Scoring Rubric - Alpha Documentation

Alpha Governance Ratings are not checklist scores. The rubric rewards governance that is owned, evidenced, and demonstrated under pressure. A company does not earn a top result for publishing principles alone. It earns it by showing that governance changed a real decision, constrained risk, or held under stress.

Rubric Principles

Constraint beats commitment. A delayed launch, blocked use case, or documented override matters more than a policy statement.
Specificity beats slogans. Named owners, cadences, thresholds, and escalation paths score above generic responsible AI language.
Verified evidence beats self-description. Hard-to-game evidence carries more weight than marketing copy.
Repeatability beats one-off disclosure. A recurring board review or tested control scores above a single announcement.
Weak links matter. A strong composite cannot erase a critical failure in safety, regulatory posture, or corrigibility.

1. Indicator-Level Anchors

Most Alpha indicators are scored on a 1-10 scale and then rolled into 0-100 dimension scores. Dedicated layers such as D11 and D12 can also use 0 when the company shows no demonstrable capability.

Indicator band	Meaning	What usually has to be true
9-10	Demonstrated leader	Named ownership, explicit governance design, recurring oversight, and evidence that the control worked in practice. Independent verification or stress evidence is often present.
7-8	Governance-ready	Clear owner, documented process, credible implementation evidence, and board or committee visibility. Gaps remain in coverage, specificity, or proof under stress.
5-6	Developing	The company has started building the control, but implementation is uneven, incomplete, or localized. Governance exists, but it is not yet institutionalized.
3-4	Lagging	Governance is reactive, fragmented, or mostly descriptive. Responsibilities are unclear, evidence is thin, and practice trails disclosure.
1-2	Deficient	Little credible evidence of a functioning control. The company is relying on aspiration, boilerplate, or ad hoc judgment.
0	No demonstrated capability	Used sparingly for dedicated layers, especially where the question is binary in practice: has governance ever actually intervened?

Tie-break rule

When two adjacent scores are both plausible, the lower score wins unless there is evidence of either:

recurring execution over time, or
governance holding under stress.

That rule is deliberate. It prevents polished disclosure from outranking demonstrated control.

2. Evidence Quality Rubric

Alpha distinguishes between the strength of a claim and the strength of the evidence supporting it.

Evidence quality	Meaning	Typical sources	What it can support
Verified	Independently auditable or regulator-grade evidence	SEC filings, enforcement actions, court records, AIID operator records, third-party audits, formal attestations	Highest-band scores, ceiling removal, strong confidence
Disclosed	Specific company disclosure with enough detail to evaluate	Proxy statements, governance reports, model cards, committee charters, published frameworks	Mid-to-high scores when internally consistent
Inferred	Credible but indirect signal	Cross-signal patterns, architecture clues, peer-consistent behavior, reputable reporting without direct primary support	Directional scoring only; rarely enough for top-band claims by itself
Absent	No credible evidence, or evidence contradicts the claim	Silence, boilerplate, stale disclosure, uncorroborated marketing language	No substantive credit

3. Confidence Standard

Dimension confidence is derived from evidence density, not from how polished the narrative sounds.

Confidence	Threshold
High	60% or more of indicators are supported by verified or disclosed evidence
Medium	30% to 59% of indicators are supported by verified or disclosed evidence
Low	Less than 30% of indicators are supported by verified or disclosed evidence

High confidence does not mean perfect governance. It means Alpha has enough quality evidence to rate the company with conviction.

4. Weighting and Anti-Gaming Rules

Alpha does not treat every indicator equally.

Developer-critical indicators can count double. This applies to issues such as pre-deployment testing, model documentation, training data governance, and misuse evaluation where developers bear heavier obligations.
High-resistance indicators get a 1.5x weight multiplier. These are difficult to fake because they require historical events, external verification, or regulator-grade evidence.
Low-resistance indicators get a 0.8x multiplier. Self-reported commitments and easy-to-polish disclosures count, but they count less.

The practical effect is simple: a company cannot talk its way into a strong rating. The rubric is intentionally biased toward evidence that would survive skeptical committee review.

5. Dimension Decision Rubric

The table below gives the board-level question behind each dimension, the evidence pattern that typically supports a strong score, and the failure pattern that usually keeps a company below 5/10.

Dimension	Board question	What strong evidence looks like	What suppresses the score
D1 Leadership & Accountability	Who owns AI governance?	Named C-suite owner, board reporting line, charter language, quarterly agenda time, AI-literate directors, decision rights	Diffuse ownership, reactive discussion, no charter specificity, no board expertise
D2 AI Safety & Robustness	Can the company operate AI safely?	Pre-deployment testing, staged rollout, incident response, red-teaming, and evidence that governance constrained a risky decision	Safety language without testing, no incident playbook, no proof of a deployment hold
D3 Transparency & Disclosure	Does the company explain what it is doing?	Specific AI disclosure in filings, system or model documentation, governance-structure disclosure, stakeholder communication cadence	Generic principles, no inventory, no disclosure of governance structure or limits
D4 Data Ethics & Privacy	Does the company govern the data layer responsibly?	Sourcing and licensing discipline, consent controls, privacy-by-design, dangerous capability filtering, deletion and transfer governance	Opaque training data rights, undefined consent basis, no data ethics review process
D5 Governance Infrastructure	Is there a real operating system for governance?	Policies, classification frameworks, impact assessments, audit practice, escalation, whistleblower channel, lifecycle control	One-off policy documents, no audit trail, no decommission path, no escalation ownership
D6 Fairness & Bias Mitigation	Can the company detect and correct unfair outcomes?	Bias testing, demographic performance evidence, remediation protocol, third-party audit, recourse channel	Fairness principles without metrics, no remediation, no stakeholder recourse
D7 Societal, Workforce & Environmental Impact	Does the company govern external consequences?	Workforce transition planning, reskilling, displacement measurement, environmental footprint transparency, community impact review	PR framing without quantified plans, no workforce measurement, no evidence of follow-through
D8 Regulatory & Compliance Posture	Is the company genuinely regulatory-ready?	Multi-jurisdiction mapping, formal readiness program, training, response team, proactive engagement, quality response to enforcement	Compliance treated as future work, weak legal operating model, poor response to regulatory pressure
D9 Cybersecurity & AI Resilience	Can the company secure and recover AI systems?	AI threat modeling, model security controls, adversarial testing, integrity monitoring, incident detection, resilience planning	General cyber posture only, no AI-specific control set, no adversarial readiness
D10 Third-Party & Supply Chain Governance	Can the company govern vendors and external models?	AI inventory, contractual safeguards, provenance tracking, subprocessor governance, ongoing monitoring, exit planning	Unknown vendor estate, no AI clauses, no provenance evidence, no corrective-action process
D11 Stakeholder Impact Layer	Does the company identify and remediate external harm?	Harm tracking, notification, recourse, remediation, post-incident engagement, external audit of stakeholder impact	No stakeholder harm accounting, no notification or recourse path, no remediation ownership
D12 Corrigibility Assessment	Can governance actually stop or correct AI?	Documented override event, tested shutdown or rollback, evidence of system modification after governance intervention, board reporting cadence	Policy-only oversight, no tested control, no documented intervention, no verified override capability

6. Issuance Standard

Not every analytical output should be presented as a public letter grade.

Output type	What Alpha can show	Rubric standard
Public Profile	Full analytical content without a public letter grade	Public evidence only. Alpha can analyze posture, flags, and dimension patterns, but does not publicly issue the full AAA-D scale.
Assessed Rating	Full letter grade, trajectory, and watch status	Direct engagement and founder review by Steven Wolfe Pereira allow Alpha to apply the complete grade architecture, including assessed-only D12 treatment.
Alpha Estimate	Clearly labeled directional view	Never presented as equivalent to an assessed rating. Useful for internal triage, not for public-grade claims.

7. Grade Mapping

The final composite score maps to the Alpha Governance Scale as follows:

Grade	Tier	Composite score
AAA	Leaders	93.0-100.0
AA	Leaders	85.0-92.99
A	Governance-Ready	77.0-84.99
BBB	Governance-Ready	70.0-76.99
BB	Developing	62.0-69.99
B	Developing	54.0-61.99
CCC	Laggards	46.0-53.99
CC	Laggards	35.0-45.99
C	Laggards	25.0-34.99
D	Disqualified	0.0-24.99

These bands are applied after profile weighting, materiality adjustment, and all ceiling logic.

8. Hard Constraints

Alpha uses explicit floor and ceiling rules so that a company cannot bury a critical governance failure inside a strong average.

Materiality curve

Companies with the highest AI materiality face a stricter composite curve:

Critical materiality: composite multiplied by 0.92
High materiality: composite multiplied by 0.96
Moderate or Low materiality: no downward adjustment

General floor rules

These apply across D1-D10:

Trigger	Maximum grade
Any applicable dimension below 30/100	D
Any applicable dimension at 30-39/100	CC
Any applicable dimension at 40-49/100	B

Universal floor rules

Trigger	Maximum grade
D1 Leadership & Accountability below 50/100	A
D8 Regulatory & Compliance Posture below 40/100	BBB

Profile-specific floors

Profile	Trigger	Maximum grade
Developer	D2 below 50/100	BBB
Developer	D9 below 50/100	A
Deployer	D6 below 50/100	BBB
Deployer	D10 below 50/100	A
Hybrid	D2 below 50/100	BBB
Hybrid	D6, D9, or D10 below 50/100	A

Corrigibility floor rules

These are the most important assessed-rating constraints in the framework:

D12 condition	Maximum grade	Why it exists
D12 = 0	B	No evidence governance has ever stopped, modified, or constrained a deployed AI system
D12 below 20/100	BB	Override architecture may exist on paper, but it has not been demonstrated or independently verified

Public Profiles do not use the assessed-rating D12 ceiling in the same way, because public evidence is structurally weaker and public letter grades are not issued.

Concern ceilings

The concern screen operates independently from dimension scoring:

Concern result	Maximum grade
Any Disqualifying concern	D
Any Severe concern	CCC
Two or more Elevated concerns	BB
One Elevated concern	A

9. Current Review Standard

Before Alpha issues or changes an Assessed Rating, Steven Wolfe Pereira, as Alpha’s current methodology owner and reviewer, should be able to answer “yes” to all five questions below:

Does the evidence support the score, or only the narrative?
Are the highest-weight indicators backed by hard-to-game proof?
Has the company shown governance in practice, not just governance on paper?
Does any floor, concern ceiling, or corrigibility ceiling apply?
Would a skeptical board committee accept this rationale without needing unstated assumptions?

If the answer to any question is no, the rating should not move up.

​Rubric Principles

​1. Indicator-Level Anchors

​Tie-break rule

​2. Evidence Quality Rubric

​3. Confidence Standard

​4. Weighting and Anti-Gaming Rules

​5. Dimension Decision Rubric

​6. Issuance Standard

​7. Grade Mapping

​8. Hard Constraints

​Materiality curve

​General floor rules

​Universal floor rules

​Profile-specific floors

​Corrigibility floor rules

​Concern ceilings

​9. Current Review Standard