Attack-vector market: parallel specialists pick structurally different angles each round

Two specialists of the same role kind now coordinate via a 2-step protocol
at the start of every specialist round, so they pursue genuinely different
attack vectors instead of producing near-duplicates.

Phase 1 (intent, parallel): each specialist emits a 1-shot JSON proposal
naming the angle it plans to take, with confidence and a list of vectors
it would pivot to. Phase 2 (referee, single call): the Director sees all
intents in one frame and reconciles overlaps, asking specialists to pivot
based on their own would_pivot_if list. Falls back gracefully (each
specialist gets first choice) when either phase fails.

Why this design: Anthropic's "Building effective agents" reports the
biggest multi-agent gains come from sectioning (different attack vectors)
rather than voting (same task, different samples). Wang et al.'s
self-consistency requires a votable answer, useless for free-form
derivations. Du et al.'s debate works but converges via sycophantic mode
collapse on long horizons. Method-specialization with a referee threads
the needle.

Implementation:

- src/halfseed/agents/attack_market.py — AttackIntent / AttackAssignment
  / AttackMarketResult types + collect_intents + referee_intents
  primitives, completely architecture-agnostic.

- _run_specialist_round now returns (artifacts, market_result | None).
  Market only fires when ≥2 enabled specialists share a role kind. Each
  specialist's task prompt gets the assigned vector injected verbatim.
  After every market round, _append_attack_market_note appends a HALFSEED-
  NOTE block to paper.tex so a human reading source can audit which
  vectors got assigned and why.

- New SSE event "attack_market_done" carrying {round, assigned, fell_back}.

- Validator surfaces "parallel_specialists_engaged" as INFO (not warn) so
  users see ≥2 same-kind specialists is supported, not a misconfiguration.

- paper.tex.j2: removed the legacy "Result 1, Result 2..." artifact-dump
  fallback path and the Supporting Artifacts appendix. Sections now emit
  a single \halfseedTODO[derive|data] placeholder when no Writer prose
  exists, so missing content surfaces as a TODO instead of a bookkeeping
  log. Old artifact data still available via the SQLite store and API.

- 14 new tests in tests/test_attack_market.py covering intent/referee/
  fallback/result rendering + 2 integration tests with a scripted backend
  driving the full _run_specialist_round path.

Verified live on heavyoperatorstatistics iter-7: 2 rounds, 4 specialists,
8 intent calls + 2 referee calls. Round 1 assigned distinct vectors
(perturbation expansion / RMT prediction / literature survey / ED
sampling); Round 2 EVOLVED them based on the round-1 output (saddle-point
/ replica trick). max_confidence rose from ~0.8 in earlier iterations to
0.95. Two HALFSEED-NOTE blocks landed in paper.tex with the full
attack-vector log.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: HaibaraAi137

frontend/src/api/sse.ts

@@ -28,6 +28,7 @@ const knownKinds = [
   "architecture_loaded",
   "plan_ready",
   "round_started",
+  "attack_market_done",
   "specialist_done",
   "skeptic_done",
   "revision_done",

src/halfseed/agents/attack_market.py

@@ -171,6 +171,31 @@ def validate_architecture(arch: ProjectArchitecture) -> ValidationReport:
                 ),
             )
         )
+    elif len(specialists) >= 2:
+        # Detect parallel specialists (≥2 sharing a role kind isn't a
+        # warning, it's the FEATURE — surface as an info note so users
+        # know the attack-vector market kicked in).
+        kind_counts: dict[str, int] = {}
+        for s in specialists:
+            k = resolve_kind(s)
+            kind_counts[k] = kind_counts.get(k, 0) + 1
+        for kind, count in sorted(kind_counts.items()):
+            if count < 2:
+                continue
+            issues.append(
+                ValidationIssue(
+                    level=ISSUE_LEVEL_INFO,
+                    code="parallel_specialists_engaged",
+                    message=(
+                        f"{count} agents share role kind={kind!r}. The "
+                        f"attack-vector market will run before each "
+                        f"specialist round so they pick structurally "
+                        f"different angles (different attack vectors). "
+                        f"You'll see one extra LLM call per specialist + "
+                        f"one Director referee call per round."
+                    ),
+                )
+            )
 
     # 6) If a skeptic exists, at least one specialist should have an edge to
     #    it — otherwise the revision round will be empty.

src/halfseed/architecture_validate.py

@@ -220,51 +220,20 @@
 
 \section{Setup and Assumptions}
 <% if section_bodies.setup %><< section_bodies.setup >>
-<% elif assumptions %><% for assumption in assumptions %>
-\paragraph{<< loop.index >>.} << assumption | latex_escape >>
-<% endfor %><% else %>
-\todo{No structured assumptions captured yet.}
+<% else %>
+\halfseedTODO[derive]{id="bootstrap-setup", what="Define notation and assumptions for this project.", needs="Writer prose grounded in current research material"}
 <% endif %>
 
 \section{Analytical Results}
 <% if section_bodies.analytical %><< section_bodies.analytical >>
-
-<% elif analytical_artifacts %><% for artifact in analytical_artifacts %>
-\subsection{Result << loop.index >>}
-<< artifact.claim | latex_paragraph >>
-
-<% if artifact.evidence %>\paragraph{Evidence.} << artifact.evidence | join_evidence | latex_escape >>
-
-<% endif %><% if artifact.failure_modes %>\paragraph{Failure modes.} << artifact.failure_modes | join_evidence | latex_escape >>
-
-<% endif %>\paragraph{Confidence.} << "%.2f" | format(artifact.confidence) >>
-
-<% endfor %><% else %>
-\todo{No analytical artifacts have reached the paper yet.}
+<% else %>
+\halfseedTODO[derive]{id="bootstrap-analytical", what="Write the analytical results section as integrated prose.", needs="Writer prose grounded in current research material"}
 <% endif %>
-<# When the Writer wrote the analytical section as prose it should weave
-   any necessary math inline; only fall back to dumping all artifact
-   equations en masse when there is no Writer prose to anchor them to.
-   The old behavior of always appending 16 \begin{equation*} blocks was
-   the main reason the rendered paper felt like a bullet dump. #>
-<% if not section_bodies.analytical %><% for artifact in analytical_artifacts %><% for eq in artifact.equations %>\begin{equation*}
-\halfseedeq{<< eq | latex_equation >>}
-\end{equation*}
-<% endfor %><% endfor %><% endif %>
 
 \section{Numerical Investigations}
 <% if section_bodies.numerical %><< section_bodies.numerical >>
-
-<% elif numerical_artifacts %><% for artifact in numerical_artifacts %>
-\subsection{Result << loop.index >>}
-<< artifact.claim | latex_paragraph >>
-
-<% if artifact.evidence %>\paragraph{Evidence.} << artifact.evidence | join_evidence | latex_escape >>
-
-<% endif %>\paragraph{Confidence.} << "%.2f" | format(artifact.confidence) >>
-
-<% endfor %><% else %>
-\todo{No numerical artifacts have reached the paper yet.}
+<% else %>
+\halfseedTODO[data]{id="bootstrap-numerical", what="Write the numerical investigations section as integrated prose.", needs="Writer prose grounded in current research material"}
 <% endif %>
 <# Sandbox-produced figures rendered after the numerical text. We pin them
    here (not in an appendix) so they appear close to the prose that
@@ -285,62 +254,19 @@
 
 \section{Prior Art}
 <% if section_bodies.literature %><< section_bodies.literature >>
-
-<% elif literature_artifacts %><% for artifact in literature_artifacts %>
-\paragraph{Reference << loop.index >>.} << artifact.claim | latex_paragraph >>
-<% if artifact.evidence %>\textit{Source:} << artifact.evidence | join_evidence | latex_escape >>.
-
-<% endif %>
-<% endfor %><% else %>
-\todo{Literature pass has not yet identified concrete prior work.}
+<% else %>
+\halfseedTODO[derive]{id="bootstrap-literature", what="Survey relevant prior work and place this paper in context.", needs="Writer prose grounded in current literature artifacts"}
 <% endif %>
 
 \section{Discussion}
 <% if section_bodies.discussion %><< section_bodies.discussion >>
-<% else %><< synthesis | latex_paragraph >>
+<% else %>
+\halfseedTODO[derive]{id="bootstrap-discussion", what="Discuss implications, limitations, and open questions.", needs="Writer prose grounded in current research material"}
 <% endif %>
 
-<% if open_risks %>\subsection{Open risks}
-\begin{itemize}
-<% for risk in open_risks | split_numbered %>  \item << risk | latex_escape >>
-<% endfor %>\end{itemize}
-
-<% endif %><% if suggested_next_steps %>\subsection{Suggested next steps}
-\begin{itemize}
-<% for step in suggested_next_steps | split_numbered %>  \item << step | latex_escape >>
-<% endfor %>\end{itemize}
-
-<% endif %>
 % PI-LOCK DISCUSSION
 [HALFSEED-LOCK:pi-discussion]
 
-\section*{Open Questions}
-<% if open_todos %>\begin{itemize}
-<% for todo in open_todos %>  \item << todo | latex_escape >>
-<% endfor %>\end{itemize}
-<% else %>No open \verb|\todo| items.
-<% endif %>
-
-\appendix
-\section{Supporting Artifacts}
-This appendix lists the structured research artifacts the iteration drew on,
-in case a reader wants to verify how the body of the paper was synthesized
-from internal claims.
-
-<% for artifact in analytical_artifacts %>\paragraph{A-<< loop.index >> (analytical, conf << "%.2f" | format(artifact.confidence) >>).} << artifact.claim | latex_paragraph >>
-<% if artifact.evidence %>\textit{Evidence:} << artifact.evidence | join_evidence | latex_escape >>.
-<% endif %><% if artifact.failure_modes %>\textit{Failure modes:} << artifact.failure_modes | join_evidence | latex_escape >>.
-<% endif %>
-<% endfor %>
-<% for artifact in numerical_artifacts %>\paragraph{N-<< loop.index >> (numerical, conf << "%.2f" | format(artifact.confidence) >>).} << artifact.claim | latex_paragraph >>
-<% if artifact.evidence %>\textit{Evidence:} << artifact.evidence | join_evidence | latex_escape >>.
-<% endif %>
-<% endfor %>
-<% for artifact in literature_artifacts %>\paragraph{L-<< loop.index >> (literature, conf << "%.2f" | format(artifact.confidence) >>).} << artifact.claim | latex_paragraph >>
-<% if artifact.evidence %>\textit{Source:} << artifact.evidence | join_evidence | latex_escape >>.
-<% endif %>
-<% endfor %>
-
 \bibliographystyle{plain}
 % Bibliography is generated by the citation gate; see citations.bib once enabled.