01
Consent and eCBSV clear first.
The identity gate is a precondition. It can control the verdict on verification failure, but it is not a corroboration axis and does not enter the index.
Formal methods
How the corroboration verdict is computed
Kenshiki's public method is not a weighted sum of signals. It is a deterministic chain: consent gate, seven graded axis judgments, categorical contradiction guard, named coherence rules, versioned policy, validation artifacts, and a sealed record a lender can inspect.
Method summary
The verdict resolves a precedence chain, not a composite score.
The core design is deliberately narrow: preserve what was asked, what evidence answered, where uncertainty remained, which rule fired, and why the resulting record is reviewable later.
02
Each source returns a structured judgment.
Axis outputs carry band, coverage, support, hypothesis, evidence class, and source attestation instead of returning raw source data.
03
Contradictions stay non-fungible.
Named rules evaluate whether the file holds together, and the categorical guard blocks approval when any admitted axis is contradictory.
The verdict equation
One equation captures the precedence chain.
The formal method is easiest to audit as a routing function. The gate is evaluated first; if it fails, it controls the verdict. If it clears or is not run for lack of consent, the policy layer reads the axis judgments and coherence rules.
Read this as: for application A, a failed gate G overrides the result; any contradictory axis blocks approval; otherwise product policy Π for product p evaluates the set of axis judgments J and product-specific coherence rules R.
- Gate first: eCBSV is a precondition, not a weighted signal.
- Seven axes stay named: each axis judgment preserves band, coverage, support, and reason scaffold.
- Contradictions stay non-fungible: a contradictory axis blocks approval instead of being averaged away.
- CCI stays outside: the index may summarize the record, but it is not an argument to the verdict function.
The index boundary
The CCI is legibility, not the decision.
The Corroboration Confidence Index summarizes the admitted evidence record so a reviewer can read the file quickly. It is not the approve, refer, or decline rule. The recommendation path uses gates, bands, coherence rules, and versioned policy thresholds.
The number makes the record easier to scan. The named rule chain makes the record defensible.
This separation matters for adverse-action and model-risk review. If a file is referred or stopped, the reason basis should trace to the specific gate, axis, missingness, or coherence rule that actually mattered — not to a generic failure to achieve a score.
Axis judgments
Each axis answers an entailment question.
An axis asks whether admitted evidence corroborates a bounded claim — for example, that presence fits the claimed timeline, that income is verified at source, or that identity elements show continuity. Each axis returns a band, coverage, support, and a reason scaffold, never raw source data. The full set of corroboration axes, and the consented identity gate, are described on the approach page.
The source evidence supports the bounded claim strongly enough for the policy layer to count it as positive evidence.
The evidence leans positive but does not carry the same assurance as a corroborated result.
The source could not answer enough of the question. Missingness stays visible rather than being imputed into a midpoint.
The evidence conflicts with a required claim. The policy layer treats material contradiction as a named risk surface.
Validation layer
Independence and calibration are measured, not assumed.
The 60-case seed corpus is useful as a methodology artifact: it regenerates the operating-point matrix, verifies that contradictory axes no longer survive into approval, and reports calibration and dependence metrics. It is not statistically powered production validation.
The current roadmap keeps the hard floor live while treating calibrated log-likelihood-ratio aggregation as a shadow-mode validation target. Before posterior-style evidence weights can influence recommendations, Kenshiki needs representative outcomes data, approved smoothing and calibration thresholds, a dependence-correction method, and model-risk review.
Regulatory posture
The record is built for specific reasons and repeatable review.
The method is designed around the practical bar that credit decision evidence must be explainable, reproducible, and reviewable. Kenshiki does not claim this alone decides legal classification; it creates the engineering posture a lender, model-risk team, and counsel can inspect.
That posture aligns with the same themes discussed in SR 26-2 and model validation: documented assumptions, versioned thresholds, third-party input review, and a record that can be replayed by people who did not build the system.
For conforming mortgage, the boundary is narrower still: Kenshiki evidence can sit beside GSE-validated decisions as lender-held context. It does not feed Desktop Underwriter or Loan Product Advisor and does not claim to upgrade conforming-loan eligibility. See the mortgage/GSE channel note.
FAQ
Common questions
What the formal method claims, and what it deliberately does not claim.
- Is the Corroboration Confidence Index a credit score?
- No. The index is a bounded legibility measure for the admitted evidence record. Kenshiki’s recommendation path is driven by gates, bands, named coherence rules, and versioned policy logic rather than the index itself.
- Does Kenshiki average all signals into one weighted score?
- No. Contradictions, missingness, and gate failures stay visible as named surfaces. A strong signal on one axis cannot simply buy back a material contradiction on another axis.
- What makes the method deterministic?
- The same attested evidence objects, policy version, decode settings, and as-of timestamp produce the same verdict record. Hash-derived audit details are stable rather than random.
- Does the eCBSV identity gate count as a corroboration axis?
- No. eCBSV is a consented precondition that can control the verdict when verification fails. It is separate from graded identity-continuity corroboration and does not enter the CCI.
- Are the axes assumed independent?
- No. The current validation layer measures dependence and effective axis count on the seed corpus, and calibrated posterior-style aggregation remains a roadmap item until representative outcomes data supports it.