v0.5.0

Anti-hallucination & methodology-aware pipeline. Full design
record at docs/V0.5_PLAN.md. Motivated by v0.4.5 live tests on
papers a6182f08, cbe8048f, eea5379b, and validated end-to-end
against fresh live runs on 2026-05-20 (234a11ea, fd6bf64d) and
2026-05-21 (525fa03c) — see docs/V0.5_LIVE_VALIDATION.md.

Lane A — Pipeline

• Programmatic anti-hallucination gate before review (new file
  src/core/pipeline/verify_numbers.py, 357 lines). Scans every number
  in \begin{tabular} blocks of paper_draft.tex and matches each
  against the flat numeric values from summary_statistics.json,
  estimation_results.json, robustness_results.json, and
  figure_spec.json. Tolerance 0.5% relative; integers ≥10 must be
  exact; signs must match. Critical mismatches (relative error >10%
  vs the closest source value) → status REJECTED and reviewers
  never spawn. Persists number_verification.json at workspace root
  on every run. Live-test paper a6182f08's "log realized variance
  falls by 0.41 ($t=-3.9$)" hallucination was caught by
  technical_reviewer only after 6 reviewers had run; this gate
  catches it deterministically, at $0, before any reviewer spends a
  token. Graceful skip when no source JSON files are present (warn +
  pass), so papers from before the analyst contract was tightened
  don't regress.
• Methodology-aware phase routing. PipelineRunner.__init__ now
  accepts methodology: str = "empirical", propagated from
  papers.methodology through _run_pipeline and resume_paper in
  the API. For methodology == "theoretical",
  _reviewers_for_methodology() drops data_reviewer from the
  6-reviewer panel and _run_replication_phase() early-returns.
  Live-test paper cbe8048f burned ~$0.34 on a data_reviewer stub
  over an empty contract plus ~$0.43 on a replication packager with
  no replication artifacts — both wasted, both gone in v0.5.
• New status PaperStatus.REJECTED, distinct from FAILED.
  FAILED is reserved for crashes; REJECTED means the pipeline
  ran successfully and the quality gate (verify_numbers,
  HARD_REJECT, MECHANISM_FAIL) returned a negative verdict.
  Resumable: transitions back to IDEA / IN_PROGRESS / REVIEW /
  REVISION / CANCELLED. _run_revision_phase's HARD_REJECT and
  MECHANISM_FAIL branches updated to emit REJECTED instead of
  FAILED. New IN_PROGRESS → REJECTED transition for the
  verify_numbers gate path.
• BudgetExceededError → PAUSED, resumable. New except BudgetExceededError branch in PipelineRunner.run(), alongside
  the existing CircuitBreakerError handler. Persists state, logs a
  paused_budget event with {spent, cap}, returns a structured
  {status: "paused", reason: "budget_exhausted", ...} payload.
  The operator raises --max-cost and POSTs
  /api/papers/{id}/resume; existing resume-from-disk logic picks
  up at the first incomplete phase. Previously a budget exhaustion
  was indistinguishable from a crash.
• PAUSED and REJECTED rows now persist last_error on the
  papers table. Pre-v0.5, only FAILED and CANCELLED rows carried
  the error/reason; PAUSED and REJECTED dropped it at the SQL layer,
  leaving the dashboard with last_error=NULL and no way to render
  the budget breakdown, circuit-breaker specialist, or review-gate
  rationale. _update_status now treats PAUSED and REJECTED the
  same way as FAILED and CANCELLED for error preservation.
  Discovered while writing the v0.5 budget-pause regression test.
• POST /api/papers/{id}/resume accepts max_cost_usd in the
  request body. Pre-v0.5 the endpoint silently ignored the body and
  read the cap from the DB row, so raising the cap on a budget-paused
  paper required a manual UPDATE papers SET max_cost_usd = ...
  beforehand (the workaround surfaced during the 2026-05-20 live
  validation). The endpoint now accepts an optional ResumeRequest
  body; a positive max_cost_usd is validated and persisted on the
  row atomically with the status reset, then passed to the runner.
  Zero or negative values 400. Calls without a body preserve the
  pre-v0.5 behaviour (use the existing row value).
• paper_drafter, section_writer, abstract_writer, and
  revisor load a new writing/cite-numbers-by-source skill that
  teaches the cite-by-JSON-key discipline: every numeric value in
  the paper must trace to a value in summary_statistics.json,
  estimation_results.json, robustness_results.json, or
  figure_spec.json. HTML-comment markers (<!-- src: file#key -->)
  let verify_numbers mismatches name the exact source path the
  drafter should have used. Reduces hallucination rate in the first
  place; complements the post-hoc gate. Includes the "empty sidecar
  → no quantitative claims" rule so the design-without-estimates
  pathway is explicit.
• Test-mock fix: MockLLMBackend._detect_specialist now matches
  on the canonical You are the <Name> specialist role line in the
  system prompt rather than searching for any specialist name
  substring. The old heuristic silently misrouted calls whenever a
  skill referenced another specialist by name (e.g. the new
  writing/cite-numbers-by-source mentions "econometrics
  specialist" → paper_drafter calls were routed to the econometrics
  output → paper_draft.tex was never produced). Now matches one
  occurrence per prompt with no skill-content interference.
• Machine-readable JSON sidecar contract for verify_numbers.
  Pre-v0.5 every specialist was told to write EXACTLY ONE file, so
  even when a skill described a JSON sidecar (e.g. data/figure-spec),
  the system prompt overrode it and the JSON never appeared. The
  2026-05-20 live runs confirmed this empirically: both papers wrote
  number_verification.json with skipped_reason="no source JSON files found" — the gate was effectively a no-op. v0.5 adds a
  SPECIALIST_SIDECAR_ARTIFACTS registry, a sidecar_artifacts field
  on WorkOrder (auto-populated by _inject_context), and a
  multi-file "Required Output" prompt block that lists every required
  file with its role + JSON validity rules. data_analyst now emits
  summary_statistics.json and figure_spec.json;
  econometrics_specialist now emits estimation_results.json
  (with optional robustness_results.json). Two new schema skill
  files (data/summary-statistics-schema,
  econometrics/estimation-results-schema) teach the JSON shapes and
  the "write {} instead of omitting when data was unavailable"
  rule that distinguishes "honest empty" from "missing" for the gate.