docs: refresh paper with construction manifest implementation

Author: jroell

info/main.pdf

@@ -118,7 +118,7 @@ \section{Auditing Faithfulness}
 \section{Reference Implementation}
 The reference implementation is written in TypeScript and is publicly available at \url{https://github.com/jroell/agentic-epistemology-framework}. It implements beliefs, justifications, frames, agents, memory, actions, tools, and observer events. The implementation now includes two benchmark harnesses: a generic root-cause benchmark and a facet-frame faithfulness benchmark.
 
-The Vurvey production codebase informed the facet extension. Its PM facet graph includes global facets and versioned facet values, `headerPrompt` and `detailPrompt` fields used to build persona prompts, `PmPersonaFragment` rows that link personas to selected facet values, mold facet configurations with weights and options, and auto-configuration services that recommend and weight facet distributions for population molds. These implementation details motivate the audit requirement: a persona response should be diagnosable back to the facet graph, prompt fragments, selected values, and composed frame that produced it.
+The Vurvey production codebase informed the facet extension. Its PM facet graph includes global facets and versioned facet values, \texttt{headerPrompt} and \texttt{detailPrompt} fields used to build persona prompts, \texttt{PmPersonaFragment} rows that link personas to selected facet values, mold facet configurations with weights and options, and auto-configuration services that recommend and weight facet distributions for population molds. The current implementation extends that path in two places: \texttt{vurvey-python} writes an AEF construction manifest into \texttt{ai\_personas.facetConfigs} when personas are generated from molds, and \texttt{vurvey-api} attaches an AEF audit summary to each simulated survey answer during phase-two generation. These implementation details motivate the audit requirement: a persona response should be diagnosable back to the facet graph, prompt fragments, selected values, construction manifest, composed frame, rendered system prompt, user prompt, and generated response that produced it.
 
 \section{Automated Root-Cause Benchmark}
 The generic benchmark asks whether structured epistemic traces make it possible to identify the cause of agent failures in scenarios where the ground-truth cause is known. It covers four injected causes: frame misweighting, source-trust error, threshold miscalibration, and justification gap. It evaluates four trace conditions: full trace, no-frame trace, no-justification trace, and event-only trace.
@@ -189,13 +189,13 @@ \section{Facet-Frame Faithfulness Benchmark}
 \label{fig:facet-frame-faithfulness}
 \end{figure}
 
-A representative facet-misweighting case illustrates the diagnostic path. The persona was expected to evaluate a premium AI subscription while constrained by crisis-flow financial outlook and a time-poverty facet. The trace showed `time\_poverty` with observed weight 0.18 against expected weight 0.78. The response gave a long, careful evaluation despite the time constraint. The diagnostic procedure predicted `facet\_misweighting` after inspecting the facet contribution event.
+A representative facet-misweighting case illustrates the diagnostic path. The persona was expected to evaluate a premium AI subscription while constrained by crisis-flow financial outlook and a time-poverty facet. The trace showed \texttt{time\_poverty} with observed weight 0.18 against expected weight 0.78. The response gave a long, careful evaluation despite the time constraint. The diagnostic procedure predicted \texttt{facet\_misweighting} after inspecting the facet contribution event.
 
 
 \section{Facet Faithfulness Audit Service Experiment}
 The benchmark above motivated a concrete implementation: \texttt{FacetFaithfulnessAuditService}. The service takes a persona response, selected facets, rendered prompt provenance, model metadata, self-reported confidence, and optional counterfactual probes. It returns a facet influence ledger, marker coverage, contradictions, counterfactual sensitivity, calibration score, final faithfulness score, primary diagnosis, causal facets, evidence, and recommended debug actions.
 
-The service implements the audit hooks proposed by the framework: selected facet IDs and values, taxonomy versions, prompt hashes, configured and runtime weights, frame composition, suppressed facets, inferred interaction terms, expected markers, observed markers, forbidden marker contradictions, confidence calibration, and counterfactual sensitivity. A Vurvey adapter maps PM facet values, facet codes, \texttt{headerPrompt}, \texttt{detailPrompt}, confidence scores, taxonomy versions, and configured weights into this audit input shape.
+The service implements the audit hooks proposed by the framework: selected facet IDs and values, taxonomy versions, prompt hashes, configured and runtime weights, frame composition, suppressed facets, inferred interaction terms, expected markers, observed markers, forbidden marker contradictions, confidence calibration, and counterfactual sensitivity. A Vurvey adapter maps PM facet values, facet codes, \texttt{headerPrompt}, \texttt{detailPrompt}, confidence scores, taxonomy versions, construction-manifest weights, and rendered prompt provenance into this audit input shape.
 
 We evaluated the audit service on eight hand-constructed Vurvey-style persona cases covering faithful responses, prompt omission, facet misweighting, facet conflict, persona drift, response-style violation, and low counterfactual sensitivity. The goal of this experiment is not to prove external predictive validity. It tests whether the implemented service can recover failure causes from increasingly rich evidence conditions.
 
@@ -249,7 +249,7 @@ \section{Live Vurvey Counterfactual Generation}
 
 The qualitative outputs were informative. One staging persona responded: ``A monthly cost is already a stretch for anything extra, and if it takes time to set up? Forget about it.'' After the \texttt{disposable\_income} facet was removed, the generated response still referenced money and family constraints, indicating that similar constraints are redundantly represented in other facets or in the generated biography. Another persona with high disposable income treated the product as plausible if it saved time, while the counterfactual version became more value-conditional. These outcomes show that facets do influence the live generation path, but effects are distributed across multiple identity layers.
 
-The main finding is that real counterfactual sensitivity is measurable, but marker coverage is currently the limiting factor. The audit service correctly avoided overclaiming faithfulness: all three live cases were diagnosed as \texttt{facet\_selection\_gap} because the expected marker catalog did not cover enough of the real facet semantics and because removing one facet did not remove all related behavioral constraints from the biography and adjacent facets. This is a useful failure mode. It shows that the next product improvement should be richer facet-level marker metadata and interaction-aware expectations, not merely better prompt inspection.
+The main finding is that real counterfactual sensitivity is measurable, but marker coverage is currently the limiting factor. The audit service correctly avoided overclaiming faithfulness: all three live cases were diagnosed as \texttt{facet\_selection\_gap} because the expected marker catalog did not cover enough of the real facet semantics and because removing one facet did not remove all related behavioral constraints from the biography and adjacent facets. This is a useful failure mode. It shows that the product improvement should be richer facet-level marker metadata, interaction-aware expectations, and persisted construction manifests. The implementation now records such manifests at persona initialization time and uses them at survey-simulation runtime, closing the main traceability gap identified by the live counterfactual experiment.
 
 \section{Discussion}
 The generic benchmark shows that structured epistemic traces can recover simple injected failures when the trace contains the relevant fields. The facet benchmark is more realistic and harder. Full traces recover many failures, but not all. This is useful: persona faithfulness cannot be reduced to checking that a prompt contains the right words. It requires tracing facet selection, prompt rendering, frame composition, expected behavioral markers, response content, and sometimes counterfactual or latent evidence.
@@ -270,7 +270,7 @@ \section{Limitations}
 The framework also depends on the quality of facet definitions, prompt fragments, weights, and constraints. If the facet graph itself encodes weak theory or biased assumptions, \method{} will make those assumptions more visible but will not make them correct. Detailed observer logs can expose sensitive persona or respondent data, so production use requires careful retention, access control, and redaction policy.
 
 \section{Code and Data Availability}
-The reference implementation is available at \url{https://github.com/jroell/agentic-epistemology-framework} under the MIT license. The implementation includes \texttt{FacetFaithfulnessAuditService}, the generic root-cause benchmark, the facet-frame benchmark, and the facet audit service experiment. Generated outputs are written under the repository's results directory. A Vurvey API branch named \texttt{AEF-implementation} contains the product-facing adapter that maps PM facet values into the audit input shape, the read-only staging/production fixture extractor, and the live Vurvey counterfactual generation runner.
+The reference implementation is available at \url{https://github.com/jroell/agentic-epistemology-framework} under the MIT license. The implementation includes \texttt{FacetFaithfulnessAuditService}, the generic root-cause benchmark, the facet-frame benchmark, the facet audit service experiment, and public benchmark documentation. Generated outputs are written under the repository's results directory. A Vurvey API branch named \texttt{AEF-implementation} contains the product-facing adapter that maps PM facet values and construction-manifest weights into the audit input shape, attaches AEF audit summaries during survey simulation, and includes the live Vurvey counterfactual generation runner. A Vurvey Python branch named \texttt{AEF-implementation} persists AEF construction manifests into \texttt{ai\_personas.facetConfigs} during mold-based persona generation.
 
 \section{Conclusion}
 The Agentic Epistemology Framework provides an audit architecture for autonomous agents and synthetic personas. By making beliefs, confidence, justifications, frames, facets, and observer events first-class constructs, \method{} supports debugging not only what an agent said, but why the system should have expected that behavior from its configured epistemic state. The facet extension connects this audit layer to synthetic human simulation: facets become inspectable contributors to frame composition, and persona faithfulness becomes a measurable, diagnosable property rather than a matter of surface plausibility.