Many AI failures reflect governance and specification gaps, not model weakness: if authorized inputs treat different cases as identical, no amount of compute can recover the missing distinction. Regulators and firms should reallocate effort from pure model scaling toward changing what information is admissible, who can attest to distinctions, and question‑relative provenance and auditing.

Main Finding

The book argues that many failures attributed to insufficient intelligence are actually failures of specification and governance: when an evaluator’s licensed information treats different cases as identical, any demanded verdict that varies across those cases cannot be recovered by more computation, larger models, ensembles, or uncertainty estimates. This is a structural (identity/persistence) failure in judgment, not a purely computational one. The Identity-Persistence Program provides a diagnostic and conceptual architecture to identify where failures arise and what kinds of remedies (specification, authorized information, realization) can or cannot work.

Key Points

• Central diagnostic: ask whether a demanded verdict varies within cases that the evaluator’s licensed information treats as identical. If it does, the missing distinction cannot be reconstructed by recombining the same licensed information.
• Distinguishes separate problems often conflated in practice: specification (what distinctions matter), identity (what counts as the same case), admissibility (what transformations or inputs are allowed), evidence (what is retained/shared), derivation (how verdicts are inferred), informational repair (how missing distinctions can be supplied), realization (how a required distinction can be produced in practice), authorization (who may supply or act on missing information), commitment (institutional promises/constraints), and verification (auditing and provenance).
• Remedies differ by locus of failure:
  • If the verdict is underspecified, fixes require upstream changes to specification or governance.
  • If the licensed information literally erases a distinction, no further inference on that same information can recover it; only new/authorized information or changes to admissibility can.
  • Required distinctions can sometimes be realized only by revealing extra information (tradeoff with privacy), or by narrowing what is demanded.
• Provenance/audit requirements are relative: identical verdicts can arise through different informational paths, so provenance obligations should be tied to the specific audit question, not a universal disclosure standard.
• The book is a synthesis and translation of prior formal work (“formal owner” papers), not a new formal theorem; it orders dependencies, maps boundaries, and supplies practical diagnostics for institutional design and judgment systems.

Data & Methods

• Methodological approach: conceptual synthesis and diagnostic framework building rather than empirical experimentation.
• Uses cross-domain illustrative examples drawn from institutional records, automated decision systems, biology, software, archives, and scientific disagreement to demonstrate the program’s distinctions and diagnostics.
• Translates and orders formal results from existing technical papers into a coherent trade-oriented lens aimed at practitioners, institutions, and designers of judgment systems.
• No primary datasets or new formal theorems are introduced; emphasis is on conceptual architecture, case analysis, and prescriptive diagnostics.

Implications for AI Economics

• Limits of scale and compute: Investments in larger models, ensembles, or confidence scoring have limited value when the core problem is erased distinctions in licensed/input data. Economically, this shifts marginal returns from model scale toward investments that change information admissibility or specification.
• Value of information and markets: There is an economic case for valuing, procuring, or licensing authorized sources that supply missing distinctions (verification services, attestations, or new measurement technologies). Pricing should reflect that some information is structurally necessary and non-substitutable by inference.
• Governance and regulation design: Regulators and firms must focus on upstream specification (what distinctions a decision must respect) and on defining who is authorized to supply additional information. Audit and disclosure rules should be tied to the audit question (which distinctions matter) to avoid unnecessary provenance burdens.
• Privacy–utility tradeoffs: Realizing required distinctions may force excess revelation of sensitive data; policy and contracts must balance privacy costs against the irrecoverable value of distinctions.
• Liability and contract design: Liability rules and contracts should recognize when failures stem from governance/specification rather than model inadequacy; this affects incentives for investment in data provenance, authorized attestations, and institutional commitments.
• Organizational investment allocation: Firms should reallocate some R&D and compliance budgets from purely modeling improvements to activities that change what information is admissible, how cases are specified, and how authorized attestations are obtained and verified.
• Auditability and verification markets: Because provenance requirements are question-relative, there is scope for specialized audit services and certification markets tailored to particular stewarded distinctions, with economic implications for entrants and standard-setters.