Company Classification Routing

Company classification and routing

Standard valuation machinery assumes a lot. It assumes the firm survives long enough to reach stable growth. It assumes debt is a financing choice rather than raw material. It assumes this year's earnings say something about a normal year, and that there are earnings at all. It assumes the owner is diversified and the shares trade.

Point that machinery at a bank, a pre-revenue startup, or a firm at a cycle trough and it still returns a number. The number is confident and wrong. That is the most expensive failure mode in this domain, because nothing downstream flags it: the arithmetic is correct, the spreadsheet balances, and the answer is nonsense.

Routing comes first for that reason. This skill converts what is knowable about a company into a route — one primary engine, a set of overlays, an ordered pipeline, and a list of methods that must not be attempted. Gate G2_classified passes only when that route is written and valid. Every stage after it reads the route and honors it.

Full framework: the special-situations routing framework bundled in valuation-playbooks; load it with skill_view("valuation-playbooks", file_path="references/special-situations-routing.md"). This skill is its operating manual, and carries the parts an agent needs in hand.

When to Use

• Before any valuation begins, to fix the route every later stage must honour and write classification.json.
• When deciding which valuation model applies: FCFF, FCFE, dividends, excess return or a revenue-driven build.
• When a company looks non-standard: a bank or insurer, a pre-revenue or loss-making firm, a distressed or declining business.
• When the subject is a private company or an IPO, a cyclical or commodity producer, an emerging-market firm, or a multi-business group.
• When compiling the hard constraints a company's type imposes, such as no-fcff-valuation or require-failure-probability.

Contents

• What this produces
• The four questions
• S0 · Evidence intake
• S1 · Signal extraction
• S2 · Sector gate
• S3 · Statement-repair gate
• S4 · Ownership and transaction gate
• S5 · Life-cycle and earnings gate
• S6 · Survival and truncation gate
• S7 · Overlays and constraint compilation
• S8 · Combination rules
• The constraint catalogue
• Writing classification.json
• Confidence and what stays unresolved

References

• Branch catalogue B1–B16 — trigger, what breaks, what replaces it, and the hard constraints for each of the sixteen branches.
• Signal extraction detail — how each S1 signal is computed from filings and market data, and the judgment each one hides.
• Combination and worked routings — precedence, exclusion pairs, composition arithmetic, and routed real companies.
• Pricing routes by branch — which multiple survives this company's defects, and which are forbidden.
• The classification artifact — field-by-field shape, validation, and the diagnosis.md companion.

What this produces

Two files, both written by company-diagnostician and by nobody else:

Downstream agents read classification.json and never re-derive it. The critic checks every produced artifact against the compiled constraints, under rule V9, route conformance. valuation-consistency-checks enforces a subset mechanically. Pass the artifact to it with --classification. It then fails the run on an FCFF valuation of a bank, a missing failure probability, or an earnings multiple on negative earnings.

The four questions

Every valuation answers four questions, and every branch below repairs one of them.

1. What are the cash flows from existing assets?
2. What value is added by growth assets?
3. How risky are those cash flows?
4. When does maturity arrive, and what can end the story before then?

A company is difficult exactly when one of those answers is missing, unstable, or mis-measured by accounting. Diagnose which one breaks and the repair names itself. A young firm breaks Q1 and Q4. A bank breaks Q1 and Q3 together, because its debt is inventory. An intangible-heavy firm breaks Q1 and Q2, because accounting expensed its capital spending. Detail: skill_view("valuation-playbooks", file_path="references/difficult-company-taxonomy.md").

Two errors follow from skipping this step. The first is repairing a cash-flow problem in the discount rate — failure risk, governance, distress and country risk all get pushed there, and all belong in expected cash flows or in probability weights. The second is treating one bucket at a time when a company sits in several. Boeing in March 2020 was mature, cyclical and distressed at once.

S0 · Evidence intake

Inputs: mandate.json, 01-data/raw-financials.json, 01-data/market-data.json, 01-data/gaps.json.

A normal valuation leans on three information sources. Record each as present, thin or absent:

1. Current financial statements.
2. The firm's own financial history.
3. Industry and comparable-firm data.

Three decision rules follow.

• Two or more sources absent — this is the point of maximum temptation toward the dark side, where analysts declare a paradigm shift and invent metrics. Force an explicit mature end-state instead: target margin, terminal ROC, stable growth. Then work backwards. Set route.requires_end_state = true.
• No cash flows exist and none ever will — a currency, a collectible, a bare commodity holding. The asset can be priced, never valued. Emit no-intrinsic-valuation and route to pricing only. See skill_view("valuation-playbooks", file_path="references/value-vs-price-gap.md").
• Fewer than three years of statements, or statements that mix personal and business expense — private-company cleanup runs before any forecast. See B13.

Return blocked when the minimum viable input set is missing and the user has supplied no substitute. Name what is needed. Do not guess and proceed.

S1 · Signal extraction

Every signal is computed from data, never asserted. Each carries a confidence and the evidence that set it. These are the only inputs the gates in S2 through S6 may consult.

The arithmetic here is mechanical. The judgment is whether a computed signal is representative: a coverage ratio taken at a trough, a regression beta on a stock that barely trades, a margin distorted by one large contract. Record that judgment where you make it. Computation detail and the traps in each signal: references/signal-extraction.md.

S2 · Sector gate

Evaluated first and outranking every other gate, because sector decides whether firm-level cash flow and firm-level leverage are even defined.

if   sector_type == financial-service          -> B5   (ends the FCFF path entirely)
elif sector_type == real-estate/REIT           -> B5b  (payout mandated)
elif macro_driver exists and R^2 is high       -> B6   (commodity)
elif sector_type == cyclical-industrial
     and earnings_status in {trough, peak}     -> B7   (cyclical normalization)
else                                           -> continue to S3

Commodity and cyclical are separated by one question: does a usable price driver exist? A high-R² regression of revenues on the commodity price sends the firm to B6, because the price essentially is the revenue model. No usable price series sends it to B7.

For a REIT, mandated payout and tax status make optimal-debt-ratio work and retention-based growth meaningless. Value on dividends or FFO, and exclude the firm from any corporate-finance financing recommendation. All other routing continues normally.

Bank statements announce themselves. Revenue arrives as net interest income and net fee income rather than sales minus cost of goods. Interest expense sits inside operations. The credit-loss provision is a recurring operating expense. Property and equipment is negligible, and equity is a thin slice of assets — HSBC in 2019 held equity worth 7.1% of total assets, which is structure rather than a warning. Detail: skill_view("valuation-playbooks", file_path="references/sector-differences-in-financial-statements.md").

S3 · Statement-repair gate

Always executed. This is what sets what "earnings" means for everything downstream. The order is fixed because each step feeds the next.

1. Update to trailing twelve months.
2. Capitalize operating leases into lease debt, restated EBIT, restated invested capital.
3. Capitalize R&D, recruiting or brand-building advertising when intangible_intensity is moderate or high. This is overlay B8.
4. Strip one-time items and personal expenses. Charge a market salary for uncompensated owner labour at a private firm.
5. Normalize only if S4 routes to B7. Normalization and the revenue-driven path are mutually exclusive.
6. Resolve the circularity. Restated EBIT sets interest coverage, which sets the synthetic rating, which sets the cost of debt, which sets the lease discount rate, which changes restated EBIT. Iterate to a fixed point.

Every downstream consumer of EBIT, invested capital, ROIC, coverage and reinvestment uses the same restated basis. Mixing bases is the most common silent error in this pipeline.

financial-statement-normalization performs steps 2 through 4 and the ratio pack; cost-of-capital-toolkit supplies the synthetic rating that closes the loop in step 6.

S4 · Ownership and transaction gate

if   ownership == private and transaction_motive != ipo  -> B13 (plus sub-scenario)
elif transaction_motive == ipo                           -> B14
elif ownership == division/subsidiary being separated    -> B10 (sum-of-the-parts)
elif transaction_motive == acquisition                   -> B15 (four-number chain)
elif transaction_motive == restructuring
     or an activist/control question is live             -> B2
else                                                     -> continue to S5

This gate fixes two things: the discount-rate identity — whose risk is being priced — and the discount stack of illiquidity, minority and key-person adjustments. It does not choose the cash-flow engine. S5 does that.

The identity question matters more than it looks. A private business does not have one value; it has a value per buyer and per purpose. The gap between a valuation for an undiversified individual buyer and one for a diversified public acquirer is the bargaining range, not an error. Quote one number to both sides of a negotiation and you have answered neither question.

S5 · Life-cycle and earnings gate

This selects the primary cash-flow engine.

if   revenue_status == pre-revenue
     or earnings_status == negative-structural
     or (life_cycle_stage in {start-up, young-growth}
         and earnings_status != profitable)        -> B1  revenue-driven, work backwards
elif earnings_status in {negative-transient,
                         cyclical-trough,
                         cyclical-peak}            -> B7  normalize, then standard engine
elif revenue_status == declining
     and life_cycle_stage == decline               -> B3  negative growth and reinvestment
elif life_cycle_stage in {mature-growth, mature-stable}
     and policies look consistent, stable and bad  -> B2  status quo versus optimal
else                                               -> standard path

The B7-versus-B1 call is the sharpest judgment in the whole gate. Normalization is legitimate only when the trouble is temporary. Evidence for temporary: the sector is in a known downturn, peers show the same pattern, the firm earned normal margins for years, and the balance sheet survives until recovery. Evidence against: falling market share, a structural demand shift, leverage forcing asset sales. All three normalization approaches will happily produce a healthy EBIT for a firm that will never earn it again.

"Consistent, stable and bad" needs evidence too. Stability is not quality. Test for ROIC < WACC, reinvestment far off the sector norm, or a debt ratio far from the sector optimum.

Standard path selection, when no special branch fires:

• Equity or firm. leverage_state == stable takes the equity route. changing, or incomplete leverage data, takes the firm route through FCFF and WACC.
• On the equity route, apply the 80% / 110% payout screen to choose dividends against FCFE.
• Stage count from the g_econ + 10% screen in S1.

S6 · Survival and truncation gate

Applied as an outer wrapper after the engine produces a going-concern value. A DCF values a company that lives long enough to reach stable growth. When it might not, the DCF overstates value, and the repair is never a higher discount rate.

survival_in_doubt = any of:
   life_cycle_stage in {start-up, young-growth}    -> B1 requires it by default
   revenue_status == declining and leverage high   -> B3 hands off to B4
   coverage < 1 or rating <= CCC                   -> B4
   traded bonds well below par                     -> B4, invert the price for pi
   bank near its regulatory capital minimum        -> B5, wipeout probability
   expropriation or regime-change exposure         -> B9 truncation
   market D/(D+E) > 50% and earnings negative      -> B4 plus equity-as-option cross-check

When it fires, route to B4, or to the truncation sub-branch of B9, or both. Source the probability from the most informative channel available. In ascending order: sector survival tables, then the rating-implied cumulative default rate, then a statistical model, then an inverted traded bond price. The last is sharpest because it is the market's own number.

Firm failure and equity wipeout are different events. A bailout can save the firm and destroy the equity. Say which one is being modelled.

S7 · Overlays and constraint compilation

Overlays are non-exclusive and compose. Assign every overlay whose trigger fires, then compile the union of the constraint sets from every branch that fired. S8 resolves any conflict.

Constraint compilation is mechanical once branches are assigned. Take every constraint listed against every fired branch in the catalogue, plus the two universal ones, and write them into constraints with a reason and a source_branch. Compile nothing that no branch emitted. A constraint with no source is an opinion.

S8 · Combination rules

Real companies sit in several buckets at once. These rules make composition deterministic.

R1 — One engine only. Precedence, highest first:

B5 financial-service  >  B13/B14 private or IPO (rate and discount identity)
                      >  B4 distress with equity wipeout
                      >  B1 revenue-driven
                      >  B7 normalized
                      >  B3 declining
                      >  B2 status quo versus optimal
                      >  standard

B10 sum-of-the-parts is a decomposition rather than an engine. It runs this precedence list once per division, then aggregates.

R2 — Overlays never replace the engine. B8, B9, B11, B12 and B16 modify inputs, the discount rate, or the bridge. They never change which cash flow is discounted.

R3 — Mutually exclusive pairs. A violation is a hard error, not a warning.

R4 — Ordering when several branches fire. The pipeline is fixed:

1. Clean accounts        S3: leases, R&D via B8, one-times, owner salary
2. Fix base earnings     B7 normalize, or B1 abandon earnings and drive from revenue
3. Build the rate stack  bottom-up beta -> total beta if B13-I
                         -> exposure-weighted ERP or lambda if B9
                         -> synthetic rating including country spread if B9
                         -> divisional rates if B10
4. Run the engine        per R1, and per division if B10
5. Outer adjustments     B4 failure, B9 truncation, B2 P(change) — each applied once
6. Equity bridge         debt including leases, minorities at market, cash,
                         B11 cross-holdings, employee options
7. Discount stack        B13 illiquidity then minority; key-person is already in EBIT
8. Uncertainty           scenario grid, then simulation
9. Price comparison      gap, catalyst, expected return

R5 — Composition arithmetic. Probability-weighted branches compose multiplicatively on the going-concern value, and each risk appears exactly once:

V = V_going_concern × Π_i (1 − p_i × loss_fraction_i) + Σ_i p_i × proceeds_i

In practice, never stack more than two without arguing that the events are genuinely distinct. A country risk premium plus a nationalization scenario plus a governance discount is three charges for one overlapping risk.

R6 — Governance and country risk are cash-flow facts. Model weak governance as low returns on capital and a low probability of change. Never as a discount-rate bump or a flat value haircut. The same holds for emerging-market risk beyond the exposure-weighted premium.

R7 — Common multi-branch shapes. Seven worked resolutions, including the distressed emerging-market bank and the cyclical firm in a shock, are in references/combination-worked-routings.md.

R8 — Mode interaction. The mandate's mode selects what runs after the engine, never the engine itself. acquisition wraps the engine in B15's four-number chain. restructuring routes to B2. corporate-finance runs the capital-structure and payout stages, and those two stages are suppressed entirely for B5 and B5b.

The constraint catalogue

Emitted by the branches, enforced by every downstream agent, checked by the critic under rule V9. The IDs are stable identifiers and are the contract between agents. Write them exactly as spelled here.

Writing classification.json

The contract in the suite design spec §5, bundled as skill_view("valuation-playbooks", file_path="references/suite-design-spec.md"), extended by the S7 route record. Both blocks belong in the one file.

{
  "life_cycle_stage": "young|growth|mature|aging|declining",
  "earnings_status": "profitable|marginal|negative|cyclical-trough",
  "sector_type": "non-financial|financial-service|commodity-cyclical|real-estate",
  "ownership": "public|private|subsidiary",
  "geography": {"incorporation": "US", "operations": [{"region": "...", "revenue_share": 0.0}]},
  "distress_markers": {"present": false, "evidence": []},
  "intangible_intensity": "low|moderate|high",
  "primary_path": "standard-fcff|standard-fcfe|dividend-discount|excess-return|revenue-driven|distress-adjusted|asset-based",
  "engine_branch": "B1..B16",
  "overlays": ["intangible-heavy", "emerging-market", "multi-business", "cross-holdings", "has-real-options", "macro-shock"],
  "transaction_overlay": "acquisition|restructuring|ipo|private-sale|none",
  "discount_stack": ["illiquidity", "minority", "key-person"],
  "constraints": [{"rule": "no-optimal-debt-ratio", "reason": "...", "source_branch": "B5"}],
  "pipeline": ["clean", "normalize|revenue-drive", "rate-stack", "engine", "outer-adjustments", "bridge", "simulate", "price-compare"],
  "confidence": "high|medium|low",
  "unresolved": ["what would change the routing"]
}

Three rules on the writing itself. primary_path and engine_branch describe the same choice and must agree — a B5 branch never carries a standard-fcff path. Exactly one engine branch appears; overlays go in overlays. Every constraint carries both a reason in plain words and the branch that emitted it, because the critic reads the reason when it finds a violation. Field-by-field guidance and the diagnosis.md companion: references/classification-artifact.md.

Confidence and what stays unresolved

Set confidence from the evidence, not from how clean the JSON looks.

unresolved lists the evidence that would change the routing, stated so someone could go and find it. "A traded bond price would replace the rating-implied default probability" is useful. "More information about the company" is not.

Low confidence is not a reason to stop. It is a reason to widen the margin of safety, run the scenario grid on the two drivers this branch turns on, and say plainly in diagnosis.md which judgment the answer rests on. The recurring pattern across every branch is that the arithmetic is mechanical and the inputs are not. Route correctly first, then spend the effort on the two or three judgments the branch actually turns on.