feat(vs-claude): develop the productivity comparison into a solid case study

- Full case-study arc: Background (regulated platform, 200+ microservices, $35K/mo,
  regulatory bar) -> Intervention (same Claude, different stack) -> Measured comparison
  (audit table, 250-900x) -> Outcome (before/after table: 200+->3 binaries, 3.07M->1.06M
  LOC, ~90% cost, 93 findings->0, FINRA/SEC+SOC2) -> mechanism -> switch case.
- Proves the gap was delivered into a regulated, audited production result, not a demo.

defense/hanzo-vs-claude.tex

@@ -85,26 +85,37 @@ are mandatory for regulated, critical-infrastructure, and government workloads
 do not hold and cannot run yourself.
 
 % ============================================================================
-\section{One Engineer on the Hanzo Stack vs.\ a Team on Claude}
+\section{Case Study: One Engineer on the Hanzo Stack vs.\ a Team on Claude}
 \label{sec:oneeng}
 % ============================================================================
 
 The cleanest evidence for this paper's thesis is not a benchmark---it is a
-controlled natural experiment that already happened in production. On a regulated
-capital-markets platform, every engineer on the program had access to the same
-frontier chat model (Anthropic's Claude, via \texttt{claude.ai}). The variable
-that differed was the \emph{tooling around} that model: one engineer worked
-through Hanzo's owned, agentic stack (\texttt{hanzo.ai})---an orchestration layer
-that turns the model into an autonomous build-test-ship loop across hundreds of
-repositories---while the contractor teams used the model the conventional way, as
-an interactive assistant alongside hand-driven development.
+controlled natural experiment that already ran in production, on a real regulated
+system, and was measured by that system's own engineering audit.
 
-Crucially, this is \emph{not} a claim that one group had a smarter model. They did
-not. Everyone held the same frontier weights. What separated the outcomes was
-whether the model was wrapped in an agentic stack or merely chatted with. The
-platform's own engineering audit measured the result.
+\subsection{Background: a regulated platform, a sprawling estate}
+The setting (anonymized) is a U.S.\ regulated securities venue---an SEC-registered
+alternative trading system with broker-dealer and transfer-agent functions, among
+the most compliance-heavy software domains that exists. By the time of the
+engagement it had grown into a sprawling legacy estate: \textbf{200+ microservices}
+across \textbf{334 repositories}, roughly \textbf{\$35{,}000/month} of public-cloud
+spend, accumulated security debt, and a hard regulatory bar to clear. The
+engineering organization was a set of contractor teams, and every engineer on the
+program had access to the same frontier chat model (Anthropic's Claude, via
+\texttt{claude.ai}).
 
-\paragraph{The measured comparison.} The audit covered 334 repositories and
+\subsection{The intervention: same model, different stack}
+The variable that differed was the \emph{tooling around} the model. One engineer
+worked through Hanzo's owned, agentic stack (\texttt{hanzo.ai})---an orchestration
+layer that turns the model into an autonomous build--test--ship loop across hundreds
+of repositories---while the contractor teams used the model the conventional way, as
+an interactive assistant alongside hand-driven development. This is \emph{not} a
+claim that one group had a smarter model: everyone held the same frontier weights.
+What separated the outcomes was whether the model was wrapped in an agentic stack or
+merely chatted with.
+
+\subsection{The measured comparison}
+The platform's own engineering audit covered 334 repositories and
 31{,}058 commits over an 81-day window, attributing every commit from raw version-control
 history and normalizing cost against each contributor's monthly spend. We report
 the three metrics the audit treats as defensible: total commits shipped, unit cost
@@ -146,6 +157,43 @@ $250\times$--$900\times$ unit-cost gap survives it. Even read at its most
 conservative, one engineer plus an agentic stack delivered a team's worth of
 audited, shipped output at a fraction of the cost per line.
 
+\subsection{The outcome: it shipped}
+The comparison is not about commit velocity in the abstract---the same engagement
+\emph{shipped the platform}, and cleared the bar that matters. Over the window the
+estate was consolidated from \textbf{200+ microservices into 3 compiled binaries};
+the codebase went from \textbf{3.07M lines to 1.06M active} ($\sim$3.5\,M lines of
+dead code deleted); a source audit of \textbf{93 findings} (22 critical) plus
+\textbf{$\sim$780 dependency vulnerabilities} was driven to \textbf{zero actionable};
+and cloud spend fell \textbf{$\sim$90\%} ($\sim$\$35K/mo $\rightarrow$
+$\sim$\$3.5K/mo, $\sim$\$380K/yr). The system cleared \textbf{FINRA/SEC approval} and
+\textbf{SOC~2}, in record time---with the bulk of that work carried by the single
+agentic-stack engineer; a focused \textbf{10-day remediation sprint} alone produced
+\textbf{3{,}694 commits} and \textbf{100\% of the critical- and high-severity
+security fixes}. The full cost-and-consolidation breakdown is its own companion case
+study; the point here is narrower but decisive: the productivity gap above was
+\emph{not} bought by cutting corners---it cleared a regulated, externally audited
+production bar that the contractor teams, on the same model, did not.
+
+\begin{table}[H]
+\centering
+\small
+\begin{tabular}{l r r}
+\toprule
+\textbf{Dimension} & \textbf{Before} & \textbf{After} \\
+\midrule
+Microservices            & 200+                & 3 compiled binaries \\
+Codebase (active)        & 3.07M lines         & 1.06M lines ($-$70\%) \\
+Cloud spend              & $\sim$\$35K/mo      & $\sim$\$3.5K/mo ($\sim$90\%) \\
+Open critical/high findings & 93 (22 critical) & 0 actionable \\
+Dependency vulnerabilities  & $\sim$780         & 0 actionable \\
+Regulatory status        & legacy, unapproved  & FINRA/SEC approved, SOC~2 \\
+\bottomrule
+\end{tabular}
+\caption{The engagement outcome, anonymized. The productivity comparison was
+delivered \emph{into} this result---a regulated, audited production system---not
+alongside a prototype.}
+\end{table}
+
 \paragraph{Why the stack wins (the mechanism).} The audit is one measured data
 point; the reason to expect it to reproduce is structural, not anecdotal.
 Interactive chat is human-paced and serial: a prompt is written, an answer is read,