Patterns that survive contact with contrast.

This is a working synthesis of the case-study corpus, not a list of timeless laws. The individual teaching notes preserve each case's evidence, rivals, counterfactuals, and boundary conditions. The generated rules/index.json controls structural rule status; substantive promotion still requires human review, independent cross-case support, a counterexample, a calibrated expected range, and acceptable false-positive costs.

1. Buy evidence before buying irreversibility

When uncertainty is high and commitment is divisible, release the smallest representative tranche, measure it, and precommit the next release to explicit economics, service, liquidity, governance, and customer-harm gates.

This pattern appears in successful transitions and platforms such as Netflix, Amazon, Microsoft, NVIDIA, TSMC, and MercadoLibre. In ASML, filed Intel NRE agreements separated EUV and 450mm and gave ASML project-specific abandonment and redirection rights; EUV later advanced while 450mm was paused when demand and timing remained unclear. The divergence is consistent with a workstream-level-options thesis, but the public record neither proves that ASML used Part A's exact gates nor identifies co-investment's causal contribution. Staging also frames the operating expansion choices in Toyota, Walmart, and Costco. Failures including WeWork, GE/Alstom, and AOL–Time Warner show why large fixed commitments, price, or integration can outrun evidence.

Agent action: separate option purchase, pilot, scale, and full integration. For each stage, state the cash at risk, evidence purchased, decision deadline, next gate, and automatic kill condition.

Boundary: staging is not a reason to wait through an imminent run, custody breach, safety hazard, or fiduciary violation. SVB, FTX, and Boeing illustrate cases where the protective action may need to precede additional learning. Nor does strategic-customer participation validate every technical roadmap: demand, readiness, net funding after concessions, and governance rights still need independent workstream gates.

2. Underwrite the smallest stable-scope decision unit

Consolidated growth can coexist with unattractive sites, cohorts, routes, contracts, products, merchants, or legal entities. The useful unit is the smallest one at which incremental revenue, fully loaded variable and step-fixed cost, working capital, service, risk, and capital can be observed without arbitrary allocation.

Examples include Walmart clusters and routes, Costco warehouses and member cohorts, MercadoLibre country-route cells, Netflix channel cohorts, WeWork locations and mature cohorts, GE contract cohorts, and Wirecard merchant-to-settlement traces.

Agent action: name the decision unit before calculating unit economics. Freeze its definition and compare matched cohorts. Reconcile the unit to consolidated statements, but do not infer unit profitability from aggregate revenue or EBITDA.

Boundary: a local unit can look attractive because centralized technology, brand, inventory, risk, support, or capital costs are omitted. Require both local contribution and a complete corporate-capacity bridge before full-scale approval.

3. Cash is a set of legal and operational rights, not a label

For material cash or liquidity, test four different propositions: existence, legal ownership or beneficial interest, restrictions and competing obligations, and practical access at the time and entity where it is needed. Then deduct haircuts, encumbrance, collateral, settlement time, customer funds, covenants, and stressed outflows.

Wirecard shows why gross consolidated cash and audit status cannot replace direct bank, trustee, ownership, restriction, and access evidence. FTX adds custody and legal-entity attribution. Lehman and SVB add time-to-cash, collateral, correlated outflows, and operational readiness. WorldCom adds the difference between headline capacity and signed drawable financing. WeWork adds the need to distinguish financing survival from business-model economics.

Agent action: build a security- and entity-level liquidity ledger. Count an asset as survival liquidity only after an independently controlled confirmation and a realistic same-day or time-bucket monetization test.

Boundary: high reported cash or committed facilities can be entirely genuine and still not be transferable, unrestricted, drawable, or available before the liability falls due.

4. Match commitment duration, funding duration, and verification speed

Long-lived, fixed, or hard-to-reverse obligations require funding and evidence that remain valid for at least the same horizon. Short customer commitments funding long leases, short confidence-sensitive deposits funding long securities, or near-term facilities supporting uncertain assets create an amplification channel.

WeWork, Lehman, SVB, and WorldCom expose different versions of duration and confidence mismatch. TSMC and Inditex show why even productive capacity should be modularized around demand, utilization, service cadence, and downside liquidity. Berkshire shows the value of reserve and liquidity gates around long-duration capital deployment. ASML adds a frontier-technology variant: an approved multiyear customer funding commitment does not eliminate technology, deployment-timing, or demand-duration risk, as the later 450mm pause illustrates.

Agent action: construct contractual and behavioral maturity ladders, include correlated runoff and margin or collateral calls, and measure the time required to verify and monetize assets. Stress trigger, amplification, and management response together.

Boundary: long-duration investment is not inherently bad. It becomes dangerous when the funding, liquidity buffer, or evidence horizon is shorter and the commitment cannot be resized. Customer funding can share duration risk, but only its net amount after discounts, support, IP, access, warranty, and termination concessions should count as risk transfer.

5. Preserve metric definition, perimeter, period, and reporting vintage

Many analytical errors are invalid joins. “Customer,” “member,” “user,” “transaction,” “volume,” “cash,” “net debt,” “organic growth,” “existing business,” and “same store” can change definition, geography, consolidation, channel, or fiscal period. Later comparatives, recasts, restatements, and court findings must not silently overwrite what was originally reported.

The rule is visible across Blockbuster, Danaher, GE, Inditex, TSMC, MercadoLibre, Costco, Walmart, AOL–Time Warner, Wirecard, and Wells Fargo. Wells Fargo adds a semantic-integrity test: a product-opening or cross-sell count cannot safely stand in for customer depth unless authorization, durable use, customer benefit, and net contribution are reconciled to the reported KPI.

Adobe adds a reporting-vintage control: its FY2013 filing reported 1.4 million paid Creative Cloud subscriptions and USD 911 million of Digital Media ARR, while its FY2014 filing displayed FY2013 comparators of 1.439 million and USD 944 million. Preserve both vintages rather than guessing whether rounding, scope, or another change explains the difference. Paid subscriptions, ARR, GAAP subscription revenue, segment revenue, deferred revenue, and operating cash flow remain distinct unless a deterministic bridge supports the join.

Best Buy and J.C. Penney add retail-calendar and sales-perimeter controls. Best Buy's original 53-week FY2012, its later 48-week FY2012 recast, the 48-week FY2013 transition period, later 52-week years, and 53-week FY2018 are not one silently continuous series; even FY2014 changed across filing vintages when businesses were presented as discontinued operations. J.C. Penney's 53-week FY2012 total sales must remain separate from the issuer's 52-week comparison, while total sales, comparable-store sales, and Internet sales measure different perimeters and Internet sales were included in comparable-store sales. Directional consistency does not make those measures interchangeable.

Agent action: attach source, public date, valid period, entity, unit, currency, definition, scope, and revision status to every material value. Join only compatible observations. Where no deterministic bridge exists, show separate directional tables or an unknown cell and abstain from a normalized trend. For compensated volume metrics, also stratify authorization, reversals, complaints, overrides, and durable use by product, manager, location, and incentive cohort before relying on the aggregate.

Boundary: preserving lineage proves what a source reported and when; it does not validate the source's accuracy. A rising KPI may be arithmetically correct yet economically misleading if the counted events were unauthorized, duplicative, inactive, or unprofitable.

6. Interpret audits, regulators, ratings, and markets within their actual scope

An audit opinion, regulatory action, investigation, rating, financing, or price is evidence about a bounded proposition. It is not transferable reassurance for every cash balance, subsidiary, transaction, product, safety system, or legal entity.

Wirecard's unqualified group opinions, BaFin trading action, issuer investigations, and market recovery coexisted with unresolved component, cash, and TPA verification. Enron and WorldCom show that governance and reporting signals require direct accounting and liquidity tests. Boeing shows that certification and program schedules do not replace an integrated system-safety case. FTX shows why brand, fundraising, and group-level resources do not prove custody or entity-level recoverability. In Theranos, one cleared assay, laboratory status, commercial scale claims, and a retail partnership could not establish performance across the full menu or its real-use workflow. The relevant perimeter was assay, specimen, collection method, device, site, operator, quality system, and intended use—not the company or platform label. Wells Fargo adds a procedural boundary: agency consent and supervisory records, company admissions, settled administrative findings, and later remediation determinations each support different propositions. The 2025 removal of the growth restriction and 2026 termination of the enforcement action do not retroactively validate the historical cross-sell metric.

Volkswagen adds a procedural-status ladder. The September 2015 EPA notice was an agency enforcement position at the cutoff, not an adjudication. A June 2016 filed proposed partial consent decree contained bounded 2.0-liter admissions; a separate EPA statement later reported court entry on October 25. The January 2017 signed plea supplied corporate admissions, while the April judgment established conviction, probation, and the criminal fine. EPA/CARB remedy approval covered specified model-year 2015 Generation 3 vehicles, not the global fleet. Later instruments strengthen the record within their own scope; none retroactively changes what was known or adjudicated at the Part A cutoff.

Agent action: for every assurance signal, record issuer, date, mandate, procedure, population, materiality, legal entity, exceptions, and explicit noncoverage. Obtain primary verification for the assertion actually driving the decision.

Boundary: the lesson is not to ignore audits, regulators, ratings, or prices. They can be high-value evidence; their weight falls when the decision claim extends beyond their scope. Likewise, later regulatory closure establishes the regulator's stated remediation status, not that earlier controls were adequate or that every economic cost has been measured.

7. A self-disruption strategy needs a funded customer and cash-flow bridge

The legacy engine often funds the successor while customers, distribution, content, developers, sales capacity, or devices migrate. Protecting the legacy profit pool too long can surrender the future; abandoning it too quickly can remove the cash and customer base needed to build the future.

Netflix and Blockbuster provide the clearest paired channel transition. Apple and Microsoft show focus, complementor support, and staged platform migration. Nokia shows that platform choice, transition design, product execution, carrier execution, and market shift must be underwritten separately. Kodak corrects the easy “failed to see digital” story: invention, digital products, disclosed substitution risk, and strategic intent can all exist without a successful successor economic model. Recognition, resource transfer, commercialization, customer migration, legacy runoff, and balance-sheet capacity are separate tests.

Adobe adds a recurring-revenue variant: early paid adoption and a cash-generative incumbent base supported continued investment, while unavailable mature-renewal, realized-pricing, customer-value, and fully loaded cohort data argued for separating product direction from withdrawal pace. Where feasible, gate irreversible product, channel, pricing, and support changes on mature-cohort retention, customer value, fully loaded contribution, and transition cash. Adobe's later outcome is compatible with this logic; the public record does not show that Adobe used these exact gates or that exclusivity was necessary.

Agent action: model legacy runoff, successor adoption, duplicated costs, cannibalization, service quality, switching incentives, and liquidity in one bridge. Release migration waves against paid retention, contribution, reliability, ecosystem, and cash gates.

Boundary: staged migration fails when the market's clock is faster than the internal gate cycle. Time-box the bridge and specify when delay itself becomes the larger risk.

8. Platform value comes from a reinforcing system, not scale alone

Durable platforms tend to reduce participant friction, preserve or increase complementor surplus, create reusable tools or infrastructure, and improve the next participant's experience. Distribution control or installed base can help, but excessive extraction, customer conflict, poor neutrality, or weak complementor economics can erode the system.

Apple, Google, NVIDIA, TSMC, Microsoft, Amazon, and MercadoLibre show distinct mechanisms: devices and complements, quality-aware auctions and distribution, developer tooling, neutral foundry services, cross-platform cloud migration, shared infrastructure, and commerce-payments-logistics integration. ASML adds customer co-investment design: foundation-held, ordinarily non-voting shares and filed Intel terms preserving ASML's development control and IP addressed some independence risks. The same record disclosed tool-purchase commitments and commercial credits, whose net value and complete cash-flow effects still require separate review. Nokia and AOL–Time Warner show why installed base, reach, or a strategic narrative is not sufficient without product execution and verified integration economics.

Agent action: map every participant, value contribution, switching cost, subsidy, take rate, conflict, and governance right. Measure retention and contribution for each side; do not assign all aggregate enterprise value to one complement without a causal and capitalization-complete bridge.

Boundary: network effects do not excuse monopoly, default, privacy, labor, credit, or customer-harm risks. Regulatory and distribution dependencies belong in the economics. Nor does customer co-investment prove neutral access, binding demand, or project return; those conclusions require the executed rights and a workstream-level cash bridge.

9. Operating systems are socio-technical systems

Tools such as pull systems, standardized work, rapid replenishment, limited assortment, continuous improvement, and acquisition playbooks work through people, incentives, training, information flow, supplier relationships, governance, and the authority to stop or correct work. Copying visible tools without those conditions can degrade quality or safety.

Toyota provides the clearest operating-system transplant case. Danaher adds acquisition selection and integration discipline. Walmart and Inditex connect logistics cadence to local density, supplier interfaces, labor, and service. Costco connects member surplus, employee capacity, assortment, and site economics. Boeing shows the cost of allowing schedule and organizational interfaces to outrun integrated safety analysis. Wells Fargo shows how a stated needs-based policy can be undermined when product-volume goals, managerial incentives, complaint escalation, and board information do not operate as one control system.

Volkswagen adds a crisis-control variant: formal governance disclosures coexisted with emissions misconduct and document destruction that occurred before the September 21 decision; the EPA's enforcement position was public at the cutoff, but the corporate admissions and criminal judgment came later. After the notice, the issuer reported Supervisory Board meetings, a special committee, and leadership changes; the signed plea recorded forensic document recovery, investigation, and several reforms while stating that remediation remained incomplete. Evaluate original controls, evidence preservation, and post-notice remediation separately. Personnel actions and new structures are inputs, not proof of operating effectiveness.

Agent action: define the required behaviors, decision rights, training, worker voice, supplier controls, information cadence, quality and safety gates, and local adaptation. Measure local mechanism outcomes before claiming enterprise-wide causality. Do not treat employee discipline as sufficient evidence of control effectiveness; reconcile the underlying incentive, complaint, override, remediation, and escalation mechanisms.

Boundary: local operational improvement can coexist with adverse enterprise shocks or weak capital allocation. Keep plant, site, program, and corporate outcomes separate. Punishment counts can reflect detection or a persistent architecture failure, and cannot distinguish the two without an incidence denominator and evidence that incentives and escalation changed.

10. In M&A, price, funding, perimeter, integration, and synergy are separate gates

An attractive target and a strong operating system do not make any price or structure acceptable. A transaction must separately clear standalone value, acquisition price, financing resilience, liabilities, control perimeter, integration capacity, customer and employee continuity, and verified net cash synergies.

Berkshire/GEICO illustrates independent price, reserve, liquidity, regulation, and management-retention gates. Danaher/Pall separates target quality and operating-system potential from price and later separation. GE/Alstom shows complexity, contract, funding, and integration risks around a transformational acquisition. AOL–Time Warner shows the importance of fixed-ratio currency, stable standalone baselines, staged commercial proof, and refusing to infer causal loss from an accounting write-down alone.

Priceline–Booking.com and eBay–Skype add a paired reading, not a controlled comparison. The Booking.com case is classified as a success only for Booking.com-led international operating-platform scaling inside Priceline through FY2011, consistent with selective integration. It does not isolate Booking.com from Agoda, rentalcars.com, foreign-exchange effects, or other perimeter changes; establish standalone acquisition ROI or IRR; prove that the purchase price was attractive; establish public-equity success; or show that selective integration caused the result.

The Skype case is classified as a failure only for eBay's commerce-and-payments strategic-fit thesis by the 2009 control sale. Skype's disclosed standalone operating path grew, and eBay called it a strong standalone business while saying it lacked synergies with ecommerce and payments. That record does not make Skype a failed product or business, and the impairment, sale proceeds, retained-interest fair value, and accounting gains do not form a complete cash-return bridge or support a total-return conclusion.

The candidate rule from the pair is to preserve the target's standalone operating core, install the minimum control and finance interfaces needed for safe ownership, and stage deeper integration behind two scorecards: target-health non-inferiority and causally attributable acquirer net contribution after direct integration cost. Predeclare pause and reversion gates; require the contribution estimate's lower confidence bound to remain above zero for two consecutive quarters; and re-underwrite ownership and integration after four quarterly reviews or 12–18 months if the thesis remains unverified. These cases do not prove that this rule caused Booking.com's outcome or that absent or weak gates caused Skype's.

Agent action: maintain distinct approval gates for bid, signing, financing, close, integration wave, and any separation. Reconcile every transaction-value definition and currency. Validate synergies through customer and cash evidence before irreversible integration. Keep target health and acquirer contribution separate so target growth cannot stand in for synergy.

Boundary: a successful integration does not prove the price was good; a later impairment does not by itself measure cash loss or prove the merger caused all value change. Staging can also delay a real opportunity, duplicate platforms, fragment accountability, weaken bargaining leverage, or defer urgent risk remediation. It is not a reason to postpone immediate integration of legal, financial-reporting, custody, security, safety, or other controls that must operate at close.

Gilead–Pharmasset and BMS–Inhibitex add a same-wave HCV nucleotide acquisition reading, not a controlled comparison or reconstruction of either board's original signing process. Both Part A packets begin after signed agreements and use the then-public record. They can test post-signing exposure control and re-underwriting, but cannot establish what either board actually knew, which alternatives were feasible before signing, or whether a contingent structure would have been accepted.

The Gilead success classification is limited to PSI-7977/sofosbuvir's later clinical validation, FDA approval, and disclosed product sales. It does not establish purchase-price attractiveness, acquisition IRR, public-equity success, ethics, sole causality, or that all later Harvoni economics belong to Pharmasset. The BMS failure classification is limited to validation of INX-189/BMS-986094 after the clinical hold, discontinuation, and reported USD 1.83 billion pre-tax impairment. BMS's later safety stop was an appropriate risk response; the impairment is not a complete acquisition cash-return bridge.

All four rule cards introduced by these two cases remain candidate. Their contrasting outcomes do not prove that Gilead used the proposed gates, BMS omitted them, or either rule transfers to a broader acquisition sample.

11. Separate accounting identity, economic value, and causal attribution

Accounting amounts can record real impairment, provisions, expenses, or balance-sheet changes without identifying the complete economic loss or its cause. Conversely, an operating metric can grow while value creation is negative after capital, subsidies, risk, or later choices.

AOL–Time Warner, Boeing, MercadoLibre, Apple, Toyota, and GE repeatedly require this separation. Kodak's reported digital-revenue crossover likewise did not establish value crossover: recurring product cash, segment profit, acquisition capital, restructuring, and legacy obligations still required a complete bridge. The same discipline applies to Wirecard's distinct KPMG evidence-status amounts, June trust-cash amount, and later court overvaluation scopes. ASML's approved EUR 1.38 billion customer R&D commitment, EUR 3,853.9 million of executed share proceeds, EUR 3,728.3 million synthetic-buyback repayment, EUR 64.4 million of 2013 CCIP contribution other income, EUR 882.0 million of consolidated R&D costs on a different basis, later consolidated results, and EUV system volumes are all different constructs; none alone measures the program's return or causal contribution. Wells Fargo's employee terminations, penalties, remediation duration, later earnings, customer harm, and counterfactual franchise value likewise cannot be summed into one loss or success measure.

Adobe adds a transition-timing case: subscription revenue rose in FY2013 while total revenue, net income, and operating cash flow fell below FY2012, consistent with the issuer's disclosed shift from upfront perpetual payments to ratable subscriptions. That sequence does not by itself establish poor economics; the later recovery and much larger consolidated subscription-weighted base likewise do not measure the causal return on the May 2013 decision.

Volkswagen adds a non-additivity control for legal crises. The issuer's EUR 16.2 billion of 2015 diesel-related expense and EUR 6.4 billion of additional 2016 expense contain provisions, additions, impairments, reversals, and currency effects; the separate USD consumer-program estimate, zero-emission-vehicle commitment, mitigation trust, and criminal fine are different constructs. The agreed and later imposed USD 2.8 billion criminal fine is one obligation at successive procedural stages, not two costs. Add only within an instrument's explicit bridge; do not combine legal amounts with accounting expenses or set them against Automotive Division net liquidity until currency, perimeter, status, period, cash timing, recoveries, and overlap are reconciled.

Best Buy and J.C. Penney add the same discipline to turnaround programs. An issuer's annualized cost-elimination claim is not cumulative cash savings, restructuring cash, free cash flow, or enterprise value. Store investment, operating cash flow, inventory, liquidity, and claimed savings remain separate lines until a deterministic bridge proves their relationship. Checkpoint claims at different dates also must not be added together when the later amount may include the earlier work.

Agent action: first reconcile the accounting bridge. Then build the cash and capitalization bridge. Only then test causal hypotheses against rivals and confounders. If the bridge is incomplete, state the observed amount and abstain from an all-value or percentage attribution.

Boundary: abstention is not a claim that causality is unknowable forever. It is the correct current state until decision-safe evidence and a reproducible model exist. Later scale or earnings are opposing evidence against a permanent-franchise-failure story, not proof that an earlier program caused the growth or that misconduct imposed no cost.

12. Model failure and success as layered mechanisms, not single events

Sequence is not causality. A useful explanation separates structural vulnerability, trigger, amplification, management or policy response, later choices, external shocks, and measurement changes. It also distinguishes process quality from realized outcome.

SVB explicitly separates deposit structure, trigger, run amplification, operational liquidity, and response. Lehman separates asset marks, funding, reporting, capital, and confidence. Wirecard separates underlying authenticity problems, verification failure, institutional delay, funding acceleration, and legal endpoints. Boeing separates design, system safety, certification, training, production, and response. Nokia, MercadoLibre, and AOL–Time Warner show why long-horizon outcomes accumulate many later choices and external forces. Wells Fargo separates KPI design, incentives, complaint escalation, metric contamination, firm-wide governance, and later remediation. ASML separates approved terms, executed equity and distribution steps, NRE commitments, recognized contribution income, commercial concessions, divergent workstreams, later execution, and consolidated company outcomes.

Adobe separates the May 2013 product-policy choice and short-run CS6 bridge from early adoption, near-term transition compression, later execution and portfolio change, changing metric scope, and the FY2025 endpoint. This sequence supports plausible contribution, not exclusivity, necessity, an effect size, or proof that management used the proposed gate process.

Volkswagen separates (i) pre-cutoff road-versus-lab anomalies, regulatory escalation, misconduct, and document destruction—some established only in later records; (ii) the September 18 notice and the partial halt and external inquiry already known at cutoff; and (iii) post-cutoff board action, expanded global and V6 scope, settlements, plea and judgment, remediation, and financial outcomes. A later approved remedy for specified Generation 3 vehicles, continued operating scale, or higher Automotive Division net liquidity is evidence about bounded response or resilience, not retroactive validation of earlier controls or proof that every configuration, customer harm, or economic cost was resolved.

Agent action: state a primary mechanism chain, at least one credible rival, opposing evidence, confidence, a falsifier, and a feasible-at-the-time counterfactual. Grade the decision before reading the outcome; grade calibration and signal recognition only after the reveal.

Boundary: a good process can lose and a bad process can win. Never promote a rule merely because one recommendation matched one realized outcome. The Wells Fargo record does not show how much a 2013 control gate would have avoided, and the ASML sequence does not show that co-investment caused EUV progress or that paused 450mm work had no residual value.

13. Turnaround sequencing must preserve the customer-response bridge

A turnaround can require simultaneous work on price, assortment, service, digital capability, stores, cost, and liquidity, but those workstreams do not all earn the same release schedule. When a legacy promotion, product, channel, or service pattern still anchors customer comprehension and cash generation, its withdrawal should be gated on evidence that the replacement works for representative customers.

The Best Buy Renew Blue and J.C. Penney Fair and Square cases form an instructive matched reading, not a controlled experiment. Both began with operating pressure, new leadership, and multi-workstream transformation claims. Best Buy's later record is consistent with a qualified operating recovery alongside customer-value, fulfillment, footprint, and cost work, but public evidence does not prove the causal contribution of the sequence or the use of Part A's proposed gates; product and category cycles, competition, portfolio and reporting-perimeter changes, and later execution remain live Best Buy rivals. J.C. Penney broadly reset pricing and promotions while its shop conversion remained phased; later filings documented severe traffic, sales, margin, cash, and liquidity deterioration. The outlet and catalog exit, merchandise and markdown transition, traditional promotional competition, inventory alignment, jcp.com in-stock and site performance, leadership change, financing constraints, later reversals, and general execution remain live J.C. Penney rivals, so the pair cannot identify a single treatment effect.

Agent action: before withdrawing a legacy customer bridge, define representative cohorts with frozen membership and maturation windows. Precommit thresholds, review cadence, and release, pause, and reversal decisions to customer comprehension, conversion, gross-profit dollars after concessions, fully loaded channel contribution, repeat purchase, service and availability, working capital, and liquidity. Keep cost claims in a separate ledger; only controller-reconciled realized net cash benefit may fund verified customer value and operating capacity.

Boundary: the rule does not require preserving a deceptive, unsafe, illegal, or structurally uneconomic legacy practice. Nor does it justify endless piloting when the market clock is faster than the learning cycle. In those settings, redesign or withdraw the harmful element promptly, cap exposure, protect liquidity, and test the safest viable replacement on the shortest decision-relevant cadence. If national messaging, cross-shopping, or operational leakage prevents representative controls, or rollback cannot be demonstrated before launch, do not infer that a nominal pilot has purchased decision-grade evidence: pause the broad reset or cap exposure through time, category, or geography and tighten the liquidity limit.

14. Pivotal validation and at-risk scale require independent gates

An emergency market clock can make waiting for complete technical validation costly, but urgency does not turn early safety, immunogenicity, funding, procurement, or manufacturing readiness into proof of clinical efficacy. Those are separate evidence streams. At-risk scale can be rational before a pivotal readout only when the decision states which risk is being purchased, caps irreversible exposure, preserves pause or redeployment rights, and does not disguise public risk-sharing or conditional demand as product validation.

The Moderna mRNA-1273 and CureVac CVnCoV cases form a useful matched reading, not a controlled experiment. At their respective cutoffs, each program had early clinical evidence, a planned large pivotal trial, public financial support or conditional procurement, and manufacturing-scale claims; neither had pivotal efficacy evidence. Their later first-generation outcomes diverged. That divergence does not identify one causal treatment: trial timing, geography, circulating variants, endpoint definitions, analysis populations, dose and formulation, assay methods, manufacturing execution, public support, procurement structure, and later management choices all differ.

Agent action: maintain separate ledgers for (i) safety and immunogenicity, (ii) the frozen pivotal protocol, endpoints, populations, and statistical plan, (iii) CMC and manufacturing readiness, (iv) grants and other risk transfer, (v) conditional and binding procurement, and (vi) irreversible net exposure after termination, redeployment, and recovery rights. Precommit release, pause, and kill decisions to the evidence each ledger can actually establish. Track protocol versions and variant or other external-regime changes without silently changing the endpoint or comparison set.

Boundary: a successful pivotal result does not prove that every at-risk commitment was necessary or optimally sized, and a failed first-generation program does not prove that a bounded, reversible emergency investment was irrational ex ante. Do not compare headline efficacy estimates across programs as if trials, populations, estimands, and circulating pathogens were standardized. When capitalization, program-level cash flows, contractual recoveries, material conflicts, or valuation inputs are incomplete, report the observed amounts separately and abstain from a program return or target price.

The Gilead–Pharmasset and BMS–Inhibitex pair extends this validation ladder from internally developed programs to acquired development assets. A regimen-level antiviral signal and short-exposure tolerability are evidence, but they do not establish longer-duration organ or cardiac safety, resistance and relapse durability, pivotal efficacy, regulatory approval, manufacturing readiness, or commercial value. The later divergence does not isolate one causal treatment because the molecules, doses, exposure durations, cohorts, development stages, protocols, portfolio options, and organizational execution differed.

Agent action: for acquired programs, maintain separate ledgers for transaction consideration, molecule- and regimen-level validation, longer-duration safety, pivotal and regulatory evidence, CMC readiness, incremental program cash, and portfolio or combination attribution. Affordability is not price attractiveness, asset validation is not acquisition return, and a prudent safety stop is not itself the failed decision.

15. Capital return is a recurring release decision, not a standing entitlement

An authorization establishes legal capacity for a capital action; it does not make the action obligatory, economically attractive, funded, or already executed. Treat every material repurchase tranche as a fresh capital-release decision. Before release, reconcile reported operating cash flow after total capex, working-capital and vendor needs, leases and other fixed charges, conditional borrowing capacity, maturities and refinancing, existing settlement commitments, gross purchases versus net share movement, and an independently reproducible price and capitalization record. Operating cash flow less total capex is only a descriptive residual when maintenance and growth capex are not separated; it is not silently renamed free cash flow.

AutoZone and Bed Bath & Beyond form a matched reading about release mechanics, not a controlled comparison of companies or endpoints. AutoZone's success classification is bounded to sustained operating activity, reinvestment, refinancing, and net share reduction through FY2024. It does not establish repurchase alpha, intrinsic-value discipline, total shareholder return, or the program's causal contribution. Bed Bath & Beyond's failure classification is bounded to the choice to accelerate a remaining USD 400 million release while the residual cash and vendor record was weakening. Later cash burn, borrowing, dilution, Chapter 11, and equity cancellation do not establish that repurchases alone caused the bankruptcy; merchandising, supply-chain, inflation, demand, inventory, vendor, financing, and later-management mechanisms remain live.

Agent action: maintain separate ledgers for authorization, open commitments, cash settlement, employee-withholding purchases, issuance and dilution, and net basic and diluted shares. Precommit automatic pause gates for negative or deteriorating period-matched residual cash after required reinvestment, inadequate fixed-charge or refinancing buffers, adverse vendor or collateral signals, and missing price support. Crossing a numerical gate should trigger a documented human re-underwriting decision; it must not authorize an external capital action automatically.

Boundary: returning genuinely excess capital can be rational, and debt growth, negative book equity, or a short-period cash deficit does not by itself establish distress or a bad repurchase. Conversely, a lower share count or higher per-share result does not by itself establish operating improvement or value creation. If complete capitalization, maintenance-capex, contractual, conflict, forecast, or valuation inputs are missing, report the observed accounting and cash movements and abstain from repurchase IRR, buyback alpha, target price, and position sizing.

16. Cyclical capacity is a funded option, not a utilization forecast

Nominal capacity, production-to-capability ratios, and issuer forecasts do not establish profitable utilization or financeability. A downturn can make capacity or acquisitions attractive, but only if period-matched cash generation and usable liquidity can fund the committed path through a downside case without relying on unsigned facilities, expected waivers, forced asset sales, or unscheduled benefit deferrals. Keep capacity, output, shipments, utilization, margin, and plant-level contribution distinct.

Nucor and Bethlehem Steel share the same June 20, 2001 knowledge cutoff and an industry downturn, but they are not a controlled experiment. Nucor's success classification is bounded to remaining profitable and liquid while funding capacity through FY2004. Bethlehem's failure classification is bounded to the legal entity, the filed plan's old-equity and unsecured-creditor treatment, and separately defined pension outcomes through the sale and disclosure process—not a later final distribution ledger. Product mix, production technology, asset age, customers and geography, labor and benefit structures, raw-material exposure, acquisitions, trade conditions, the 2002 steel tariff, the 2004 price cycle, and later management choices all limit causal transfer from the pair. Production-technology and fixed-cost differences are unquantified hypotheses in this packet, not a sourced causal decomposition.

Agent action: maintain separate ledgers for (i) period-matched operating cash flow, (ii) total capital expenditure with maintenance and growth left unknown when not disclosed, (iii) acquisitions and other committed cash uses, (iv) cash and short-term investments, unused facilities, covenant headroom, borrowing-base eligibility, and legal-entity access, (v) funded debt and leases, (vi) pension and OPEB accounting measures and their actual cash schedules, and (vii) plant- and product-level contribution. Operating cash flow less total capital expenditure is a descriptive residual, not free cash flow when maintenance capital is unknown. Require executed financing documents and precommit release, pause, reduction, and exit gates; a breach should trigger documented human re-underwriting, not an automatic external action.

Boundary: Nucor's bounded outcome cannot prove that every capacity release or acquisition earned an attractive return, that its shares were attractively priced, or that one operating model caused resilience. Bethlehem's bounded outcome does not imply that its operating assets disappeared or that labor, imports, retirement obligations, trade policy, or any other single mechanism caused Chapter 11. Bankruptcy can preserve operating assets while the legal entity and common equity fail. When complete capitalization, maintenance capital expenditure, contractual cash schedules, material conflicts, forecasts, or valuation inputs are missing, abstain from target price and position sizing.

17. LTL density is a service-and-liquidity system, not a terminal count

A less-than-truckload network does not become economically dense merely because it owns many terminals, combines brands, or reports a larger shipment base. Decision-grade density evidence must connect service reliability and direct routing to customer retention, price realization, shipment contribution, operating-ratio improvement, and cash-funded capacity. Terminal count, tonnage, yield, and operating ratio remain separate measures until their scopes and periods are reconciled.

Old Dominion and YRC/Yellow share a March 16, 2010 knowledge cutoff and an LTL decision frame, but they are not a controlled experiment. Old Dominion's success classification is bounded to later service, network, operating, profitability, and financeability evidence through FY2022. It does not establish total shareholder return, valuation, project IRR, or that network density alone caused the outcome. YRC/Yellow's failure classification is bounded to the later operational wind-down, network cessation, and Chapter 11 filing. It does not establish final common-equity recovery, final creditor distributions, disappearance of the operating assets, or a single causal verdict about labor, acquisitions, integration, financing, policy, or management. Customer mix, geography, labor structure, acquisitions, capital structure, financing access, reporting-perimeter changes, external shocks, and later choices all limit causal transfer. In particular, Yellow's later consolidated One Yellow measures cannot be treated as a continuous series with the National and Regional segment measures reported at the cutoff.

Agent action: maintain matched ledgers for service failures and claims, shipment and weight mix, price and fuel effects, direct versus handled moves when disclosed, shipment contribution, operating ratio, operating cash flow, total capital expenditure, cash, executed and conditional borrowing capacity, debt and leases, pension and other fixed obligations, collateral, maturities, and legal-entity access. Reconcile the chain from service to retention, pricing, contribution, and cash before releasing network capital. Precommit release, pause, reduction, and exit gates; crossing one requires documented human re-underwriting and cannot itself authorize an external capital action.

Boundary: gross facilities are not usable liquidity, a consolidation plan is not realized integration, and a better reported operating ratio does not by itself prove economic density or adequate financeability. Differences in metric scope, fuel treatment, reclassifications, acquisitions, and corporate perimeter can defeat apparently simple time-series or cross-company comparisons. When complete capitalization, maintenance capital expenditure, contractual cash schedules, material conflicts, forecasts, or valuation inputs are missing, report the observed operating and financing evidence and abstain from target price and position sizing.

18. Semiconductor roadmaps are evidence chains, not moat claims

A process or product roadmap is not validated by an architecture name, node label, capital-spending plan, manufacturing model, or management assertion. Underwrite the chain from technical validation through roadmap milestones and customer adoption to stable segment economics, manufacturing and capital requirements, funding resilience, and valuation. Breaks at different links imply different decisions; consolidated growth or positive cash generation cannot identify which link is working.

AMD and Intel share the same evidence cutoff, but they are independent underwrites rather than a controlled horse race. AMD's turnaround classification is bounded to the operating, roadmap, segment, and financeability evidence observed through FY2025. Intel's failure classification is bounded to the roadmap, manufacturing, segment, capital, and funding evidence observed through FY2025. Neither classification establishes total shareholder return, security valuation, or a terminal verdict on either company.

Causal transfer from the pair is limited by fabless-versus-integrated manufacturing, starting scale, acquisitions and divestitures, segment definitions and mixes, leadership and capital-policy changes, competitors and customer exposures, and non-comparable process node labels. Cloud and AI demand, crypto cycles, the pandemic, supply constraints, and other external regimes also changed during the horizon. Preserve these differences rather than treating one company's later path as the counterfactual for the other.

Agent action: maintain separate evidence ledgers and precommitted gates for technical validation, roadmap milestones, customer adoption, stable segment economics, manufacturing and capital, funding, and valuation. Treat issuer roadmap, capacity, and capital-expenditure statements as claims until the relevant milestone and economic record verify them. Operating cash flow minus gross cash property-and-equipment additions is only a descriptive scale check when maintenance, growth, working-capital, acquisition, and restructuring perimeters are unresolved; do not call it free cash flow. Crossing a gate must trigger documented human re-underwriting, not autonomous trading or capital release.

Boundary: the pair does not prove that Zen alone caused AMD's turnaround, that either fabless or integrated manufacturing is universally superior, or that roadmap execution alone explains the divergent classifications. Do not infer total shareholder return, historical or current valuation, target price, position size, permanent Intel impairment, or a terminal competitive outcome without the separate point-in-time evidence those claims require.

19. Serial acquisition is a product-cash-and-governance system, not an adjusted-earnings rollup

A serial acquirer is not validated by deal count, proprietary-product mix, reported growth, an issuer-defined organic-growth label, adjusted earnings, or the ability to refinance. Underwrite the full chain from product necessity and customer alternatives through price, units, mix, access, retention, required reinvestment, acquisition-cohort cash returns, GAAP reconciliation, and instrument-level funding resilience. Each link can fail independently, and consolidated results cannot identify a cohort's return.

TransDigm and Valeant share a February 25, 2015 knowledge cutoff and both announced material signed acquisitions on February 20, 2015. They are independent underwrites, not a controlled horse race. TransDigm's success class is bounded to later operating, acquisition, and funding capacity through FY2025. It does not establish acquisition IRR, valuation, total shareholder return, customer surplus, or ethical innocence. Valeant's failure class is bounded to the 2015 acquisition-pricing-accounting-leverage thesis. It does not establish insolvency, failure of every product, a final verdict on every later business, or a single causal explanation.

Causal transfer is limited by aerospace versus pharmaceutical product economics, government and commercial customers versus patients, prescribers and payers, contract and regulatory regimes, starting leverage, acquisition histories, reporting perimeters, customer switching and lifecycle dynamics, and later management choices. Industry demand, interest rates, divestitures, separations, accounting changes, and other external regimes also changed during the horizon. Preserve those differences instead of treating either company's later path as the counterfactual for the other.

Agent action: maintain separate, versioned ledgers for (i) product-level price, units, mix, access, retention, alternatives, lifecycle and customer outcomes; (ii) acquisition price, financing, purchase accounting, integration, required R&D and commercial investment, working capital, tax and realized after-tax cohort cash; (iii) GAAP line items, every adjusted-measure bridge and earnings-to-cash reconciliation; (iv) debt instruments, accessible cash, maturities, rates, collateral, covenants, refinancing, distributions and stress liquidity; and (v) complaints, quality, regulatory findings, customer harm, conflicts and board remediation. Precommit release, pause, reduction and exit gates. A breach must trigger documented human re-underwriting and cannot itself authorize a trade, distribution, acquisition, financing, or other external action.

Boundary: issuer-defined organic sales are supplemental classifications, not standardized organic growth or causal attribution. Operating cash flow less total capital expenditure is only a descriptive residual when maintenance, growth, acquisition, integration and working-capital needs are unresolved; do not call it free cash flow. Debt carrying values and consolidation perimeters can change across filings, so apparent debt reduction or growth must not be treated as like-for-like operating evidence without a reconciled instrument and legal-entity schedule. When complete capitalization, attributable cash flows, material conflicts, contractual obligations, forecasts, or valuation inputs are missing, report the observed operating and financing evidence and abstain from target price and position sizing.

20. Acquisition underwriting must price standalone truth before synergy

A strategic acquisition is not validated by an accretion forecast, a large addressable market, reported post-close revenue growth, continued goodwill, or management's list of product integrations. Underwrite the chain from the target's versioned standalone financials, revenue quality, customer economics and cash conversion through price, premium, financing, purchase accounting, operating autonomy, integration interfaces, regulatory constraints, and attributable post-close cash returns. A break at any link cannot be repaired by assigning unmeasured synergy to the residual purchase price.

Microsoft–LinkedIn, HP–Autonomy, Disney–Pixar, and Daimler–Chrysler use different historical cutoffs and are independent underwrites, not a controlled comparison. Microsoft's success class is bounded to closing, continued operating identity, and later disclosed revenue and member scale through FY2026. It does not establish LinkedIn standalone profit or cash flow, purchase-price IRR, valuation, total shareholder return, or that Microsoft caused all later growth. HP's failure class is bounded to the acquisition's later impairment and the carefully separated accounting and legal record. It does not establish that fraud caused the entire impairment, that every accused person was guilty, that the impairment equals cash loss or damages, or that one cause explains the outcome.

Disney–Pixar's success class is bounded to completion, continued Pixar identity, and later disclosed Pixar-labeled film and franchise scale. Disney's broad segment results, multi-title quarterly income, consumer box-office gross, and proprietary Toy Story title return are not standalone Pixar cash flow or acquisition IRR; the title set includes pre-acquisition films. Daimler–Chrysler's failure class is bounded to severe observed Chrysler operating deterioration, restructuring, transfer of control within nine years, the separately reported deconsolidation loss, and retained exposure. The announced USD 92 billion combined-company value is neither consideration nor value destroyed, and the court's failure-to-prove disposition on a specific trial record is not a universal finding that every merger representation was true or false.

Causal transfer is limited by transaction timing, target maturity, business model, product and customer mix, purchase-price perimeter, acquirer funding capacity, management stability, integration design, regulatory conditions, market growth, and later product investment. Purchase accounting and pooling-of-interests are different reporting perimeters; absence of a purchase-price allocation under pooling does not mean zero economic cost. Economic ownership is not governance parity. Announced and closing exchange ratios, contributions, seller withdrawals, intercompany repayments, aggregate cash-flow lines, accounting losses, retained interests, funded loans, commitments, and guarantees must remain distinct. DM, USD, and EUR figures stay in their native currencies unless a dated sourced exchange rate supports conversion. The HP and Daimler records also demonstrate why issuer allegations, settlement orders, civil judgments, criminal dispositions, and operating outcomes remain separate evidence classes with their own parties, burdens, scopes, and findings.

Agent action: freeze the target's standalone fact base before assigning synergy; retain reported and later-corrected versions; map every material representation to independent diligence evidence; and maintain distinct ledgers for price and premium, financing, purchase accounting, customer and product cohorts, integration interfaces, regulatory commitments, attributable profit and cash, impairments, litigation, and remediation. Precommit approval, pause, renegotiation, integration, and exit gates. A failed gate must trigger documented human re-underwriting and cannot itself authorize an acquisition, security trade, accounting adjustment, legal conclusion, or publication.

Boundary: cumulative revenue is not a return, goodwill is not proof of value, absence of an impairment is not proof of success, and a non-cash impairment is not itself a complete cash-loss estimate. Court-assessed damages and accounting impairment measure different objects and cannot be netted without a governed bridge. When standalone profit, cash flows, complete capitalization, integration costs, material conflicts, forecasts, or valuation inputs are missing, report the observed operating record and abstain from acquisition IRR, target price, and position sizing.

The meta-lesson: uncertainty must change the action

An uncertainty section is not useful if the recommended commitment is identical whether the missing evidence arrives or not. Convert uncertainty into one of five controls:

1. obtain a specific primary record before committing;
2. reduce or cap exposure while the record is missing;
3. stage an irreversible commitment and buy evidence with the first tranche;
4. set a dated reversal or kill trigger; or
5. abstain from valuation or action when the decision cannot be made safely.

Use AGENT_EVALUATION_PROTOCOL.md to apply and score these principles without hindsight. Use rules/index.json to inspect the underlying case-rule cards and their structural promotion gaps.