Foundational Scoring &
Validation Engine

No system may claim certainty it cannot justify.

FSVE v3.6

CEV AUDITED · 8 FINDINGS RESOLVED

M-MODERATE

AXIS INPUTS — 6 DIMENSIONS

WEIGHTS: 1.000 ✓ (auto-normalized)

k_bottleneck — Bottleneck Multiplier 1.50

Default: 1.5 · Range: [1.0, 2.0] · Safety-critical override: 1.0

SUPERVISED

EV 0.500

Score is within the supervised zone. Human review required before deployment-grade use.

EV FINAL

0.500

UNCERTAINTY MASS

0.500

EV BASE

0.500

CRA (AGREEMENT)

0.000

GINI (LAUNDERING)

0.000

BOTTLENECK AXIS

—

THRESHOLD PROXIMITY

—

⚠ LAUNDERING DETECTED — Gini G > 0.15. Score distribution is highly unequal. One axis may be inflating EV_base. Review axis inputs before treating score as valid.

SENSITIVITY — HIGHEST LEVERAGE AXIS

—

CEV v1.0 — FINDINGS SUMMARY · v3.5 → v3.6

FINDING	LOCATION	TYPE	STATED (v3.5)	CORRECTED (v3.6)	SEVERITY
CEV-F001	§9.1 Gini	Formula error — sign	Negative for all inputs	G = (2Σ(i×s_i))/(nΣs_i) − (n+1)/n	⛔ CRITICAL
CEV-F002	§11.4 CRA	Domain error — no floor	CRA can be −0.131	CRA = max(0, 1−σ/μ)	⛔ CRITICAL
CEV-F003	§11.2 Teleology	Division by zero	Unguarded	if sim_T+sim_M=0 → TS:=0	🔴 STRUCTURAL
CEV-F004	§17 projected EV	Arithmetic error	0.845	0.8227 (bottleneck shifts to L)	🟠 MAJOR
CEV-F005	§17 min E for VALID	Incorrect inference	E ≥ 0.75	E ≥ 0.62 (L is next bottleneck)	🟠 MAJOR
CEV-F006	§17 EV_base	Rounding	0.788	0.786	🟡 MINOR
CEV-F007	ODR-007 5-year	Unit error	157,248,000 s	157,680,000 s (−5 days)	🟡 MINOR
CEV-F008	§4.2 ES bottleneck	Edge case — empty set	Undefined	Guard added	🟡 MINOR

GINI FORMULA — DETAILED ERROR RECORD

The v3.5 Gini formula G = 1 − (2×Σ(i×s_i))/(n×Σs_i) is archived as GINI-FSVE-ERR-001. The standard rank-sum Gini formula is G = (2×Σ(i×s_i)) / (n×Σs_i) − (n+1)/n. The v3.5 formula applied 1 − [numerator term] instead of subtracting (n+1)/n. For a uniform distribution of n=5 items: the correct term equals (n+1)/n = 1.2, making correct G = 1.2 − 1.2 = 0. The v3.5 formula produced 1 − 1.2 = −0.2. The laundering check G < 0.15 could never have been triggered. All v3.5 scores that invoked the Gini laundering check were missing this detection layer entirely.

INPUT	v3.5 RESULT	v3.6 RESULT	EXPECTED
[0.5 × 5]	−0.200	0.000	0.000 ✓
[0,0,0,0,1]	−1.000	0.800	(n−1)/n = 0.800 ✓
[0.8 × 5]	−0.200	0.000	0.000 ✓
[0.55 × 5]	−0.200	0.000	0.000 ✓

⚠ v3.5 CLEARANCE VOID NOTICE — Any laundering clearance issued under FSVE v3.5 using the Gini primary check is void. The Gini layer was mathematically non-functional for all inputs. Re-run §9.1 primary check under v3.6 for all affected ScoreTensors. ScoreTensors that did not invoke the Gini check are unaffected.

BOTTLENECK SHIFT ANALYSIS — EV PROJECTION CORRECTION

v3.5 stated: raise E from 0.35 to ≥ 0.75 to achieve projected EV = 0.845. Both figures were wrong. When E rises above L = 0.62, the binding constraint shifts from E to L. At that point EV = min(EV_base, 1.5 × 0.62) = min(0.786, 0.93). Since EV_base = 0.786 < 0.93, the cap is not binding — EV = 0.786 at E=0.62, rising to 0.8227 as E increases further. Corrected path to VALID: E must rise from 0.35 to ≥ 0.62. VALID status (EV ≥ 0.70) is achievable sooner than v3.5 stated. Every FCL entry that raises E above 0.62 moves EV_base upward from there.

CEV v1.0 VERDICT

FSVE v3.5 contained two critical formula errors (Gini, CRA domain) that were non-functional. The Gini check was mathematically incapable of detecting laundering in any configuration. The CRA formula could produce results outside its implied domain [0,1]. Both are corrected in v3.6. This is a Minor release — all v3.5 ScoreTensors are valid except those that invoked the Gini laundering check, which must be re-evaluated. EV = 0.525 is unchanged. v3.6 fixes formula errors, not evidence gaps. Path to VALID requires FCL entry accumulation.

CEV Audit conducted: Sheldon K. Salmon (AI Reliability Architect) + Claude (Anthropic) · February 2026

FIVE FOUNDATIONAL PRINCIPLES

PRINCIPLE 1

No Free Certainty

Certainty must be earned through evidence, consistency, and bounded assumptions. If certainty increases in one dimension, something measurable must account for the gain.

PRINCIPLE 2

Uncertainty Is Conserved

Uncertainty may be reduced, bounded, transferred, or deferred. It may never be erased silently. A score that omits uncertainty is not a score — it is a false claim.

PRINCIPLE 3

Scores Are Claims, Not Truth

Every score is a claim about reality. Therefore: every score must be explainable, every score must be reversible on new evidence, and every score must degrade under contextual stress.

PRINCIPLE 4

Invalidatability Is Required

Any scoring system that cannot produce the output "this score is invalid" is not a scoring system. It is decoration.

PRINCIPLE 5

Structural Honesty Precedes Numerical Accuracy

A structurally honest score of 0.40 is more valuable than a structurally dishonest score of 0.90. The architecture of how a score was produced matters as much as its value.

SIX SCORE CLASSES

CONFIDENCE — How well the system understands intent or claim structure

Cannot exceed Information Completeness Score. Cannot ignore unresolved contradictions. Failure mode: false alignment — high confidence with misunderstood intent.

CERTAINTY — How likely a claim is to remain valid under structured challenge

Cannot increase without a corresponding reduction in Uncertainty Mass. Cannot be high when System Fragility is high. Failure mode: overconfidence collapse.

VALIDITY — Whether a score itself is legitimate (meta-scoring)

Hard Rule: If Validity < 0.40 → all downstream scores from this system are suspended until remediated. Failure mode: scoring nonsense confidently.

COMPLETENESS — What fraction of the scoring surface has been assessed

Completeness cannot imply correctness. A complete but wrong score is still wrong. Failure mode: coverage confusion — high Completeness mistaken for quality signal.

CONSISTENCY — Internal coherence across definitions, rules, and outputs

Each unresolved contradiction applies a ceiling reduction. Failure mode: self-contradiction that sounds internally coherent — locally consistent, globally incoherent.

RISK EXPOSURE — Potential damage magnitude × likelihood across failure modes

Cannot be averaged away or hidden by high confidence on other axes. Failure mode: low-probability catastrophic blindness — rare but severe failure modes excluded.

CURRENT ENGINE STATE

DEGRADED — EV = 0.525

v3.6 corrected two critical formula errors and six additional findings. The formula corrections do not change EV — they restore the engine to accurate operation. EV = 0.525 reflects the current evidence gap, not a formula problem. Path to VALID: E-axis must rise from 0.35 to ≥ 0.62 via FCL (Finding Confirmation Log) entries. At E = 0.62, EV reaches 0.811 — VALID status. Every confirmed finding above 0.62 moves EV_base upward. FCL entry accumulation is the only path.

DOCUMENTATION & ATTRIBUTION

FULL SPEC — AION-BRAIN STACK OVERVIEW EIGHT LAWS AIONSYSTEM.GITHUB.IO

FSVE v3.6 · Foundational Scoring and Validation Engine · Author: Sheldon K. Salmon — AI Reliability Architect · February 2026 · CEV Audit: Claude (Anthropic) · Supersedes: FSVE v3.5