◂ Back to cross-company loss  ·  Mission Control SAA Alliance · KOKON Methodology Contract
SAA ALLIANCE KOKON METHODOLOGY CONTRACT — CONFIDENTIAL · PRE-CLIENT
DIAGNOSTIC ENGINE · METHODOLOGY & RENDER CONTRACT  ·  JUNE 2026

How KOKON dollarizes a finding, scores severity, suppresses inapplicable detectors, and proves the report did not contradict itself

Phase 524.9 closure of the SAA Independent Audit Memorandum (KOK-F-01 through KOK-F-07; RND-01 through RND-10). Published as a controlled contract — these rules are enforced in code and exercised by golden-file regression on every renderer change.
1Summation rules — recoverable cash never double-counts

The four typed-dollar columns (recoverable cash · exposure at risk · peer-benchmark gap · not scored) are never summed across types. This was the existing discipline; it is enforced on every aggregation surface.

Phase 524.10 (auditor verdict #2 closure, KOK-F-08) replaces the earlier "keep parent, drop components" rule. The replacement: within each lever family the headline is the largest coherent lever — the maximum of (parent impact, sum of components). The reason: parent and components are dollarized on different denominators (CCC drag on COGS/day base, DSO drift on revenue/day base) and cannot strictly nest just because the underlying day-metrics intersect. For revenue ≫ COGS entities (NVDA), a DSO component dollar can legitimately exceed the CCC parent dollar.

Operational rule. Each finding carries lever_family and lever_role ∈ { parent · component · standalone }. The sum_deduplicated field on every bucket is the auditor-aligned headline:
family_max = max(parent_impact, Σ component_impacts)
headline    = Σ family_max + Σ standalone

This reads as "the upper-bound recoverable cash from any single execution path that doesn't double-count" — executing the parent strategy and the component strategy will not double the recovery, so the headline takes the larger of the two coherent measures.

Example. NVDA recoverable findings fire as { CCC parent $232M · DSO component $600M · AP early-pay component $36M }. Gross sum = $869M; components sum = $636M. Deduplicated headline = max($232M, $636M) = $636M.

2Confidence taxonomy — four tiers, every figure inherits its lowest input

Every figure inherits the lowest tier among its inputs. A derived_cash value computed on an assumed_baseline reports as assumed_baseline. No more "hard_fact · disclosed inputs" tag on values that were neither disclosed nor calculated from disclosed primaries.

TierDefinitionExample
hard_factDisclosed primary data, unmodified.NVDA revenue $60.92B from the 10-K.
derived_cashArithmetic on hard_fact inputs only.Carrying cost = trapped working capital × cost-of-capital.
researched_estimateThird-party / OSINT / supplier-side estimate, not in primary disclosure.NVDA top-vendor share 59.3% inferred from supply-chain research; peer headcount median.
assumed_baselineAnalyst-set threshold or baseline driving the trigger.DSO "excess days" baseline · 45-day CCC target · 15% inventory-old-stock threshold.
3Severity normalization — materiality, not absolute dollars

Legacy bands keyed severity to absolute dollars (>$5M Critical). On a megacap that converts every finding to Critical regardless of materiality. Phase 524.9 keys severity to the ratio of finding to revenue (or working capital / equity where appropriate), with absolute dollars shown alongside.

TierBandExample
Critical> 1.00% of baseVeridian vendor concentration $4.9M on $95M revenue = 5.16% of revenue.
Material0.25% – 1.00% of baseNVDA CCC drag $232M on $61B revenue = 0.38% of revenue.
Monitor< 0.25% of baseNVDA AP early-pay $36M on $61B revenue = 0.06% of revenue.

Bands are calibrated against the operator's investment-grade peer set.

For vendor concentration specifically, severity additionally consults concentration depth (Phase 524.11 KOK-F-12 closure, revising Phase 524.10 KOK-F-11): <40% top-vendor share = Monitor · 40-60% = Material · ≥60% sole-source = Critical. The final tier is MAX(materiality_tier, concentration_tier) — both rules can RAISE severity, neither lowers it. A $10B realistic LGD cannot collapse to Monitor merely because concentration is shallow (materiality dominates); equally a sole-source 80% dependency cannot collapse to Monitor merely because its absolute dollar is small (concentration dominates). The earlier MIN rule was rejected by the auditor on exactly this asymmetry.

Honest gap · the MAX-rule is asymmetric by design but ungrounded by outcome data today. The Phase 85 BrainOutcomeLedger schema accepts confirmed / refuted / uncertain per finding, and the /api/control/brains/observation endpoint accepts writes — but no production code path automatically records the verdict of a KOK-F finding after the operator acts on it. So there is no FP-rate per severity-class measured today, and the MAX ratchet has no symmetric anti-inflation check derived from real outcomes. Phase 524.23B2 wires the operator confirm/refute UI on every finding and surfaces a per-class FP-rate; until that lands, treat KOK-F-12 as conservative by construction, not validated by feedback. The 22-agent ARIN council brain calibration (Phase 85 / 85b) is a separate ledger; it has writers via /ui/agents-live and is unrelated to KOK-F finding outcomes.

SAA Alliance  ·  KOKON Methodology Contract · Phase 524.9 · auditor-aligned closure PAGE 1 / 2
SAA ALLIANCE KOKON METHODOLOGY CONTRACT — CONFIDENTIAL · PRE-CLIENT
4Detector × vertical applicability — suppression is not silent

Some detectors are mathematically inapplicable for entire business models. Running them on the wrong vertical produces dollarized findings the engine cannot defend. The matrix below is enforced at dispatcher time; suppression produces a visible NOT_SCORED finding tagged inapplicable (vertical=…).

DetectorManufacturingSaaSFinancialRetail / Wholesale
DSO / receivables long tail
Inventory aging excessNOT_SCOREDNOT_SCORED
Cash conversion cycle dragpartial (no DIO)NOT_SCORED
AP terms mismatch
Vendor concentration
Headcount / revenue ratio
Period-over-period flux
5Exposure — realistic loss-given-disruption, not 100% of annual spend

Vendor concentration exposure was historically reported as top_vendor_share × total_spend — i.e. 100% of annual supply through the vendor for a full year. Phase 524.9 replaces that headline with a realistic loss-given-disruption computed as gross × switching-cost / lead-time-gap factor over a typical recovery cycle.

Operational rule. Headline dollar_impact = realistic LGD. gross_dollar_impact is retained as secondary disclosure. The LGD factor scales with concentration depth: 20% baseline; 35% when concentration ≥ 60% (sole-source territory where recovery is materially harder).

Example. AAPL: 35.7% × $140B = $50B gross exposure → $10B realistic LGD at 20% factor. The board reads the realistic number; the gross stays auditable as secondary.

6Render Invariant Spec — what the renderer refuses to ship

Machine-checked invariants enforced at render time. Behaviour on violation: BLOCK refuses to emit · AUTO-FIX applies a deterministic correction from available data. Every renderer change is diffed against a golden-file regression set before release.

RND-01Declared counts equal rendered counts (gates, agents, findings).BLOCK
RND-02Any value referenced in one section that exists elsewhere in the document must match it exactly.BLOCK
RND-03No / blank placeholder where the value exists in the same artifact.AUTO-FIX
RND-04Self-audit findings reconcile to the rendered surface they describe.BLOCK
RND-05No hedging lexicon (may refer, possibly) in a control / audit finding.BLOCK
RND-06Truncation only at sentence boundary, with [full text in JSON] marker.AUTO-FIX
RND-07Date / identifier lineage complete and single-format (input → signed → rendered).AUTO-FIX
RND-08Blocking gate metrics distinguish NO_DATA from a real 0.BLOCK
RND-09Severity / confidence labels conform to the published taxonomy (sections 2 and 3).BLOCK
RND-10Test fixtures carry the non-production watermark.BLOCK
RND-11Cross-vertical AND cross-detector boilerplate. A finding whose narrative references vertical-specific entities (e.g. specialty pharma in a SaaS run, Phase 524.12 KOK-F-19) OR phrases from another detector's template (e.g. "switching-cost + lead-time-gap" on an inventory finding instead of vendor, Phase 524.14 KOK-F-24) is template leakage — the renderer blocks emission.BLOCK
RND-12Working-capital sanity-floor visibility. Every scored DSO / CCC / AP / inventory finding must mention the balance-sheet sanity floor in its dollar basis — "trapped $X ≤ disclosed $Y ✓" or an explicit "check not run". Silent absence is indistinguishable from "floor not run" and the renderer refuses to emit. Phase 524.14 KOK-F-25 closure.BLOCK
7Golden-file regression · the closing mechanism (Phase 524.15)

The auditor's exit prescription: "По существу аудит закрыт. Дальше — не новые findings, а регрессионный механизм." Phase 524.15 lands that mechanism. Seven canonical scenarios — five demo clients (Veridian, AAPL, NVDA, MSFT, Cascade) plus two edge cases (no_inputs, dso_floor_fail) — are frozen as input / expected pairs under internal/handlers/testdata/golden/.

Operational rule. Every commit that touches EnrichFinding, RealDispatcher, or buildRunResult runs make golden-test. Any drift fails the suite with a structural diff (severity / confidence / lever_role / dollar_impact per finding plus a full JSON dump). A TestPhase52415_GoldenFile_DriftIsDetected meta-test injects a synthetic drift and asserts the comparator catches it — so a silently-broken suite cannot exist.

Regenerating goldens after an intentional change: make golden-update. The Makefile prints the diff for review before commit. The auditor's contract: changes to the surface must be diffed; surface drift must not pass silently.

What's locked. Every Phase 524 rule — typed-dollar discipline (524.9 §1) · max-coherent-lever dedup (524.10 KOK-F-08) · materiality + concentration MAX severity (524.11 KOK-F-12) · vertical applicability (524.9 KOK-F-04) · 4-tier confidence taxonomy (524.9 KOK-F-02) · balance-sheet sanity floor (524.12 KOK-F-17) · realistic LGD on every exposure detector (524.13 KOK-F-20) · always-visible floor (524.13 KOK-F-21) · detector-specific narrative (524.14 KOK-F-24) · WC floor mention (524.14 KOK-F-25 / RND-12) · cross-detector / cross-vertical boilerplate detection (524.12+524.14 / RND-11) · data sufficiency (524.13). Future changes that drift any of these without explicit golden-update will fail.

7Statement of evidence

The four controllable gaps the auditor identified — aggregation integrity, confidence taxonomy honesty, calibration normalization, and render integrity — are closed in code under Phase 524.9. Lock-in tests TestPhase5249_AuditorClosure_* exercise each rule against the live engine. The five reviewed demo runs (Veridian Cloud · Apple · NVIDIA · Microsoft · Cascade Outfitters) re-render under these rules and are available at /ui/clients.

This contract supersedes the implicit methodology used in pre-524.9 artifacts. Where a re-rendered figure differs from a prior version, the auditor-aligned figure governs. Earlier figures remain in the WORM ledger for audit, prefixed with the pre-524.9 generation timestamp.

SAA Alliance  ·  KOKON Methodology Contract · Phase 524.9 · auditor-aligned closure PAGE 2 / 2