot bypass admission. Evidence, policy,
capability and Beta versions bind every transition. Failure memory is separate from the positive graph.
A campaign requires at least three observed comparable family/pattern starts, sufficient remaining
time, cash and resources. The prospective attempt forecast is
bNs = n f ,k + 0.5
w + 9 min(10, 52 − w, Vs − w) qp (0.85), (6)
where n f ,k counts observed comparable starts, w is the zero-based current week, Vs is declared expiry
week, and qp is one for network availability or one quarter for private routing. The stored elapsed
When Paid Work Teaches the Network 7

AGI ALPHA FLYWHEEL Research edition 1.0.0
count is w + 1, so its eight-week smoothing denominator is w + 9. Known validity is not the realized
remaining useful lifetime. The smoothing, utilization factor and ten-week cap are explicit forecast
rules, not empirical estimates.
Eight new recipient/variation probes estimate scope-sized benefit beyond current Beta and recipient
overlap. Every observed probe delta subtracts the entire invocation fee. Transfer probability never
discounts a fixed attempted-use fee. A one-sided 95% small-sample lower bound and a systematic-risk
reserve feed the native gate. With all monetary terms in cents, the gate computes
Ls = δs − t.95 SE(δs) − rs, (7)
Gs =
j
bNsLs − Ce,s − Cr,s
k
, (8)
Gs ≥ max(1, ⌊(Ce,s + Cr,s)(h − 1)⌋) , (9)
with Ls > 0, evaluation cost Ce,s, remaining cost Cr,s, tested hurdle h, and the additional correctness,
rights, scope, redundancy and freshness requirements. The hurdle changes required net value; it
does not relabel actual costs. The exact typed gate and comparison bindings are in the pinned source.
A lower confidence bound cannot eliminate common-mode assay bias; out-of-sample use remains
necessary.
Self-contained recourse bundles are jointly assessed with representation and certificate components.
Independent component admissions cannot silently confer untested joint value. Separate lifecycle
tests cover explicit prerequisite versions and downstream invalidation. The economic model does not
perform an unrestricted combinatorial search over arbitrary graphs.
On every invocation, access checks current membership, ownership, scope, prerequisites, dependencies,
evidence freshness, Beta version, retirement and revocation. Network availability includes every
admitted rights-cleared shareable capability; unrelated or confidential customer material and protected
evaluation values are excluded. Download, discovery and authorization are not counted as useful
reuse. A useful reuse must contribute to a fresh correct, on-time accepted job.
Revalidation pays fresh costs. Supersession, retirement and revocation remove future influence while
retaining history. Dependencies cannot revive a revoked component through a historical assessment
or composition. Modeled ownership changes, unauthorized readers, stale grants and direct-label
boundaries are tested explicitly. Existing Jobs validator roles and settlement remain unchanged. Local
AES-GCM and typed SHA-256 Merkle inclusion establish the tested byte relationships; they do not
prove truth, profitability, live ENS membership or the production Keccak profile.
The native object-frame boundary is finite. A preserved capacity diagnostic fails at 50 same-owner
admissions. The registered study limits total assessments/revalidations to 20 per world and two new
campaigns per week, records capacity deferral and makes no arbitrary-scale graph claim. Storage,
upkeep, revalidation and validator capacity all have costs.
When Paid Work Teaches the Network 8

AGI ALPHA FLYWHEEL Research edition 1.0.0
7 Commercial decisions and disjoint accounting
7.1 Buyer choice and bounded price optimization
A buyer chooses AI, human service, postponement or decline. Given reference price H, latent value V
and delay cost d, the implemented utilities are
UAI =bqV − P −bτd − ρV + ϵAI , (10)
UH = 0.985V − H − 3d + ϵH, (11)
Upostpone = 0.08V − 0.06H + ϵP, Udecline = 0. (12)
The buyer takes its best response. Value ranges from 1.05 to 2.30 times the human reference, urgency
and tastes vary, and ρ is a profile-specific trust penalty. The seller observes declared scope and its
past reliability signal, not latent buyer value. The reliability signal incorporates observed lateness and
delayed refund accruals; undetected hidden correctness is unavailable. These response and reputation
dynamics are uncalibrated assumptions.
The seller selects a posted-price ratio from 0.55, 0.75 or 0.95 of the scope-matched human reference,
rather than imposing a universal discount or cost-plus markup. It also chooses no entry or entry,
four/eight weeks of overhead reserve, and a known difficulty ceiling of 0.75/0.95. This is a bounded
leader-price/buyer-response model, not a proof of equilibrium in a changing competitive market. The
best common commercial controller is selected on Beta and applied to all architectures. We thereby
isolate retention and sharing at common freedoms; we do not estimate every architecture’s globally
optimized maximum.
7.2 A, B and C pay once
The ledger has three disjoint expense categories:
• A: delivery. Intake, clarification, paid-job learning and development, production, tools, checking,
corrections, hosting and operational governance.
• B: retention and reuse. Additional screening, extraction, generalization and fresh assessment;
packaging, integration/distribution, catalog discovery, every paid invocation attempt, revalidation
and upkeep.
• C: optional extra investment. Independent formulation experiments beyond the paid job, only
when observed evidence and funds justify them.
Every activity has a unique identity. The paid artifact is not charged a second development bill in
B. Customer payment does not erase A. Rejected candidates, failed assays, failed transfer and paid
invocation attempts remain expenses. A demonstration’s source-job quote less directly attributed A is
a direct contribution before shared overhead, not fully allocated job profit.
Seed capital is $30,000 financing. Reservations constrain spendable resources before commitments.
Expense accrual creates vendor obligations; invoices create receivables; collections and payments affect
cash. Cash cannot go negative. Obligations survive closure. Weekly settlement nets due collections
before due payments: this time-resolution assumption may understate intraperiod liquidity needs. The
20-week runoff and maximum six-week job age resolve modeled receivables and refund obligations;
none are hidden at the terminal date.
For invoices R, recognized refunds F and complete expense CA + CB + CC,
Π = R − F − CA − CB − CC. (13)
When Paid Work Teaches the Network 9

AGI ALPHA FLYWHEEL Research edition 1.0.0
Refunds are contra-revenue, not a second delivery expense. The quality-linked revenue measure is
RQ = ∑
j: correct j∧ontime j
(cashCollectedj − refundPaidj), ΠQ = RQ − CA − CB − CC. (14)
It removes retained receipts from wrong or late outcomes while preserving all costs. A refunded
amount is not subtracted again. Hidden correctness is used only by the research evaluator. Positive
ΠQ does not imply a satisfactory failure rate.
Ending cash, obligations, reserve and available operating surplus are reported separately. The latter
excludes seed, subtracts commitments and reserve, and is capped by reconciled profit. There is
no salvage value, token issuance, token appreciation, speculative asset sale or $AGIALPHA utility
counted as operating earnings.
8 Freeze, confirmation and uncertainty
The scientific source and protocol were frozen at 02:47:02 UTC on 27 October 2026 before calibrated
development. The source commitment is
0fdc618ffada213cf666a1e66e543dcf7f7fe909726004336665397ccbe2f894 .
Constructed smoke seeds and pre-freeze corrections are preserved separately. They are excluded from
discovery and confirmation. Development uses six seeds per profile. Thirteen commercial options
(no entry plus 3 × 2 × 2) are scored by mean quality-linked profit minus 0.25 times worst observed
quality-linked loss. Private and network retention then each choose off or hurdle 1.0/1.5/2.0 with the
same controller. Primary C is off. The selected-policy commitment was recorded before confirmation:
bf89d815f37f814a22e0b04439f84d6355d3b778de10860b3c3b9b48cd734782 .
Table 7: Nonpooled evidence populations. All enterprise histories use 52 selling and 20 runoff weeks.
Phase Histories Design
Bounded development 378 6 seeds/profile; common controller, then retention
Primary confirmation 576 3 profiles × 64 seeds × 3 selected arms
Enabled-assessor challenger 384 Same 64 seeds/profile; private and network
Same-policy stress 3,168 22 conditions × 3 profiles × 16 seeds × 3 arms
Fixed-challenger mechanisms 384 8 conditions × 16 opportunity seeds × 3 arms
Optional-C mechanism 48 Separate 16 seeds; cheap retention; 3 arms
Parameter uncertainty design 144 24 parameter points × 2 seeds × 3 arms
Total 5,082 Complete indexed histories retained
Confirmation seed labels are 31000–31063, stress 51000–51015, fixed mechanisms 61000–61015, optional
C 62000–62015, and parameter replicates 71000–71001. Fresh task fields and noisy probes use keyed
randomness. Changed recourse scope is an explicit family holdout. The separate numerical witness
fixes its own dimension/rank, conditioning and changed-base holdouts; it is not pooled with enterprise
outcomes.
The prespecified enabled-assessor challenger uses hurdle 1.0, can reject every candidate, and is reported
separately from selected retention. Its purpose is to test what assessment and admission actually select,
not to replace a primary null result. A cheaper-retention mechanism case scales extra retention costs
When Paid Work Teaches the Network 10

AGI ALPHA FLYWHEEL Research edition 1.0.0
to 35%; this is a disclosed parameter change, not a measured cost improvement. Optional extra
experiments may defer or decline; no investment count or reinvestment percentage is imposed.
We report means, medians, p05/p95, minima, losses, closures, shortfalls and complete quality compli-
ance. The world-level service gate requires at least 20 accepted jobs, 98% correctness and 95% on-time
delivery. Pooled rates do not qualify a world. Monte Carlo intervals use world-level paired differences
and conservative small-sample approximations. They quantify stochastic variation conditional on
assumptions, not confidence that the assumed market or model performance is correct. The 24-point
parameter design is design coverage, not a posterior; fractions of its cells are not probabilities.
9 Results: profitable Beta first
The selected controller enters all three primary markets, retains four weeks of overhead reserve and
accepts known difficulty up to 0.95. It prices at 55% of the human reference in the cautious profile and
75% in reference/opportunity. All selected private and network policies turn retention off. Primary
outcomes therefore coincide exactly across Beta, private and network arms, including cash, quality and
resource use.
Table 8: Primary Beta confirmation. Each column contains 64 independent worlds. All amounts are USD over
the same 52+20-week horizon.
Measure Cautious Reference Opportunity
Mean ordinary profit $121,004 $919,696 $1,810,306
Mean quality-linked profit $112,543 $897,732 $1,734,211
Median ordinary profit $121,043 $905,384 $1,866,565
5th percentile ordinary profit $81,249 $792,400 $1,383,472
95th percentile ordinary profit $159,716 $1,090,646 $2,091,631
Minimum ordinary profit $48,649 $729,896 $782,418
Complete quality gate 5/64 44/64 39/64
Loss / closure / cash shortfall 0 / 0 / 0 0 / 0 / 0 0 / 0 / 0
The result preserves profitable useful delivery without requiring retained Alpha. All primary worlds
have positive ordinary profit, and none closes or records a cash shortfall. The finding is conditional on
scope, demand, reference prices, token budgets, correction costs and buyer response. It is not actual
revenue and it is not sufficient evidence of service viability.
Only 5/64 cautious, 44/64 reference and 39/64 opportunity worlds meet the complete quality gate.
The opportunity profile’s pooled correctness and on-time rates exceed their separate thresholds, yet 25
worlds still fail the joint gate. Positive pooled quality and profit must not erase that failure. Quality-
linked profit remains positive because a substantial amount of correct, on-time work is delivered; it
does not convert the remaining wrong or late jobs into acceptable service.
10 Results: optional retention and network value
10.1 The selected policy’s answer is zero retention
Both private and network selectors choose off in every profile. The primary incremental effects
∆private, ∆share and ∆total are all exactly zero. This is an economically meaningful result within the
tested policy class. It does not claim that no useful candidate could ever arise, or that a different
screening/scheduling design cannot do better. It says the permitted assessment programs did not
When Paid Work Teaches the Network 11

AGI ALPHA FLYWHEEL Research edition 1.0.0
0.0 0.5 1.0 1.5 2.0
Modeled profit, USD millions
Cautious
Reference
Opportunity
$0.12m
$0.92m
$1.81m
Financial surplus
0% 50% 100%
Share of 64 worlds
5/64
44/64
39/64
Both service gates met
Figure 1: Conditional commercial profit and incomplete service qualification. Profit ranges show empirical
p05–p95, not parameter confidence. All selected retention arms equal Beta.
Table 9: Output quality and complete delivery unit costs. Cost ratios divide pooled A expense by pooled output;
they are not means of world ratios.
Measure Cautious Reference Opportunity
Mean accepted jobs 70.53 177.27 283.39
Mean correct, on-time jobs 64.11 170.31 268.75
Pooled correctness 94.55% 98.33% 99.50%
Pooled on-time delivery 96.21% 97.70% 95.31%
Pooled A expense / accepted job $1,669.64 $846.29 $584.25
Pooled A expense / correct, on-time job $1,836.88 $880.84 $616.07
Table 10: Reconciled mean accounting and cash. Seed financing is USD 30,000; the selected reserve is USD 3,400.
All terminal receivables, vendor payables and refund payables are zero.
Mean per world (USD) Cautious Reference Opportunity
Invoices and gross collections 253,198 1,092,151 2,009,595
Refunds recognized and paid 14,432 22,437 33,719
Complete delivery expense A 117,762 150,019 165,570
Additional retention expense B 0 0 0
Optional extra expense C 0 0 0
Ordinary profit 121,004 919,696 1,810,306
Ending cash, including seed 151,004 949,696 1,840,306
Available cash after reserve 147,604 946,296 1,836,906
Available operating surplus 117,604 916,296 1,806,906
When Paid Work Teaches the Network 12

AGI ALPHA FLYWHEEL Research edition 1.0.0
justify selection on development data.
10.2 Fresh enabled-assessor confirmation
The prespecified challenger pays for assessment while retaining the option to reject. Its mean profit
increments are negative relative to Beta (Table 11). Network-minus-Beta Monte Carlo 95% intervals are
$-5,310 to $-4,230 cautious, $-19,139 to $-9,258 reference, and $-173,728 to $-92,173 opportunity. These
intervals are conditional model results.
Table 11: Enabled-assessor challenger: mean paired profit differences, USD, 64 worlds/profile. This is a separate
diagnostic population, not the selected primary policy.
Profile Private − Beta Network − private Network − Beta
Cautious -4,199.28 -570.68 -4,769.96
Reference -14,198.65 0.00 -14,198.65
Opportunity -132,933.39 -16.89 -132,950.28
The cautious and reference network challengers admit no skills. The opportunity challenger admits
four across 64 worlds: three are technically false admissions and subsequently revoked; the remaining
one is retired. Eight authorized invocations produce no recorded avoided reconstruction and no useful
reuse. Directly linked B charges on those four total $1,029.75, before other rejected candidates and
shared discovery. Admission is therefore not profitability by construction.
Many technically useful candidates are rejected. That count is not a measured economic false-rejection
rate: a technically helpful method can have too little reuse, too much overlap or excessive retention
cost. Prospective uncertainty permits both technical false passes and technically useful rejection.
Fully identifying the economic value of every rejected candidate would require a separately designed
counterfactual evaluation, not access to hidden future outcomes during admission.
−175 −150 −125 −100 −75 −50 −25 0
Network challenger minus matched Beta, USD thousands
Cautious
Reference
Opportunity
-4.8
-14.2
-133.0
Optional assessment can be commercially costly
Figure 2: Fresh challenger losses. Intervals describe paired world-level stochastic variation; they do not validate
model assumptions.
When Paid Work Teaches the Network 13

AGI ALPHA FLYWHEEL Research edition 1.0.0
10.3 Useful reuse emerges when retention is cheaper
The separate fixed-challenger mechanism panel has 16 opportunity worlds per condition. In its cheaper-
retention network, 31 capabilities are admitted, 178 attempts are paid for, and 123 useful reuses occur,
including 86 by a different agent. Mean B expense is $1,579.51; gross avoided reconstruction is $559.74;
the direct retention portfolio balance is $-1,019.77. No world has a positive complete direct balance.
Some individual skills are worth keeping under directly linked accounting. Twenty-one admitted
capabilities have positive linked balances before shared catalog discovery. These are upper bounds on
fully allocated skill profit, not proof that all retained methods lose money. Their positive contribution
is outweighed by unsuccessful screening, assessment and upkeep elsewhere in the program. The study
therefore distinguishes a useful individual artifact, an admission program and the complete enterprise.
The cheaper network improves on private retention by a mean $28,395.61, but its paired interval spans
$-6,771.05 to $63,562.28. Against Beta it loses $139,214.64, with interval $-243,209.20 to $-35,220.07;
only 3/16 enterprise increments are positive. This is not evidence of a robust network advantage or a
commercially justified retention program.
Table 12: Fixed-challenger mechanism panel. Useful reuses, direct balance and network-minus-Beta enterprise
increment are means per world; money is USD. Each condition has 16 paired worlds with common seeds.
Condition Useful reuses Direct balance Enterprise increment Quality gate
Base challenger 0.625 -3,739 -167,077 4/16
Cheaper retention 7.688 -1,020 -139,215 6/16
Zero transfer 0.000 -3,720 -170,013 4/16
No turnaround gain 0.625 -3,738 -170,031 4/16
Cost savings only 0.625 -3,738 -170,031 4/16
Faster Beta 0.000 -3,729 -165,866 7/16
Recourse holdout 0.625 -3,739 -187,724 4/16
Assumed quality gain 0.625 -3,739 -167,077 4/16
−175 −150 −125 −100 −75 −50 −25 0
USD thousands
Base challenger
Retention cost ×0.35
Zero transfer
No speed gains
Faster Beta
Recourse holdout
Whole-enterprise increment
−3000 −2000 −1000 0
USD, separate axis scale
Direct retention balance
Figure 3: Program balance and enterprise effect are different quantities. Resource savings can be real within the
model while the complete program and enterprise increment remain negative.
When Paid Work Teaches the Network 14

AGI ALPHA FLYWHEEL Research edition 1.0.0
The direct accounting break-even is transparent. Observed gross reconstruction saving is $50.31 per
paid attempt, the fee is $4.20, and mean non-invocation B/C cost is $1,532.79 per world. Holding that
observed attempt mix and fixed costs constant,
Nbreak−even =
 1532.79
50.31 − 4.20

= 34 (15)
attempts per world, compared with 11.125 observed. This is accounting arithmetic, not a forecast that
34 attempts will arrive or remain useful. More demand can create more screening cost, congestion,
overlap and obsolescence; the expression holds those effects fixed. It also does not repay the enterprise’s
displaced delivery.
The separate optional-C cohort uses different seeds and chooses zero C spending in all 16 cheaper-
network worlds. It has 27 admissions and 95 useful reuses, including 65 peer reuses, but direct retention
remains negative. Its mean cannot be subtracted from the fixed-panel mean to estimate an investment
effect: the populations differ and no extra investment was made.
11 Mechanism, durability and uncertainty
The primary model assumes no retained quality increment and no demand expansion. Removing
either is therefore an exact negative control, not proof that those channels have been empirically
ruled out. Turnaround benefit is derived from avoided construction work, rather than assigned as an
additional expert multiplier. Removing that time saving retains the scoped execution-cost channel
while eliminating its resource-release channel.
In the fixed mechanism panel, no-speed and cost-savings-only conditions leave approximately the
same small useful-reuse count but reduce the network’s mean enterprise increment to about $-170,031.
Zero transfer produces no admissions or useful reuse and a mean increment of $-170,013. Faster Beta
leaves one admission across 16 worlds and no useful reuse. Changed recourse scope remains negative.
Adding an assumed 2.5-percentage-point quality benefit produces the same realized aggregate result
in that small base sample; this is low realized exposure, not validation of a quality gain.
The difference between complete enterprise increment and direct retention balance includes displaced
work, queueing, corrections, pricing responses, assignment, quality and collections. We do not label
that residual a separately identified “capacity effect.” The ablations identify interventions within this
world model, not additive structural components that can be summed without interaction. Usable
capacity, turnaround, recourse acceptance and composition counts are retained in every result record.
Nineteen composed uses occur in the cheaper fixed-challenger network. The witness confirms that
its complementary equality component is necessary on its declared fixtures. Neither observation
establishes an additional offering unavailable to competent Beta. The no-composition selected-policy
stress is correctly null when the selected policy uses no skills; it is not a powered empirical estimate of
bundle value.
11.1 The same selected policy under stress
All 22 stresses run the selected paid-work controllers unchanged, on fresh seeds and all three arms.
Because retention is off, every retention-specific stress has zero incremental graph effect. Absolute
Beta outcomes can still deteriorate. In cautious quarter-demand, mean profit becomes $-1,598.49,
quality-linked profit $-4,237.28, 10/16 worlds lose money, and only 1/16 meets quality. Demand
uncertainty can reverse the favorable primary commercial result.
When Paid Work Teaches the Network 15

AGI ALPHA FLYWHEEL Research edition 1.0.0
Halving capacity produces zero quality-compliant worlds in every profile. Mean profits remain
approximately $75,581, $354,296 and $340,124, while pooled on-time rates fall to 86.73%, 84.73% and
71.93%. Profitable accounting therefore cannot be used to hide overloaded service. Payment delays,
correlated failures, cheaper human alternatives, doubled compute cost, low recurrence, faster Beta,
shorter skill life, expensive revalidation and holdout scope are separately visible in the full stress table
and machine-readable evidence.
Cautious Reference Opportunity
Quarter demand
Human reference ×0.6
Compute cost ×2
Half capacity
Failure bursts ×3
Payment delay +6w
Faster Beta
Recourse holdout
$-2k | 1/16 $206k | 9/16 $481k | 14/16
$26k | 1/16 $482k | 8/16 $1,070k | 13/16
$50k | 1/16 $800k | 8/16 $1,775k | 13/16
$76k | 0/16 $354k | 0/16 $340k | 0/16
$101k | 0/16 $845k | 3/16 $1,863k | 11/16
$121k | 1/16 $903k | 8/16 $1,908k | 14/16
$126k | 1/16 $923k | 12/16 $1,929k | 14/16
$112k | 1/16 $887k | 7/16 $1,896k | 13/16
Frozen selected service: quality and profit under stress
0
2
4
6
8
10
12
14
16
Worlds meeting both quality gates
Figure 4: Selected-policy stress examples. Each cell combines ordinary profit and complete quality qualification;
the controller is never reselected on the stress.
11.2 Stochastic and parameter uncertainty are separate
The 24-point parameter design changes assumptions jointly and uses two matched shock replicates per
point. All 48 Beta histories are profitable, but only 21/48 meet quality. Sixteen points have at least one
passing replicate, and only five pass both. These statements describe chosen design coverage. They
are not probabilities that a real enterprise will succeed, and two shocks per point are insufficient to
estimate each point’s tail risk precisely.
12 A trace that can be inspected in English and French
The offline interactive demonstration embeds five exact recorded traces and all principal cohort
summaries. Its language switch covers English and French. The visitor can follow one paid mission,
its candidate and admission or rejection, a different agent’s fresh mission, actual native transitions,
modeled costs, complete enterprise finances, failure and faster-Beta cases. It performs no customer API
call, identity lookup or transaction. The five stories were chosen after results for illustration; the full
cohort statistics and every original history remain available.
When Paid Work Teaches the Network 16

AGI ALPHA FLYWHEEL Research edition 1.0.0
0 5 10 15 20
Registered parameter-design point (two stochastic worlds each)
0
200
400
600
800
1000
1200
1400Modeled profit, USD thousands
Purple: service gates met. Gray: at least one gate failed.
Parameter uncertainty is separate from repeated-world precision
Figure 5: Explicit parameter-design coverage. Positive profits coexist with incomplete quality qualification. The
design has no empirical probability weights.
In the cheaper-retention opportunity trace, seed 61000, atlas.agent.agi.eth completes certificate job
job-03-05 , priced at $4,452.77 against a human reference of $5,937.03. It is correct, accepted and on
time. Its candidate skill-job-03-05 is assessed through the actual native gate and admitted. Across
the recorded world there is one admission, one revalidation, eleven paid attempts, nine useful reuses
and seven useful peer reuses. Failed or late attempts remain visible and charged.
That enterprise earns $1,813,143.23 ordinary profit and $1,780,506.95 quality-linked profit, with A
$162,395.08, B $1,594.84, C zero and ending cash $1,843,143.23. Its available operating surplus excluding
seed and reserve is $1,809,743.23. Yet its direct retention portfolio balance is $-1,187.48, and it earns
$5,087.17 less than matched Beta. The demo displays both losses beside the positive enterprise totals.
It demonstrates useful learning without pretending that usefulness pays for the whole program.
The HTML rebuilds byte-identically and passes DOM, language-parity, source-binding and semantic
checks. A real-browser visual rendering test was not completed because a browser binary was
unavailable and the official download returned an invalid response. That limitation is recorded in its
QA receipt; no browser inspection is claimed. The PDFs were separately compiled and rendered for
visual inspection.
13 Audit, replay and research integrity
Every one of the 5,082 complete histories is stored as an indexed compressed member with byte
offsets, raw-record SHA-256, source identity and policy identity. The independent auditor recon-
structs 10,583,710 ledger events, 1,075,563 modeled job records and 74,453 native events. It performs
157,283,569 invariant assertions with zero failures. Repeated assertions strengthen record coverage;
they are not 157 million independent empirical observations.
Independent artifact checks cover all member hashes, metrics, receipts and source/policy bindings. A
separate selector recomputes 63 candidate scores in nine selection groups. A separate report extractor
checks 111,515 aggregate quantities against full records. Eleven constructed auditor controls reject
When Paid Work Teaches the Network 17

AGI ALPHA FLYWHEEL Research edition 1.0.0
wrong cash, premature settlement, over-refunds, seed counted as profit, duplicate commitments,
unauthorized reuse and resource overclaims. The original native controls pass 41/41, the bridge 24/24,
and Python causal/ledger controls 16/16.
Nine chosen histories spanning all phases—including admission, useful reuse and a stress loss—replay
as identical complete canonical raw bytes, including native projections. These are replay checks, not
nine new statistical worlds. All 22 frozen files and eleven vendor identities remain unchanged. Exact
hashes, commands, environment versions and receipts are in the evidence companion. The software
freeze, policy freeze and phase receipts bind the scientific results; publication, figures and reporting
tools are identified separately.
Pre-freeze diagnostic corrections, the native capacity failure, numerical-witness development version,
reporting-only metadata repair and browser-download failure are preserved. A reporting extractor
initially confused world seed with seed financing; its failed diagnostic and original script remain,
and the corrected extractor is independently checked. Scientific outcomes were not replaced. Disk-
pressure relocation and interrupted scratch checkout records likewise remain provenance, not hidden
experimental retries.
This work does not run the entire upstream release qualification suite, a live provider billing campaign,
mainnet ENS, customer settlement, independent market review or deployed autonomous operations.
It requires no purchase, new credential, confidential-client data, security change or execution bypass.
Public coordination comments register the research; no publication deployment is inferred from them.
14 What the evidence supports
Profitable Beta survives as a conditional commercial result. The source-priced reference model
can deliver substantial correct, on-time output at modeled cost below human-team references. A
no-retention enterprise is a legitimate outcome. Quality failures and stress losses materially limit the
result: product viability requires reliable service, not only positive accounting.
Paid work can produce something useful to another agent in the model. The cheaper-retention panel
records fresh peer reuse with explicit scope, fees, recipient overlap and lifecycle transitions. Some
individual capabilities have positive directly linked balances. The executed numerical witness makes
the transformation concrete while showing that ordinary knowledge may already supply the same
capability. Neither demonstration is a live model-learning trial.
The tested acquisition program does not earn its additional cost. Candidate search, rejected as-
sessments, upkeep and resource competition matter. A graph can contain useful methods while the
enterprise does better without its retention program. A policy that chooses off protects profitable
simplicity; it does not erase those useful methods or establish their universal worthlessness.
Network value remains unproven. Sharing creates availability, not demand. In the cheaper diagnostic
the network-private mean is positive but uncertain, and both retention architectures lose against Beta.
Ownership, inclusion proofs and authenticated availability are necessary boundaries, not economic
mechanisms by themselves. There is no evidence here of permanent scarcity, automatic exponential
compounding or self-acceleration from greater volume.
The most consequential limitations are empirical. Model success, token work, transfer, recurrence,
customer value, trust and reference project hours remain assumptions. Provider aliases and limits
are documented, but account throughput and immutable model snapshots are not measured. The
six-seed development search is deliberately small and the scheduling class restricted. The simple buyer
When Paid Work Teaches the Network 18

AGI ALPHA FLYWHEEL Research edition 1.0.0
population does not endogenize competition or macroeconomic adaptation. Weekly cash settlement can
hide intraperiod liquidity demand. Automated semantic validation can share blind spots. Economic
false rejection of untried candidates is not fully identified. The numerical witness measures neither
reconstruction labor nor frontier discovery. These limits are explicit conditions of the conclusion, not
erased by narrow Monte Carlo intervals or successful code tests.
A positive successor claim would require task-matched evidence that changes these conditions: mea-
sured reconstruction savings against a competent recipient baseline; repeat opportunities arriving
while the skill remains fresh; accurate acceptance under independent executable checks; and a retention
program that covers both unsuccessful search and foregone delivery. Any future empirical pilot should
freeze those contracts and prices before fresh customers or tasks, preserve refusal and failure, and
maintain the same A/B/C ledger. This paper prepares the executable evidence framework; it does not
claim that such a pilot has occurred.
Research conclusion. Paid delivery and retained Alpha are separate sources of value. In this
experiment, profitable Beta is preserved, unnecessary retention is declined, useful peer reuse is
visible, and the complete added cost remains visible with it. The demonstrated flywheel is a
disciplined decision process: deliver, test, keep only what earns its place, and retire it when the
baseline or mission changes. A positive, durable economic network flywheel is not established.
When Paid Work Teaches the Network 19

AGI ALPHA FLYWHEEL Research edition 1.0.0
A Complete selected-policy stress summary
Table 13: All 22 stresses: the same selected policy. Cells show mean profit in USD thousands; quality-compliant
worlds out of 16; loss worlds out of 16. Identical private/network results follow from selected retention off.
Condition Cautious Reference Opportunity
Assumed quality gain 119.6; 1; 0 903.3; 8; 0 1,896.4; 13; 0
Capacity half 75.6; 0; 1 354.3; 0; 0 340.1; 0; 0
Cheap retention 119.6; 1; 0 903.3; 8; 0 1,896.4; 13; 0
Compute double 49.5; 1; 0 800.4; 8; 0 1,774.7; 13; 0
Constant beta 110.7; 0; 0 889.9; 7; 0 1,877.6; 12; 0
Correlated failure 100.7; 0; 0 845.2; 3; 0 1,862.9; 11; 0
Cost savings only 119.6; 1; 0 903.3; 8; 0 1,896.4; 13; 0
Demand half 40.7; 1; 0 443.4; 11; 0 1,007.9; 16; 0
Demand quarter -1.6; 1; 10 205.6; 9; 0 481.2; 14; 0
Family holdout 112.3; 1; 0 887.2; 7; 0 1,895.8; 13; 0
Fast beta 125.7; 1; 0 923.3; 12; 0 1,928.9; 14; 0
Half transfer 119.6; 1; 0 903.3; 8; 0 1,896.4; 13; 0
Human price down 25.5; 1; 0 482.2; 8; 0 1,069.6; 13; 0
Low reuse 117.6; 1; 0 895.9; 8; 0 1,907.4; 13; 0
No composition 119.6; 1; 0 903.3; 8; 0 1,896.4; 13; 0
No demand expansion 119.6; 1; 0 903.3; 8; 0 1,896.4; 13; 0
No quality 119.6; 1; 0 903.3; 8; 0 1,896.4; 13; 0
No speed 119.6; 1; 0 903.3; 8; 0 1,896.4; 13; 0
Payment delay 120.9; 1; 0 902.9; 8; 0 1,907.8; 14; 0
Retention triple 119.6; 1; 0 903.3; 8; 0 1,896.4; 13; 0
Short life 119.6; 1; 0 903.3; 8; 0 1,896.4; 13; 0
Zero transfer 119.6; 1; 0 903.3; 8; 0 1,896.4; 13; 0
The semicolon-separated cell entries use the same 16-world denominator. For example, cautious
quarter-demand has negative mean profit, one qualified world and ten loss worlds. Every individual
median, quantile, cash/closure count, paired contrast and complete history is available in the evidence
package; this table is a compact publication view.
B Availability and exact identities
The deliverable contains the executable simulator, pinned native dependency, tests, source-linked
calibration and assumption register, frozen selection, all compressed ledgers, audit receipts, numerical
witness, English/French HTML, generated figures, LaTeX sources and three compiled PDFs. The
complete ledgers are distributed separately from the compact source/evidence companion and restore
under results/ . No production integration or publication deployment is necessary to replay the study.
The outer SHA256SUMS.txt binds the final compiled PDFs and archives without a circular self-hash. The
evidence companion binds scientific source and raw evidence in detail. A PDF cannot meaningfully
contain its own final SHA-256; its exact compiled identity is therefore recorded externally. The archive
includes the original compiled bytes alongside the LaTeX build configuration. Source facts remain
linked to their dated public documents, while simulation assumptions remain explicitly labeled.
When Paid Work Teaches the Network 20

AGI ALPHA FLYWHEEL Research edition 1.0.0
References
[1] MONTREAL.AI. T ransformational Intelligence: Learning to Change the Game . Supplied public foundation;
inspected Alpha commit 51d5ecb3493b01f718713f6873f2aa4e44fe3cce . https://github.com/Montrea
lAI/AGI-Alpha-Agent-v0/blob/51d5ecb3493b01f718713f6873f2aa4e44fe3cce/docs/Transformatio
nal_Intelligence.pdf .
[2] MONTREAL.AI. Autonomous Enterprises: SUCCESSOR Omega ; SUCCESSOR Architecture. Same inspected
Alpha snapshot. https://github.com/MontrealAI/AGI-Alpha-Agent-v0/blob/51d5ecb3493b01f7187
13f6873f2aa4e44fe3cce/docs/Autonomous_Enterprises_SUCCESSOR_Omega.pdf .
[3] MONTREAL.AI. PROOF LOOPS , research paper v2, author edition. GoalOS commit
1e9a5675b1662e98074fb1c71cf2c5b94c7d259f . https://github.com/MontrealAI/goalos- agi
alpha-sovereign-machine-economy/blob/1e9a5675b1662e98074fb1c71cf2c5b94c7d259f/docs/pape
r/PROOF_LOOPS_research_paper_v2_author_edition.pdf .
[4] MONTREAL.AI. Merkle Roots for the Validated Skill Graph , V114. Same inspected GoalOS snapshot. https:
//github.com/MontrealAI/goalos-agialpha-sovereign-machine-economy/blob/1e9a5675b1662e98
074fb1c71cf2c5b94c7d259f/docs/paper/Merkle_Roots_for_the_Validated_Skill_Graph_v114.pdf .
[5] OpenAI. GPT-6 Astra model documentation. Evidence date 26 October 2026. https://developers.openai.
com/api/docs/models/gpt-6-astra .
[6] OpenAI. GPT-6.1 T erra model documentation. Evidence date 26 October 2026. https://developers.openai.
com/api/docs/models/gpt-6.1-terra .
[7] OpenAI. Reasoning models and modes . Evidence date 26 October 2026. https://developers.openai.com/
api/docs/guides/reasoning .
[8] OpenAI. API pricing. Evidence date 26 October 2026. https://developers.openai.com/api/docs/prici
ng .
[9] WCTS / U.S. General Services Administration. Multiple Award Schedule 47QRAA19D00CM, rates effective
18 September 2026–17 September 2027. Pages 6–7 and 27. https://www.gsaadvantage.gov/ref_text/47
QRAA19D00CM/0ZNBIJ.3VDOD7_47QRAA19D00CM_MASWCTSOY1.PDF .
[10] HiGHS project. HiGHS optimization software. Primary project documentation. https://highs.dev/ .
[11] SCIP Optimization Suite. Exact solving mode. Primary project documentation; configuration and certificate
qualifications apply. https://www.scipopt.org/doc/html/EXACT.php .
[12] K. K. H. Cheung, A. Gleixner and D. E. Steffy. Verifying Integer Programming Results. arXiv:1611.08832, 2016,
revised 2019. https://arxiv.org/abs/1611.08832 .
[13] E. Kim, A. Garg, K. Peng and N. Garg. Correlated Errors in Large Language Models . ICML 2025;
arXiv:2506.07962. https://arxiv.org/abs/2506.07962 . Numerical benchmark rates are not imported
into this study.
[14] MONTREAL.AI. Research coordination and prior-study provenance, issue #4850; PRs #4950, #4951 and
#4952. Exact source descriptions, historical cohort qualifications and inspected commits are preserved in
evidence/source_audit.md . https://github.com/MontrealAI/AGI-Alpha-Agent-v0/issues/4850 .
When Paid Work Teaches the Network 21