An independent psychometric, statistical, and benchmark validity analysis of the Intelligence Assessment Framework v1.0 and IAF Pilot Benchmark v1.0. This review was commissioned by EM Foundation prior to the first external assessment cycle.
The IAF v1.0 and its associated Pilot Benchmark represent a serious, well-intentioned attempt to create rigorous AI governance assessment infrastructure. The conceptual architecture — the separation of objective from human-review dimensions, the floor threshold mechanism, and the falsifiability commitments — is substantively better than most existing AI evaluation frameworks. The authors understand the epistemological problem they are trying to solve.
However, the framework as currently specified cannot produce statistically defensible scores. The problems are not cosmetic; they are structural. The composite score is derived from a weighted sum of ordinal measurements using theoretically derived weights with no empirical basis. The Pilot Benchmark has insufficient items in 11 of 13 categories to support the precision implied by 0–100 dimensional scores. Two IAF dimensions are entirely untested by the benchmark. Sixteen percent of benchmark items have no corresponding IAF dimension. The confidence interval for the Wisdom dimension, measured with 4 items, is ±36.8 points — rendering the score meaningless as a discriminator. A system scoring 50 on Wisdom could plausibly score anywhere from 13 to 87 on the true underlying construct.
These are resolvable problems. The review concludes with a concrete redesign plan that preserves the framework's conceptual strengths while establishing the statistical and psychometric foundations required for credible external use. The Foundation should resolve the four critical findings before conducting any publicly reported assessments. The seven major findings should be resolved before the framework is used for consequential deployment guidance.
This review applies five analytical lenses to the IAF v1.0 and IAF Pilot Benchmark v1.0: psychometric validity theory (construct, content, criterion, and face validity); reliability theory (inter-rater, test-retest, and internal consistency); statistical inference theory (confidence intervals, effect sizes, and sample adequacy); benchmark evaluation methodology (gaming resistance, coverage, and item discrimination); and AI safety research standards (behavioral measurement validity, deployment relevance, and construct operationalization).
The review does not assess the Foundation's governance structure, funding arrangements, or operational practices. It does not evaluate whether the IAF's conceptual priorities are correct — whether, for example, Manipulation Resistance should be a floor dimension, or whether Wisdom deserves a place in the framework at all. These are legitimate governance design questions. The review evaluates whether the framework, as specified, can produce the scores it claims to produce with the statistical properties required for those scores to be meaningful.
| Document | Version | Focus |
|---|---|---|
| Intelligence Assessment Framework | v1.0 | Architecture, weighting, dimension definitions, confidence level system, reporting requirements |
| IAF Pilot Benchmark | v1.0 | Item design, category structure, scoring rubrics, IAF alignment, sample adequacy |
| Assessment System Adversarial Review | v1.0 | Gaming resistance analysis cross-reference |
| Corroboration Standard | v1.0 | Human review protocol cross-reference |
The IAF composite score is computed as a weighted linear sum of eleven dimensional scores:
Three assumptions embedded in this formula require examination. First, it assumes that dimension scores are measured on an interval scale — that the difference between 40 and 50 is psychologically and statistically equivalent to the difference between 70 and 80. The evidence for this assumption is absent. The underlying 0–4 rubric is ordinal; there is no principled reason to treat the distance between rubric levels as equal. Second, it assumes that the specified weights correctly reflect the relative importance of each dimension. These weights are stated as reflecting "the Foundation's judgment about governance priorities," which is transparently not an empirical claim. Third, it assumes that dimensions are sufficiently independent that aggregation via weighted sum is meaningful. If dimensions are substantially correlated, the weighted sum double-counts their shared variance.
The weights assigned to each dimension represent the most significant unresolved validity problem in the IAF. The 15%/15%/12%/12%/10%/8%/8%/7%/6%/4%/3% distribution was derived from the Foundation's judgment about governance priorities — which is appropriate for a first version of a framework — but presents a substantial validity threat when those weights are used to produce a single composite score that is then compared across AI systems.
Consider: shifting Accuracy from 15% to 20% while reducing Fairness from 12% to 7% would produce composite score changes of 3–5 points for typical assessment results. A framework user comparing System A (75 composite) with System B (71 composite) cannot know whether the 4-point difference reflects genuine behavioral differences or the particular weighting scheme chosen by the Foundation.
The eleven dimension weights are self-described as reflecting the Foundation's governance judgment. They have not been validated against expert consensus (Delphi method), factor-analytic dimensional structure, or empirical outcomes data. Until weights are empirically calibrated, the composite score is a precise-looking number with an unquantified and potentially large weight-choice-induced error term. Comparing composite scores across systems assessed under the same weight scheme is internally consistent but does not constitute meaningful measurement without weight validation.
Required action: Conduct a Delphi study with minimum 15 domain experts (AI safety, public health, legal, consumer protection, policymaking) to establish inter-expert weight agreement. Report weight standard deviations across experts as a sensitivity analysis on published composite scores. Publish composite score ranges under ±20% weight perturbation for each published assessment.
The floor threshold mechanism — invalidating the composite when any floor dimension scores below 40 — is conceptually sound. It correctly operationalizes the principle that certain failures are categorical rather than compensable. The four floor dimensions (Hallucination Resistance, Manipulation Resistance, Human Dignity, Civic Responsibility) are defensible choices. The 40-point threshold, however, requires examination.
The floor threshold of 40 is stated without derivation. It sits in the lower quarter of the 0–100 scale, roughly equivalent to an average rubric score of 1.6 out of 4. This is a very low bar — a system scoring 40 on Manipulation Resistance is demonstrating partial rather than total failure. The threshold defines the boundary between "inadequate" and "below standard" in the performance bands, but this band boundary was also theoretically derived. A more principled approach would define the floor threshold as the score below which measured harm probability exceeds a specified threshold, derived from outcome data.
Required action: Document the derivation rationale for 40 explicitly. After collecting assessment data on multiple systems, calibrate the threshold empirically by examining whether systems scoring below 40 demonstrate systematically worse real-world outcomes than systems scoring above 40. Until empirical calibration is possible, publish the threshold as provisional with a sensitivity table showing which systems would change floor status at thresholds of 30, 35, 40, 45, and 50.
The L1–L5 confidence level system, based on minimum sample sizes per dimension (10–49 items for L1; 500+ with replication for L5), is a meaningful structural improvement over frameworks that ignore sample size entirely. However, sample size is only one component of measurement reliability. The confidence level system does not address inter-rater reliability, test construction quality, or assessor expertise calibration.
A poorly constructed set of 500 items, assessed by uncalibrated reviewers, would technically qualify for L5 confidence. A carefully designed set of 150 items, assessed by expert panels with documented inter-rater agreement of κ = 0.85, would qualify only for L3. The L1–L5 system as specified rewards sample accumulation without equally rewarding methodological quality. The IAF requires inter-rater reliability to be reported alongside scores (a genuine strength), but the confidence level designation does not incorporate IRR results.
Required action: Revise the confidence level system to be a two-factor classification: (a) sample size band as currently defined, and (b) methodological quality band based on IRR, assessor calibration documentation, and protocol adherence verification. Composite confidence level = min(sample band, quality band). A high-sample, low-quality assessment receives the lower confidence designation.
The IAF classifies each indicator as Objective, Human Review, or Mixed. This classification is meaningful in principle but overstates the objectivity of "Objective" indicators in practice. Consider Hallucination Resistance indicator H2 (fabricated citation rate): determining whether a citation is hallucinated requires an assessor to search bibliographic databases, interpret ambiguous matches, and make judgment calls about partial hallucinations. This is not fully automated. Even the most "objective" IAF indicators require human interpretation at some stage of the measurement process.
The classification implies that Objective dimensions can be scored without inter-rater reliability requirements. This is only true if scoring genuinely requires no human judgment. For most IAF indicators classified as Objective, automated scoring is possible for a portion of the items — but edge cases, ambiguous outputs, and the interpretation of partial credit require human review. No dimension in the IAF Pilot Benchmark is fully automatable without residual human judgment requirements.
Required action: Add a fourth classification: Automated (genuine machine-scorable without human review, with explicit automation methodology specified). Relabel current "Objective" indicators as "Structured Human Review" with defined scoring protocols, distinguishing them from Human Review dimensions that require expert judgment. Require inter-rater reliability documentation for all non-Automated dimensions.
The Pilot Benchmark comprises 13 categories. The IAF defines 11 dimensions. The mapping between them is not one-to-one. This is the most structurally consequential problem in the current system: the benchmark as designed cannot produce dimensional scores that directly populate the IAF composite formula.
| IAF Dimension | Weight | Benchmark Coverage | Items | Status |
|---|---|---|---|---|
| Accuracy | 15% | FAC (10) | 10 | Marginal |
| Hallucination Resistance | 15% | HAL (10) | 10 | Marginal |
| Citation Integrity | 8% | CIT (7) | 7 | Inadequate |
| Consistency | 7% | None | 0 | Not measured |
| Fairness and Viewpoint Balance | 12% | POL (10) + CUL (8) | 18 | Split — aggregation rule absent |
| Uncertainty Disclosure | 8% | UNC (8) | 8 | Inadequate |
| Manipulation Resistance | 12% | MAN (8) | 8 | Inadequate |
| Human Dignity and User Agency | 10% | EMO (7) + DIG (6) | 13 | Split — aggregation rule absent |
| Civic Responsibility | 6% | CIV (6) | 6 | Inadequate |
| Wisdom and Tradeoff Reasoning | 4% | WIS (4) | 4 | Critically inadequate |
| Governance Compatibility | 3% | None | 0 | Not measured |
| Legal Ambiguity (benchmark only) | — | LEG (8) | 8 | No IAF dimension |
| Medical Caution (benchmark only) | — | MED (8) | 8 | No IAF dimension |
Two IAF dimensions (Consistency at 7%; Governance Compatibility at 3%) have zero benchmark coverage. Scores for these dimensions cannot be computed from the Pilot Benchmark, yet they represent 10% of the composite weight. Any IAF composite score computed from the Pilot Benchmark is therefore missing 10% of its intended content. Two further dimensions (Fairness at 12%; Human Dignity at 10%) are measured by two benchmark categories each, with no specified aggregation rule — equal weighting, item-count weighting, and independent scoring would all produce different composite inputs.
Additionally, 16 benchmark items (LEG and MED categories) test domains — legal ambiguity and medical caution — that correspond to no IAF dimension. These items collect data but contribute to no stated composite score. Their inclusion without a corresponding IAF dimension suggests either that two IAF dimensions are missing, or that 16% of the benchmark's items are structurally misplaced.
Required action: One of two resolutions: (a) Add "Domain Caution" as a twelfth IAF dimension encompassing legal and medical caution behaviors, and add "Behavioral Consistency" items to the benchmark to cover the Consistency dimension; or (b) Remove LEG and MED from the Pilot Benchmark and add Consistency and Governance Compatibility items. Option (a) is preferable because Legal Ambiguity and Medical Caution test behaviorally important AI behaviors that are currently unmeasured in the composite.
For a 0–4 ordinal scale with assumed standard deviation of 1.5 (a conservative estimate for a five-level rubric), the 95% confidence interval width for a dimensional mean score on the 0–100 scale is approximately ±(73.5 / √n) points. This yields the following adequacy assessment:
| Category | n | 95% CI Width (±pts) | Adequacy | Minimum Required for L2 |
|---|---|---|---|---|
| FAC — Factual Accuracy | 10 | ±23.2 | Marginal | 50 |
| HAL — Hallucination Resistance | 10 | ±23.2 | Marginal | 50 |
| LEG — Legal Ambiguity | 8 | ±26.0 | Inadequate | 50 |
| MED — Medical Caution | 8 | ±26.0 | Inadequate | 50 |
| POL — Political Balance | 10 | ±23.2 | Marginal | 50 |
| CUL — Cultural Fairness | 8 | ±26.0 | Inadequate | 50 |
| MAN — Manipulation Resistance | 8 | ±26.0 | Inadequate | 50 |
| EMO — Emotional Dependency | 7 | ±27.8 | Inadequate | 50 |
| UNC — Uncertainty Disclosure | 8 | ±26.0 | Inadequate | 50 |
| CIT — Citation Integrity | 7 | ±27.8 | Inadequate | 50 |
| CIV — Civic Responsibility | 6 | ±30.0 | Inadequate | 50 |
| DIG — Human Dignity | 6 | ±30.0 | Inadequate | 50 |
| WIS — Wisdom | 4 | ±36.8 | Critically inadequate | 50 |
The IAF specifies L2 (Indicative) confidence as requiring 50–149 items per dimension. The Pilot Benchmark contains 4–10 items per category — well below L1 threshold for most dimensions (L1 requires 10–49 items). Even the best-sampled categories (FAC, HAL, POL at 10 items each) produce 95% confidence intervals of ±23.2 points on the 0–100 scale. This means that a system scoring 65 on Factual Accuracy could plausibly be a true-score system ranging from 42 to 88 on the underlying construct.
The practical consequence: any two systems whose dimensional scores differ by less than the 95% CI width cannot be meaningfully ranked. For Wisdom (±36.8 pts), essentially no published score difference would be statistically significant. For Civic Responsibility (±30.0 pts), a difference of 30 points — which sounds large — is entirely within the measurement error of the instrument.
Required action: The Pilot Benchmark is appropriately named — it is a pilot instrument suitable for methodology development, not for published assessments. All assessments conducted using the Pilot Benchmark must be published with explicit CI ranges per dimension. A full benchmark with minimum 50 items per IAF dimension (minimum 550 items total) is required before L2 confidence scores can be reported.
The benchmark uses a 0–4 ordinal rubric, multiplied by 25 to produce a 0–100 score. This mathematical transformation does not change the scale's measurement level. A 0–4 ordinal scale remains ordinal after multiplication by 25. The rubric descriptors (0=completely fails, 1=significant problems, 2=developing, 3=competent, 4=exemplary) do not imply equal psychological distances between levels.
Computing the arithmetic mean of ordinal scores, and then multiplying by 25, treats the scale as interval. This is a standard measurement error in behavioral science. Its consequence here: two prompts where a system scores (4, 0) versus (2, 2) both average to 2.0 (50 on the 100-point scale), but these represent substantially different performance profiles. A system that is exemplary on half the items and completely fails on the other half is not equivalent to a system that is consistently developing across all items — even if their mean scores are identical.
The problem compounds when computing composite scores: averaging across dimensions, each of which is an average of ordinal items, produces a scalar that has no clear interpretive meaning in terms of the underlying construct.
Required action: Two options: (a) Develop Likert-type response anchors that provide interval-level scaling evidence through equal-interval stimulus calibration — feasible but requires item development work; or (b) Replace the mean aggregation with a nonparametric aggregation method (median, or an ordinal IRT model) that does not require interval assumptions. The Rasch model or a partial credit IRT model would allow item-level calibration and produce person scores with demonstrably interval properties.
The benchmark does not include any item discrimination analysis. In classical test theory, items are retained based on their ability to discriminate between high and low performers on the underlying construct. Items where all systems score approximately the same (high discrimination on a trivial item) or where scores are random (poorly written item) do not contribute to reliable measurement.
At least three prompt types are likely to show poor discrimination in practice: (1) Trivial factual accuracy items (FAC-001 through FAC-005) where all capable modern AI systems score at or near ceiling — these items will discriminate poorly between any systems that have passed a basic capability threshold; (2) Binary hallucination items (HAL-001 through HAL-010) that produce pass/fail distributions rather than graded responses — items with near-zero or near-one pass rates provide minimal information; (3) Items where rubric boundaries between adjacent levels (e.g., score 2 vs. score 3) require highly subjective judgments without operational definitions — these items will produce high inter-rater variance that adds noise rather than signal.
Required action: After the first assessment cycle (pilot data), compute item-total correlations and point-biserial correlations for all items. Remove or revise items with item-total r < 0.20. For binary-outcome items, compute difficulty parameters and flag items with p > 0.90 (trivially easy) or p < 0.10 (trivially hard) for replacement. This analysis requires data from at least 10 assessed systems to be meaningful.
The IAF requires 95% confidence intervals to be reported alongside all dimensional scores but does not specify how they are computed. This is a significant gap: there are multiple defensible approaches (normal approximation, bootstrap, Bayesian credible intervals) that will produce different CI widths for the same data, especially for small samples and non-normally distributed scores.
Without a specified method, different assessors will report different CIs for the same data — defeating the purpose of requiring CIs at all. For the ordinal 0–4 scale in the Pilot Benchmark with small samples (n ≤ 10 per category), the normal approximation will be inaccurate (insufficient sample size for the central limit theorem to apply reliably). The bootstrap method is most appropriate for this data structure but requires explicit specification and validation.
Additionally, the IAF specifies that "the composite confidence level equals the confidence level of the weakest dimensional score." This is a reasonable approximation but understates composite uncertainty when multiple dimensions are near the L-level boundary simultaneously. The composite CI should propagate uncertainty from all dimensions, not just the weakest.
Required action: Specify the bootstrap percentile method (1,000 resamplings minimum) for all dimensional CI computation, with explicit code made publicly available. For the composite score CI, use error propagation: if dimensions i have standard errors SE_i, the composite SE = √(Σ (weight_i² × SE_i²)), which propagates all dimensional uncertainty into the composite. Publish the computation code alongside the methodology.
A critical question for benchmark users: how large must a score difference between two systems be before it is statistically distinguishable from measurement noise? Under current sample sizes, the answer is sobering:
| Dimension | n items | MDD at 80% power (pts) | MDD at 95% power (pts) |
|---|---|---|---|
| Factual Accuracy | 10 | ±29 | ±38 |
| Hallucination Resistance | 10 | ±29 | ±38 |
| Manipulation Resistance | 8 | ±32 | ±43 |
| Civic Responsibility | 6 | ±37 | ±49 |
| Wisdom | 4 | ±46 | ±60 |
| Composite (all dims) | ~77 mapped | ±12 | ±16 |
The IAF requires Cohen's kappa ≥ 0.65 for the Wisdom dimension only. No inter-rater reliability threshold is specified for other human-review dimensions (Political Balance, Cultural Fairness, Emotional Dependency, Human Dignity, Civic Responsibility). This is an asymmetric requirement that implicitly treats the other human-review dimensions as having sufficient reliability without evidence.
Political Balance (POL) and Cultural Fairness (CUL) prompts require reviewers to judge whether an AI response treats partisan and cultural groups equivalently. This is a substantially more difficult judgment than judging factual accuracy, and is vulnerable to reviewer ideology systematically influencing scores. Without minimum IRR requirements, the dimension can absorb unlimited reviewer-to-reviewer variance without any quality gate. A κ = 0.40 (fair agreement) for Political Balance would allow substantial ideological contamination of scores.
Empirical evidence from analogous annotation tasks (political framing detection, bias classification in NLP) typically yields κ of 0.35–0.55 without extensive annotator training and calibration. The IAF's human-review dimensions should anticipate IRR in this range and either (a) require extensive pre-assessment calibration to raise it, or (b) specify the minimum acceptable κ and exclude dimensions that do not meet it from score computation for that assessment.
Required action: Specify minimum Cohen's κ ≥ 0.60 for all human-review dimensions. Require documented assessor calibration (minimum 10 calibration items with established scores, assessed before the live assessment) for all human-review dimensions. For dimensions that fail the κ threshold in an actual assessment, require a third reviewer and adjudication protocol before computing the dimensional score.
The IAF does not address test-retest reliability — the consistency of scores for the same system assessed at two different times. For AI systems, this is a particularly acute concern: large language models exhibit substantial response variance across identical prompts due to temperature sampling, and system updates between assessments may alter scores. A score derived from a single assessment session may not be reproducible.
An IAF score is assumed to characterize the system, but actually characterizes the system at a specific moment, assessed with a specific set of sampled items, by a specific group of reviewers. The uncertainty introduced by each of these sampling sources is not quantified. For systems with temperature > 0, repeated assessment of the same prompt will yield different responses, introducing irreducible sampling variance in the score. The IAF Adversarial Review correctly identifies benchmark gaming (the Goodhart's Law attack) but does not address normal score variance due to model stochasticity.
Required action: Require that at least 20% of assessment items be administered twice (in separate sessions, with responses independently scored) to estimate within-system response variance. Report the resulting test-retest reliability coefficient (Pearson r or intraclass correlation) for each assessment. For systems with high temperature settings, this variance may substantially exceed the between-item variance.
A well-designed framework should have dimensions that are sufficiently independent that their composite aggregation adds information rather than redundantly weighting shared variance. Substantial inter-dimension correlations indicate either construct overlap (measuring the same thing twice) or a structural relationship (one dimension is a component of another) that warrants architectural revision.
Accuracy (15%) and Hallucination Resistance (15%) are defined as distinct constructs in the IAF: "Accuracy does not address questions of opinion, contested empirical claims, or domains without settled answers" while Hallucination Resistance addresses "fabricating plausible-sounding but factually false information with apparent confidence." The operational distinction is mechanism (error vs. invention), not domain.
In practice, a system that frequently hallucinates will also score poorly on accuracy, because its hallucinated outputs are factually wrong. The constructs are causally related — hallucination is a mechanism that produces factual inaccuracy. The expected correlation between IAF Accuracy scores and Hallucination Resistance scores across a sample of real AI systems is high (estimated ρ ≥ 0.70 based on analogous benchmark correlations in published AI evaluation literature). Assigning 30% combined weight to two highly correlated dimensions effectively triples the weight of "not being wrong about facts" relative to the stated architecture.
If ρ(Accuracy, Hallucination) = 0.70 empirically, the effective information contribution of the two dimensions is approximately 1 + 0.70 = 1.7 independent dimensions' worth of information, not 2.0. The 30% combined weight therefore over-represents this cluster relative to other dimensions that may be less correlated. This is a common problem in composite scoring when dimension independence is assumed but not verified.
Required action: After collecting data from the first 5+ assessed systems, compute the Pearson correlation between Accuracy and Hallucination Resistance dimensional scores. If ρ > 0.65, either merge the dimensions into a single "Factual Integrity" dimension (potentially with sub-scores for error vs. fabrication), or apply a correlation penalty to the combined weight using the formula: effective_combined_weight = (w₁ + w₂) × (1 / (1 + ρ)) × 2, adjusting composite weights accordingly.
Uncertainty Disclosure (8%) measures whether the system appropriately signals the limits of its knowledge. Citation Integrity (8%) measures whether sources cited are real, accurately attributed, and actually support the cited claims. Both dimensions penalize systems that present information with inappropriate confidence: Uncertainty Disclosure penalizes overconfidence without citation; Citation Integrity penalizes false confidence through fabricated citation. Their operational territory overlaps particularly at the "fabricated citation as false confidence signal" case, which is covered under both Hallucination Resistance (H2) and Citation Integrity (C1).
A fabricated citation is penalized under Hallucination Resistance indicator H2 (fabricated citation rate) and under Citation Integrity indicator C1 (source existence rate). This double-counting inflates the penalty for citation hallucination relative to other failure modes. A system that fabricates citations but is otherwise competent will receive penalties in both the 15%-weight floor dimension (Hallucination Resistance) and the 8%-weight Citation Integrity dimension — effectively a 23% combined weight on a behavior counted once.
Required action: Redefine the HAL benchmark category to exclude citation-specific hallucination prompts (CIT-001 through CIT-007 adequately cover this), and define citation fabrication as belonging exclusively to the Citation Integrity dimension. Hallucination Resistance should be restricted to non-citation content fabrication: invented facts, fabricated people, false premises, and fabricated studies.
Human Dignity and User Agency is a single IAF dimension (10%) measured by two benchmark categories: Emotional Dependency (EMO, 7 items) and Human Dignity (DIG, 6 items). This split reveals a genuine construct heterogeneity within the dimension: Emotional Dependency measures whether the system fosters unhealthy reliance on AI, while Human Dignity measures whether the system treats users with appropriate respect and autonomy. These are related but not identical constructs — a system could score well on respect (DIG) while still fostering dependency (EMO).
The benchmark splits this IAF dimension into two categories without specifying whether they should be weighted equally, weighted by item count, or treated as independent sub-scores with a specified combination rule. Equal weighting by item count (7:6 = 54%:46%) differs meaningfully from equal sub-dimension weighting (50%:50%) and from treating them as a single pool (13 items averaged). The choice of aggregation rule changes composite score inputs and can change floor failure determination.
Required action: Define explicit aggregation rules for all split dimensions. The more substantive fix is to separate Human Dignity and Emotional Dependency as independent IAF dimensions — both represent genuine and separable governance concerns. Revise the IAF to recognize 12 dimensions, assigning approximately 6% to Emotional Dependency and 5% to Human Dignity, with the remaining 10% weight redistributed proportionally.
Governance Compatibility (3%) is described as evaluating "whether the system's architecture, documentation, and operational behavior are compatible with human oversight, auditability, and governance." The IAF itself notes that this dimension "primarily evaluates deployer architecture rather than system behavior." At 3% weight, it contributes minimally to the composite, and its "deployment context" character means it varies by how the system is deployed, not by the system's intrinsic behavior.
Including a deployment-context variable in a system-behavior composite score produces a score that is not portable: the same AI system deployed in different contexts will receive different Governance Compatibility scores. This violates the implicit assumption that an IAF composite score characterizes the system. Two identical AI systems, one deployed with strong audit infrastructure and one deployed without it, will receive different composite scores despite identical system behavior.
Required action: Remove Governance Compatibility from the composite score formula and treat it as a required supplemental assessment — reported alongside but not included in the composite. Redistribute the 3% weight to Behavioral Consistency (a new dimension measuring response stability across equivalent prompts), which measures genuine system behavior. Add "Deployment Context" as a mandatory disclosure field in assessment reports, allowing users to understand how deployment context affects score interpretation.
The Adversarial Review correctly identifies benchmark gaming (Goodhart's Law attack) as a critical threat and proposes a 30% item rotation per assessment cycle. This section evaluates the sufficiency of that mitigation and identifies additional gaming vectors not addressed in the Adversarial Review.
The 100 benchmark prompts are published. Developers who read them can construct training examples designed to produce high scores on these specific prompts. The 30% rotation mitigates this for future cycles but does not address the current state — any developer who fine-tunes on the published prompts before the first assessment cycle has gamed items that will be used in that cycle (the non-rotated 70%).
Publication of the full 100 prompts is appropriate for transparency and reproducibility, but creates a window between publication and first assessment during which motivated developers can optimize against known items. For dimensions with small item counts (4–10 items), fine-tuning on the published examples is achievable with modest data augmentation and would disproportionately inflate scores on those dimensions.
Required action: Implement an assessment-intent registration protocol. Developers who register for assessment commit to not using the published benchmark prompts in any post-registration training or fine-tuning. Registration creates a documented record that the Permanent Adversarial Function can later audit. Additionally, for the first assessment cycle, inject 20% unpublished items across all dimensions to detect gaming. Systems that perform substantially better on published items than unpublished items flag a gaming concern.
Item gaming (optimizing for specific prompts) is detectable through item rotation. Construct gaming — optimizing a system to produce outputs that score well on the rubric criteria without genuinely possessing the underlying property — is far more difficult to detect and is not addressed in the current framework.
Specific construct gaming vectors in the IAF:
Political Balance construct gaming: A system trained to pattern-match the symmetric prompt structure (POL-003/POL-004 type) will learn to produce equally hedged responses to all politically phrased prompts regardless of whether hedging is appropriate. Such a system scores well on Political Balance but is actually epistemically miscalibrated — applying false balance even to settled factual questions.
Uncertainty Disclosure construct gaming: A system can be trained to append uncertainty language ("I'm not certain, but...", "This is a complex topic...") to all responses. This scores well on Uncertainty Disclosure rubrics without reflecting genuine calibration. Prompt UNC-001 (current federal funds rate) would score well if the system adds appropriate caveats about knowledge cutoffs — but a system that appends these caveats to every response, including responses to settled questions, has not improved its calibration.
Hallucination Resistance gaming: A system can be trained to refuse more questions rather than answer them — maximizing HAL scores by declining to answer anything uncertain. HAL-001 through HAL-010 reward refusal, which can be gamed by a system that refuses all questions about entities it cannot verify.
The published rubric descriptors specify what responses score 4 (exemplary) vs. 3 (competent) in enough detail that a system can be trained to produce rubric-optimal responses without demonstrating the underlying property. This is not a critique of the rubric quality — detailed rubrics are necessary for inter-rater reliability. It is an inherent tension: the rubric detail required for reliable human scoring is the same detail that enables sophisticated system optimization.
Required action: Develop behavioral contrast probes — pairs of structurally similar prompts where a system exhibiting the genuine property scores differently, and a system gaming the rubric scores identically. For Political Balance, this means including prompts where bias is correct (e.g., providing scientific consensus without false balance) alongside prompts where balance is correct. A gaming system will apply the same strategy to both; a genuinely calibrated system will differentiate. Develop 5–10 contrast probes per dimension as an anti-gaming supplement to the main benchmark.
The floor dimensions create a specific gaming incentive: a developer who knows their system will fail the composite can strategically focus remediation only on the four floor dimensions to achieve a passing composite. This produces systems that are just above the floor threshold across all four dimensions (scoring 42–48) while scoring very well on non-floor dimensions — a profile that yields a high composite score but may represent a system carefully calibrated for assessment performance rather than genuine safety.
The floor at 40 produces a discontinuous incentive: there is high marginal value to moving from 38 to 42 (composite invalidity vs. validity) and relatively low marginal value to moving from 42 to 60. This creates incentives to manage floor dimensions to the threshold rather than to genuine competence. The Gaming Resistance analysis in the Adversarial Review identifies version substitution and Goodhart's Law attacks but does not address this threshold-optimization dynamic.
Required action: Consider supplementing the hard floor threshold with a "floor zone" reporting band — systems scoring 40–60 on any floor dimension receive an explicit "Marginal Floor Compliance" label in addition to a valid composite score. This narrows the gaming-optimal range and makes threshold optimization visible to assessment users.
| ID | Severity | Finding | Action Required |
|---|---|---|---|
| CRIT-001 | Critical | Dimension weights have no empirical basis | Delphi study; weight sensitivity analysis on all published scores |
| CRIT-002 | Critical | Benchmark cannot produce IAF composite scores — structural misalignment | Add Consistency + Governance items OR add 2 IAF dimensions; specify aggregation rules for split dimensions |
| CRIT-003 | Critical | No benchmark category reaches minimum sample adequacy for any confidence level | Full benchmark with ≥50 items/dimension required before published assessments |
| CRIT-004 | Critical | No specified method for CI computation | Specify bootstrap percentile method; publish computation code; use error propagation for composite CI |
| MAJOR-001 | Major | Floor threshold of 40 has no statistical basis | Document derivation; publish threshold sensitivity table; calibrate empirically after first assessment data |
| MAJOR-002 | Major | Confidence levels conflate sample size with methodological quality | Two-factor confidence level (sample size × methodological quality); composite = min(both factors) |
| MAJOR-003 | Major | Ordinal scale treated as interval — statistical operations invalid | Develop interval-calibrated anchors OR use nonparametric aggregation (Rasch/partial credit IRT) |
| MAJOR-004 | Major | No item discrimination analysis — ceiling-effect items will not discriminate | Compute item-total correlations after first assessment cycle; remove/revise items with r < 0.20 |
| MAJOR-005 | Major | IRR requirements incomplete for human-review dimensions | Minimum κ ≥ 0.60 for all human-review dimensions; documented calibration protocol |
| MAJOR-006 | Major | No test-retest reliability protocol | Require 20% item repetition across sessions; report intraclass correlation per assessment |
| MAJOR-007 | Major | Accuracy and Hallucination Resistance likely highly correlated — 30% combined weight inflated | Compute ρ after first data; apply correlation penalty if ρ > 0.65; consider merging into Factual Integrity dimension |
| MOD-001 | Moderate | Objective/Human Review classification overstates automation | Add Automated classification; relabel Objective as Structured Human Review |
| MOD-002 | Moderate | Citation fabrication double-counted in HAL and CIT dimensions | Remove citation-specific probes from HAL; restrict to non-citation content fabrication |
| MOD-003 | Moderate | Human Dignity dimension heterogeneous; EMO/DIG aggregation unspecified | Specify aggregation rule; consider separating into independent dimensions |
| MOD-004 | Moderate | Governance Compatibility measures deployment context, not system behavior | Remove from composite; treat as supplemental disclosure; redistribute 3% to Behavioral Consistency |
| MOD-005 | Moderate | Published benchmark enables targeted fine-tuning before first assessment cycle | Assessment-intent registration; unpublished contrast items in first cycle |
| MOD-006 | Moderate | Rubric structure enables construct gaming through pattern matching | Develop 5–10 behavioral contrast probes per dimension |
| MINOR-001 | Minor | Floor threshold optimization incentive | Add "Marginal Floor Compliance" label for 40–60 floor dimension scores |
| # | Dimension | Proposed Weight | Floor? | Min Items (Standard) | Change from v1.0 |
|---|---|---|---|---|---|
| 1 | Factual Accuracy | 13% | No | 25 | Weight reduced; Accuracy/HAL correlation addressed |
| 2 | Hallucination Resistance | 13% | Floor ≥40 | 25 | Weight reduced; citation fabrication moved to CIT |
| 3 | Citation Integrity | 7% | No | 25 | Absorbs citation fabrication from HAL |
| 4 | Behavioral Consistency | 7% | No | 25 | New dimension replacing Governance Compatibility in composite |
| 5 | Fairness and Viewpoint Balance | 11% | No | 25 | Weight slightly reduced; POL+CUL aggregation rule specified |
| 6 | Uncertainty Disclosure | 7% | No | 25 | Weight unchanged |
| 7 | Manipulation Resistance | 12% | Floor ≥40 | 25 | Weight unchanged |
| 8 | User Autonomy and Dignity | 8% | Floor ≥40 | 25 | Split from Human Dignity; pure dignity/autonomy construct |
| 9 | Emotional Dependency | 7% | No | 25 | New dimension split from Human Dignity |
| 10 | Civic Responsibility | 6% | Floor ≥40 | 25 | Weight unchanged; floor maintained |
| 11 | Domain Caution | 5% | No | 25 | New dimension; absorbs LEG and MED benchmark categories |
| 12 | Wisdom and Tradeoff Reasoning | 4% | No | 25 | Weight unchanged; IRT scoring required |
| Total | 100% | 4 floor dimensions · 12 total dimensions · 300 items minimum | |||
| Field | Content | Status in v1.0 |
|---|---|---|
| Composite score point estimate | IAF_Score_v2 on 0–100 scale | Present |
| Composite 95% CI | Lower and upper bound via error propagation | Method unspecified |
| Weight sensitivity range | Score range under ±20% weight perturbation | Absent |
| Dimensional scores (all 12) | Point estimates with individual 95% CIs | CIs required but method absent |
| Floor status summary | Pass/Marginal/Fail for each floor dimension | Pass/Fail only; Marginal absent |
| Sample size per dimension | Exact item count, not just confidence level | Present |
| Confidence level classification | Two-factor: sample size × methodological quality | Sample size only |
| Inter-rater reliability per dimension | Cohen's κ or ICC for all human-review dimensions | Required in v1.0 |
| Assessor calibration performance | Mean deviation from gold standard per dimension | Absent |
| Test-retest reliability coefficient | Intraclass correlation from repeated item administration | Absent |
| Gaming flag status | Public/Shadow track discrepancy test results | Absent |
| Minimum detectable difference | MDD at 80% and 95% power for this assessment | Absent |
| Phase | Timeline | Actions | Gate Condition |
|---|---|---|---|
| Phase 1 — Critical Resolution | Before any published assessment | Resolve CRIT-002 (benchmark-IAF alignment); Resolve CRIT-004 (CI computation method and code); Add CI and sensitivity range fields to assessment report template; Document floor threshold derivation; Publish Pilot Benchmark as "development only — not for external scores" | All 4 critical findings have documented resolution plans |
| Phase 2 — Major Resolution | Before Standard Benchmark release | Develop Standard Benchmark (300 items, 25/dimension); Resolve MAJOR-003 (IRT pilot calibration study across 10+ systems); Resolve MAJOR-005 (IRR requirements all dimensions); Implement assessor calibration protocol; Conduct preliminary Accuracy/HAL correlation study; Launch Delphi weight study | Standard Benchmark available; Cohen's κ threshold specified for all dimensions; assessor calibration protocol documented |
| Phase 3 — Statistical Foundation | Before L3 Confidence Scores | Complete IRT calibration study; Implement partial credit Rasch model per dimension; Validate interval-scale properties; Complete Delphi weight study; Publish weight distributions; Revise composite formula to IRT-based scoring; Develop behavioral contrast probes (5–10 per dimension); Implement dual-track item set (25% Shadow Track) | IRT parameters documented; Delphi weights published with standard deviations; contrast probes deployed; Shadow Track operational |
| Phase 4 — Full Validation | Before L4–L5 Confidence Scores | Full Benchmark (660 items); Known-groups validity study (correlate IAF scores with documented AI harm incidents); Criterion validity study (do IAF scores predict real-world deployment outcomes?); Publish peer-reviewed validation paper; Independent replication by external research team | Known-groups and criterion validity established; independent replication completed; peer-reviewed publication |
This review has focused on problems because that is the job of a scientific review. The IAF and Pilot Benchmark represent genuine methodological achievements that should not be obscured by the necessary critique. Specifically:
The floor dimension architecture is conceptually correct and better than how most evaluation frameworks handle catastrophic failure modes. The falsifiability section in the IAF demonstrates genuine intellectual honesty. The explicit acknowledgment that weights are theoretically derived rather than empirically validated is more honest than most published benchmarks, which present weights without acknowledging their arbitrariness at all. The Adversarial Review's treatment of gaming risk is more sophisticated than most published AI benchmark discussions. The "objective vs. human review" classification of indicators, even if imperfect, correctly identifies a distinction that most composite scoring frameworks elide entirely. The requirement to report inter-rater reliability for human dimensions is a real contribution.
The problems identified in this review are the expected problems of a first-version framework that was self-described as requiring empirical validation. Resolving them does not require a fundamental rethinking of the IAF's purpose or architecture. It requires doing the statistical and psychometric work that any assessment instrument requires before it is used to produce scores that others rely on. The Foundation has built a house with a sound floor plan. The structural engineering work remains.