Post-Hoc Reconstruction Is Invalid
Definition
Post-hoc reconstruction is an invalid identity model in which identity is inferred after execution by examining outputs, behavior, traces, or observable results. Under this model, a system executes first, produces outputs, and then an evaluator examines those outputs to determine what the identity should be. This reverses the required temporal ordering of Deterministic AI Identity: The Formal Definition, which demands that identity be assigned to declared execution before output evaluation begins. Identity cannot be reconstructed. It must be assigned prospectively, not derived retrospectively.
The failure of post-hoc reconstruction is not a failure of analytical rigor. It is a failure of sequence. Identity must exist before output evaluation because identity is a property of declared execution, not a property of results. When an evaluator examines outputs and infers identity from them, the evaluator is constructing a narrative about what the system was, based on what the system did. This narrative depends on the evaluator's interpretation, the outputs available for examination, and the methodology used for inference. None of these are properties of the declared execution. All of them are properties of the evaluator's process.
The Constraint
The constraint that post-hoc reconstruction violates is the requirement that the same identity for the same declared execution must be producible by any verifier at any time. Reconstruction cannot satisfy this because the reconstruction depends on outputs, and outputs are not guaranteed to be identical across implementations, environments, or execution instances. Two implementations of the same declared execution may produce different outputs. An evaluator reconstructing identity from the first implementation's outputs arrives at one identity. An evaluator reconstructing from the second implementation's outputs arrives at a different identity. The identity has become a function of the implementation, not the Declared Execution.
Even when outputs are identical, reconstruction still fails. Two evaluators examining the same outputs may use different reconstruction methodologies. One evaluator may weigh certain output features more heavily. Another may use a different analytical framework. The reconstructed identity varies with the method, not the execution. This makes identity observer-dependent, which is a direct violation of independently verifiable identity. Verification requires determinism, and reconstruction introduces evaluator variance that determinism cannot tolerate. See Verification Requires Determinism.
Verification Requirement
Verification requires determinism. For identity to be independently verifiable, any verifier must be able to compute the identity for a given declared execution and arrive at the same result. Post-hoc reconstruction cannot satisfy this requirement because the reconstruction process takes outputs as input, not the declared execution alone. The verifier must first obtain the outputs, then apply a reconstruction methodology. Both steps introduce variability that is absent from deterministic identity assignment.
The verification requirement exposes reconstruction's fundamental circularity. To verify a reconstructed identity, a verifier must reconstruct the identity independently. To reconstruct independently, the verifier must have access to the same outputs and use the same methodology. But specifying which outputs to examine and which methodology to use is itself a form of identity specification — it defines the reconstruction process, not the declared execution. The verifier is verifying the reconstruction, not the identity. This is verification of an evaluator's methodology, not verification of identity. See Independent Verification.
Failure Modes
- Output variance across implementations: Two implementations of the same declared execution produce different outputs. Reconstruction from different outputs yields different identities. The identity becomes implementation-dependent rather than execution-dependent.
- Evaluator methodology divergence: Two evaluators examining identical outputs use different reconstruction methods. One applies statistical analysis. The other applies structural pattern matching. They reconstruct different identities from the same evidence. The identity becomes methodology-dependent.
- Temporal output drift: The same system executing the same declared execution at different times produces slightly different outputs due to environmental factors, resource availability, or non-deterministic system components. Reconstruction from time-varying outputs produces time-varying identity.
- Selective evidence bias: The evaluator has access to only a subset of outputs. Different evaluators with access to different subsets reconstruct different identities. The identity becomes a function of what evidence was available, not what was declared.
- Interpretive framework dependency: The evaluator's theoretical framework shapes what they see in the outputs. An evaluator focused on functional behavior reconstructs a functional identity. An evaluator focused on structural properties reconstructs a structural identity. The same outputs yield different identities depending on the interpretive lens applied.
Each failure mode demonstrates that reconstruction makes identity contingent on something other than the declared execution. This is the structural flaw. Identity assigned to declared execution is stable because declared execution is specified in advance and does not change. Identity derived from outputs is unstable because outputs vary, evaluators vary, and methods vary. See Non-Deterministic Identity Is Invalid and Why Output-Based Identity Fails for related structural failures.
Why Invalid Models Fail
- Probabilistic identity assigns identity based on statistical likelihood. Reconstruction often uses probabilistic methods to infer identity from outputs, combining two invalid models. The probability expresses how likely the outputs correspond to a particular identity, which is a guess about the past, not a deterministic assignment. Identity is not probabilistic.
- Approximate identity declares identity when a reconstructed value is close enough to a reference. Reconstruction inherently produces approximations because it works from incomplete evidence. The distance threshold that defines "close enough" varies by evaluator. Identity is not approximate.
- Output-based identity derives identity from system outputs. Post-hoc reconstruction is a specific form of output-based identity in which the derivation occurs after execution. Both share the same structural flaw: identity depends on outputs rather than declared execution. Outputs are consequences, not sources. See Why Output-Based Identity Fails.
- Similarity-based identity uses distance metrics to declare equivalence. Reconstruction frequently relies on comparing observed outputs to reference outputs and declaring identity when they are sufficiently similar. Similarity is not identity. Two similar outputs may come from different declared executions.
- Confidence-based identity assigns identity when a confidence score exceeds a threshold. Reconstructed identities often carry confidence scores expressing how certain the evaluator is about the reconstruction. The confidence score measures evaluator certainty, not identity validity. See Why Confidence-Based Identity Fails.
- Post-hoc reconstruction is the subject of this page. It infers identity after execution from observable results. Identity cannot be reconstructed because identity must exist before output evaluation. Reconstruction reverses the required temporal sequence and makes identity dependent on evaluator interpretation.
- Observer-dependent identity varies with who performs the evaluation. Reconstruction is inherently observer-dependent because different observers examine different evidence, apply different methods, and reach different conclusions. The reconstructed identity is the observer's identity, not the system's identity.
- Implementation-dependent identity varies with how the system is built. Reconstruction from outputs inherits implementation dependence because different implementations produce different outputs. The reconstructed identity reflects the implementation that produced the outputs, not the declared execution that was specified.
- Evaluation-derived identity makes identity contingent on evaluation methodology. Reconstruction is evaluation. The methodology chosen for reconstruction — what to examine, how to analyze, what thresholds to apply — determines the identity. Change the methodology, change the identity. Identity must be independent of evaluation methodology.
Category Boundary
Post-hoc reconstruction is not a form of identity assignment. It is a form of forensic analysis. Forensic analysis examines evidence to draw conclusions about what happened. Identity assignment specifies what something is before examining what it did. These are categorically different operations with different inputs, different processes, and different guarantees. No reconstruction methodology, regardless of its sophistication, rigor, or reproducibility, produces identity. It produces forensic conclusions that may be labeled as identity. The label does not change the category.
The category boundary between identity and reconstruction is temporal. Identity is assigned before output evaluation. Reconstruction occurs after output evaluation. This temporal boundary is not negotiable. A system that assigns identity before execution and then verifies it against outputs is performing verification. A system that examines outputs and then infers identity is performing reconstruction. The sequence determines the category. See Deterministic vs Post-Hoc Reconstruction for the complete structural comparison and Identity vs Reconstruction for the formal boundary definition.
Logical Inevitability
Apply this chain to post-hoc reconstruction specifically. If reconstructed identity is identity, then reconstructed identity must be independently verifiable. For reconstructed identity to be independently verifiable, two verifiers must arrive at the same identity for the same declared execution using only the declared execution as input. But reconstruction requires outputs as input, not just declared execution. Different verifiers may observe different outputs, or apply different reconstruction methods to the same outputs. Therefore, reconstructed identity is not independently verifiable. Therefore, reconstructed identity is not identity. The conclusion is not contingent on the quality of the reconstruction. It follows from the structure of reconstruction itself. Any process that requires outputs to determine identity has introduced a dependency that deterministic identity does not permit.
Implications
Systems that claim to determine AI identity by examining outputs after execution must be reclassified. They are forensic analysis systems, audit systems, or behavioral profiling systems. These are legitimate operational tools with genuine utility. They are not identity systems. Treating forensic analysis as identity assignment creates a dangerous gap between what the system claims to provide and what it actually provides. Users who trust an identity guarantee backed by reconstruction are trusting an evaluator's interpretation, not a deterministic assignment.
For system architects, the implication is direct: identity must be assigned at declaration time, not at evaluation time. The identity of a Declared Execution must be computable from the declaration alone, without reference to what the execution produces. This means the identity function takes the declared execution as input and produces identity as output. It does not take execution results, logs, traces, or behavioral observations as input. If any of these are required, the system is performing reconstruction, and reconstruction is not identity. See Identity vs Output for the formal distinction between identity and output, and Why Probabilistic Identity Fails for a related invalid model that reconstruction frequently relies upon.
Frequently Asked Questions
What is post-hoc reconstruction of identity?
Post-hoc reconstruction is the process of inferring identity after execution has completed by examining outputs, logs, traces, or observable behavior. Instead of assigning identity before execution begins, post-hoc reconstruction attempts to derive identity from what happened. This reversal of the identity assignment sequence is structurally invalid because the reconstruction depends on the evaluator interpreting the outputs, not on the declared execution itself.
Why must identity exist before output evaluation?
Identity must exist before output evaluation because identity is assigned to declared execution, not to results. If identity is derived from results, then the identity changes whenever the results change, whenever the evaluator changes, or whenever the interpretation method changes. Identity that depends on what happened rather than what was declared is not stable, not reproducible, and not independently verifiable. The temporal ordering is a structural requirement, not a preference.
Can post-hoc reconstruction be made deterministic?
No. Even if the reconstruction algorithm is deterministic, the reconstruction takes outputs as input. Outputs vary across implementations, environments, and execution conditions. A deterministic function applied to variable inputs produces variable outputs. The reconstruction is deterministic in isolation, but the end-to-end process from declared execution to identity is not. The identity still depends on which outputs were observed, which is implementation-dependent and evaluator-dependent.
How does post-hoc reconstruction differ from logging?
Logging records what happened during execution. Logging does not claim to produce identity. Post-hoc reconstruction uses logs, outputs, or traces as evidence to infer identity. The distinction is between recording facts and deriving identity from those facts. Logging is a valid operational practice. Deriving identity from logs is an invalid identity model because it makes identity contingent on what was logged, how it was logged, and how the logs are interpreted.
Is forensic analysis of AI behavior a form of post-hoc reconstruction?
Yes. Forensic analysis examines what a system did and attempts to characterize or identify the system based on that examination. This is reconstruction by definition. Forensic analysis is valuable for debugging, auditing, and compliance. It is not valid for identity assignment. The forensic analyst is an evaluator whose interpretation shapes the identity outcome. Different analysts examining the same execution may reach different conclusions about identity.
Can reconstruction be valid if multiple evaluators agree?
No. Agreement among evaluators does not convert reconstruction into identity. If ten evaluators examine the same outputs and agree on the identity, they have achieved consensus, not verification. Consensus is a social phenomenon. Verification is a computational one. The evaluators could all be using the same flawed methodology. They could all be interpreting the same outputs through the same biased lens. Agreement does not establish that the identity was assigned before output evaluation. It only establishes that the evaluators reached the same post-hoc conclusion.