v0.3.0

Added

Los Alamos Exploration Layer (optional, auto-triggered addon)

• Hypothesis tree/DAG materialized view (scripts/los_alamos/hypothesis_tree.py),
  event-sourced from an append-only message bus (bus.py); four evaluation
  tracks: Track 0 (semantic screening), Track R "Bletchley" (red-team stress
  test), Track 1 (entropy-weight + TOPSIS objective ranking), Track 2
  (qualitative pairwise judge panel, Copeland ranking, Shannon-entropy
  disagreement)
• Six autonomous trigger criteria (SKILL.md path C Step 0), including a new
  "problem-complexity initial judgment" checkable as early as
  problem_analysis — no longer requires an explicit human request or
  literature-survey-stage evidence to fire
• New LOS_ALAMOS_METHOD_CATALOG.md — six-category MCM paradigm taxonomy
  (optimization / prediction / evaluation / simulation / classification /
  network) used by Alsos literature surveys and Director's first-layer branch
  selection to widen candidate coverage beyond ad-hoc search
• Elastic branch width (2 branches under time pressure, 3–4 when time budget
  allows) and deeper branching support (same-paradigm technical-choice
  sub-branches, depth + --parent-ids)
• Governance rule (SOP §9.2.6): branch decisions must query
  hypothesis_tree.py status before acting, not rely on narrative memory —
  wired into SKILL.md at Step 5 (build/verify dispatch) and Step 9 (finals)
• Fixed: adjudicate.py screen/redteam previously left the
  hypothesis_tree_{N}.json materialized-view snapshot stale until the next
  hypothesis_tree.py invocation; both now call _sync_tree() immediately

RAND Delphi — Track 2 anonymous multi-round convergence

• adjudicate.py panel-vote --round / panel-entropy --round /
  delphi-summary; capped at 2 retry rounds (SOP §9.6)

Project Skunk Works — lightweight mode

• pipeline_manager.py init --skunk-works, relaxes citation gate (≥1 vs ≥2)
  and Checkpoint report verbosity; §4 self-verification mandate is explicitly
  not relaxed
• Autonomous trigger extended (SOP §10.1): Director may self-trigger based on
  a time-pressure signal alone (not only explicit human request), but must
  announce the switch before executing — never silent

Andon Cord — universal emergency stop

• pipeline_manager.py andon-pull/andon-clear/andon-status; freezes advance
  pipeline-wide until cleared; AP mode cannot auto-clear (one of only two
  hard human-in-the-loop exceptions, alongside Checkpoint LA)

NASA Mission Control Go/No-Go — final_compile pre-flight check

• quality_gate.py launch-check (gate 8): polls Andon / stage completeness
  (now filtered by problem_count, see Fixed) / image existence / placeholder
  residue / anonymity / Los Alamos assumption-ledger integrity
• Mandatory before final_compile (new absolute-prohibition item, SOP §7)

Project Kaizen — post-verification quality-polish loop

• pipeline_manager.py kaizen-assess/kaizen-round-start/kaizen-status;
  4-dimension self-assessment (significance / robustness / assumptions /
  completeness, threshold 4.0/5), capped at 2 rounds, gated by the same
  objective time-budget signal as Skunk Works (opposite direction: adds
  rigor when time allows rather than relaxing it under pressure)
• Hard boundary vs. writing-style polish (§16): Kaizen may revise numeric
  conclusions and model content; style passes may not

Work Log — single-file complete session record

• scripts/worklog.py — append-only, Simplified Chinese,
  CUMCM_Workspace/memory/worklog.md; stage transitions / rework / Andon /
  Kaizen / gate results auto-logged from pipeline_manager.py and
  quality_gate.py at zero extra token cost; user messages and undocumented
  Agent decisions logged manually via worklog.py append

Citation Authenticity Verification & Shared Pool

• scripts/cite_check.py — register (dedup by DOI/URL/title into a shared
  citations.bib across sub-problems) / verify (DOI via CrossRef API, URL
  via HEAD request — catches hallucinated citations, not just format-shape
  matching) / list / export-bibitems
• quality_gate.py lit gate now prioritizes the verified citation pool over
  the legacy prose-regex scan; legacy path still works with a migration nudge
• Fixed: export-bibitems did not escape LaTeX special characters — a real
  DOI containing an underscore broke xelatex compilation with a fatal error

Writing-Style / "AI-Generated Feel" Reduction

• scripts/style_check.py scan — six objective, non-LLM checks: list-density
  in narrative sections, stock-phrase repetition, mechanical
  first/second/finally transition skeletons, sentence-length coefficient of
  variation, paragraph-length CV, cross-section n-gram repetition; thresholds
  informed by public research (mrshibly/Humanizer, Wikipedia:Signs of AI
  Writing, Pangram) and recalibrated against real Chinese academic writing
• SOP §16 writing protocol: plan-then-segment writing, ≤2 style-polish rounds;
  hard boundary — style polish may only change how something is said, never
  numeric conclusions (Kaizen's territory)
• Fixed: LaTeX comment lines (%% ==== section divider ====) were leaking
  into n-gram frequency analysis as false "repeated phrases"
• Fixed: overlapping sliding-window n-gram duplicates of the same longer
  phrase (e.g. three 6-char windows of one 8-char term) now merge into one
  finding instead of three

Official Format Compliance (SOP §17)

• CUMCM: templates/latex_template.tex synced to the 2026 format spec —
  electronic-version-compliant cover-page structure (title+abstract combined
  as page 1, page numbering starts there), \tableofcontents removed (official
  rule: no TOC), commented-out print-only 承诺书/编号页 block
• New anonymity heuristic gate quality_gate.py anon-check (self-referential
  / labeled-field phrase matching), wired into launch-check
• New scripts/ai_usage_doc.py: generate (builds a standalone
  《AI工具使用详情》PDF from worklog.md per the 2025-trial AI-tool-usage
  regulation's four required sections), cite-format (CUMCM reference-line
  formatter), mcm-entry (COMAP Report on Use of AI Tools entry formatter)
  — none of these auto-insert into the paper; declaring AI usage is a
  deliberate, manual decision by design
• MCM/ICM: templates/mcm_template.tex gets \setcounter{tocdepth}{2}
  (keeps the required TOC to ~1 page) and a commented-out
  Report on Use of AI Tools placeholder section (COMAP format, appended
  after the 25-page solution, not counted toward the limit)
• compile_pdf.py now reports page counts after a successful compile
  (PDF-bookmark-based body/appendix split) with contest-specific soft targets
  and hard-cap warnings — CUMCM body target 20–25 pages / hard cap 30
  (appendix unlimited); MCM/ICM hard cap 25 pages total including appendix
  code — and recommends compacting appendix core code first when over budget

§18 — Research domain-conventional visualization forms before drawing

• New AutoMCM_SOP.md §18 ("画图前先查领域惯例"): before designing any result
  figure, search (web-search, keywords abstracted per S3) for the
  conventional visualization form for that problem archetype — e.g.
  trajectory/geometry-optimization problems → 3D trajectory + key-moment
  annotation; coverage/siting/allocation → map/scatter + Gantt-style timeline;
  multi-objective/multi-scenario comparison → Pareto front or radar chart;
  sensitivity analysis → tornado chart or multi-parameter curve family;
  classification/clustering/ranking → heatmap or a re-sorted stacked bar
  chart. Findings recorded in thought_process.md; falling back to
  common-sense judgment is permitted when no clear convention is found, but
  — same spirit as §7's "never silently skip a failed check" — that fallback
  must also be recorded, never silently taken.
• Wired into .claude/skills/auto-mcm/SKILL.md's 图表风格规范 section (new
  subsection preceding 标准用法) and cross-referenced across all three lean
  runtime bindings (.opencode/, .dsh/, .agents/).
• Codex binding's existing "known capability gap: no native web search"
  section extended to explicitly cover this new requirement too — the same
  underlying gap (no WebSearch/WebFetch tool) blocks the domain-convention
  research step, not just literature review; the existing "disclose at
  Checkpoint① rather than silently skip" guidance now applies to both.

Two new fully-autonomous demo cases (replacing the previous single showcase)

• demo/2020C/ — real 2020 CUMCM Problem C (Credit Strategy for Small and
  Micro Enterprises), official attachments (~1.1M invoice rows) downloaded
  and verified from mcm.edu.cn. Full dsh headless run (zero human
  checkpoints) through all three sub-problems + sensitivity analysis:
  logistic-regression default-probability model (5-fold CV AUC 0.935),
  entropy-weight/TOPSIS cross-validation track, PSI transfer-stability check,
  KMeans industry-behavior clustering, expected-profit-maximizing credit
  allocation. 30 pages (21-page body), 18 figures, all 2D, zero 3D plots.
• demo/2025A/ — 2025 CUMCM Problem A (smoke-screen decoy deployment),
  restyled end-to-end into a NASA technical-report / orbital-mechanics
  chart theme (white background, dense fine grid + bold major gridlines,
  black/white primary lines, NASA Blue #0B3D91 / NASA Red #FC3D21
  reserved for key-annotation accents only, fully boxed axes) — implemented
  as a workspace-local override of scripts/plot_style.py's palette/rcParams/
  save() grid post-processing, not a change to the repository's shared
  style module. 28 pages (20-page body), same modeling conclusions as before
  restyling.
• demo/mind_reader_ui.png — the Mind-Reader tool screenshot (previously
  nested inside the single demo's CUMCM_Workspace/) relocated to a
  case-independent top-level path, since it illustrates the tool, not either
  specific paper.

Runtime Bindings

• New: DeepSeek Harness (dsh) binding — .dsh/skills/auto-mcm/SKILL.md
• New: opencode binding — .opencode/skills/auto-mcm/SKILL.md
• New: Codex CLI binding — .agents/skills/auto-mcm/SKILL.md
• All three verified via real executed tasks (not documentation inference),
  cross-checked through structured runtime session logs (opencode export,
  decompressed .zstd dsh session logs, plain-.jsonl Codex rollouts)

Fixed

• templates/latex_template.tex and templates/mcm_template.tex both
  failed to compile out of the box — both loaded \usepackage{natbib}
  while their bibliography section used plain \bibitem/\cite (not
  natbib's \citep/\citet), which is a fatal incompatibility
  (Package natbib Error: Bibliography not compatible with author-year citations). Neither template body actually uses any natbib-specific
  command, so natbib was unnecessary; removed from both. Verified with a
  from-scratch standalone xelatex compile of each template (CUMCM: exit 0,
  6 pages, zero errors once a placeholder figure exists; MCM/ICM: confirmed
  the natbib error is gone, remaining failure is the separate, already-
  documented mcmthesis.cls external-package requirement). This bug predates
  this release — every fresh template compile would have hit it — and had
  only ever been patched ad hoc inside individual generated papers'
  main.tex, never at the template source; this is the first fix at the
  actual source. README prerequisites also updated to document the
  mcmthesis package requirement for the MCM/ICM template specifically
  (most minimal TeX Live installs don't ship it; CUMCM doesn't need it).
• plot_style.py: correlation-matrix heatmap used a one-directional
  sequential colormap for signed (−1…1) data instead of diverging_cmap() —
  negative and positive correlation were nearly indistinguishable at a
  glance; fixed.
• Two instances of a "single-x-tick stacked bar chart degenerates into a
  solid-color rectangle with no visible structure" bug, both triggered when
  an expected-profit-maximizing credit allocation concentrated almost
  entirely into one rating tier (2020C demo, problem1_credit.py and
  problem2_credit.py) — replaced with per-enterprise sorted horizontal bar
  charts (amount + PD annotated per bar), which is both correct rendering
  and more informative than the aggregate single bar it replaced.
• LaTeX verification-summary table (2025A demo main.tex) overflowed the
  page's right margin by ~415pt (~5.7in) — a plain lcl tabular with long
  cell text and no wrap points; fixed with array package p{}-width
  wrapped columns, plus \allowbreak inserted after underscores in
  monospace filenames (which otherwise have no natural line-break point and
  still overflow their column even when wrapping is enabled).
• 3D scene-overview and trajectory figures (2025A demo) had large dead
  whitespace bands where set_box_aspect'd axes were never tightened to
  their actual rendered content — added plot_style.fit3d_axes() (two-pass
  render + get_tightbbox()-based repositioning) and wired it into the
  affected 00_data_eda.py/problem1_geometry.py figure functions; also
  replaced two 3D-space text labels that sat close together in this
  scenario's real geometry (and so collided under certain view angles) with
  projection-immune screen-space text2D anchors.
• Documented, and deliberately reverted rather than shipped: a third 3D
  inset panel + taller figure canvas intermittently caused
  bbox_inches='tight''s save-time crop to drop unrelated existing labels
  from a sibling subplot on the same figure — reproduced twice independently
  (including after reducing label offsets), suspected to be an mplot3d
  Text3D bounding-box quirk rather than a usage error in this codebase;
  reverted to the previously-verified two-panel layout instead of risking a
  silent regression for a marginal enhancement.
• gate_launch_check() stage-completeness check now filters by
  problem_count — previously always scanned model_1–model_3 regardless
  of actual sub-problem count, producing a false permanent NO-GO for any
  1- or 2-problem contest
• Hardcoded /tmp/ absolute paths in quality_gate.py example commands
  replaced with in-workspace paths — broke under sandboxed runtimes'
  external_directory permission checks

Testing

• Every bug above (degenerate allocation charts, dead 3D whitespace, table
  overflow, wrong colormap) was found by direct human/Claude visual
  inspection of the rendered PDF pages — none were caught by the dsh agent's
  own completion report, and none could have been: the model dsh was running
  (deepseek-v4-flash) has no image-input support, so its claimed "visual
  self-review" was necessarily only ever a file/pixel-level programmatic
  check, disclosed honestly in worklog.md rather than misrepresented.
• Multiple full real-agent pipeline runs (dsh, opencode) exercising the
  complete new addon set together, independently re-verified rather than
  trusting agent self-reports (PDF validity/page counts, launch-check
  re-run, cite_check verify re-run, session-log decompression for
  tool-call audit)
• Los Alamos genuinely self-triggered via multiple distinct criteria across
  runs (explicit "compare methods" problem wording, autonomous
  complexity judgment with no such wording); one run surfaced two genuine
  Checkpoint LA two-track conflicts (one directional contradiction, one
  Condorcet cycle) — resolved with fully documented headless-degradation
  reasoning; flagged as a known test-methodology limitation that Checkpoint
  LA's true human-wait path remains unverified in a genuinely interactive
  session