v2.1.0 — Currency + citability: Fable 5 refresh, doc-currency fixes, /submission-disclosures, CITATION.cff

.claude/references/model-versions.md

@@ -1,23 +1,26 @@
-<!-- CURRENT: Opus 4.8 | Sonnet 4.6 | Haiku 4.5 -->
+<!-- CURRENT: Fable 5 | Opus 4.8 | Sonnet 4.6 | Haiku 4.5 -->
 
 # Current Model Versions (single source of truth)
 
-**Last verified against Anthropic docs:** 2026-05-31
+**Last verified against Anthropic docs:** 2026-06-10
 
 This file is the **one place** that names current Claude model point versions. Everything else in the template should either refer to tiers abstractly ("newest Opus", "the Haiku tier") or point here. `scripts/check-model-versions.sh` flags any **superseded** version that is presented as **current** in the template's user-facing surfaces.
 
 The machine-readable `<!-- CURRENT: ... -->` marker at the top is parsed by the checker — keep it in sync with the table.
 
 | Tier | Current version | Model ID | Notes |
 |------|-----------------|----------|-------|
-| Opus (high-judgment) | **Opus 4.8** | `claude-opus-4-8` | newest; API default; GA 2026-05-28; $5/$25 per MTok; 1M context; defaults to `high` effort |
+| Fable (Mythos-class; hardest, long-horizon) | **Fable 5** | `claude-fable-5` (alias `fable`; 1M variant `claude-fable-5[1m]`) | most capable model in Claude Code; **opt-in** (`/model fable` or the `best` alias) — NOT the default on any account type; GA 2026-06-09; $10/$50 per MTok; 1M context (128k max output); defaults to `high`; requires Claude Code ≥ 2.1.170; falls back to Opus 4.8 on flagged cyber/bio content |
+| Opus (high-judgment) | **Opus 4.8** | `claude-opus-4-8` | current Opus tier; **API/account default**; GA 2026-05-28; $5/$25 per MTok; 1M context; defaults to `high` effort; the content-fallback target for Fable 5 |
 | Sonnet (workhorse) | **Sonnet 4.6** | `claude-sonnet-4-6` | 1M context |
 | Haiku (fast / mechanical) | **Haiku 4.5** | `claude-haiku-4-5-20251001` | fast tier; ID is the snapshot-pinned alias |
 
 **Retiring 2026-06-15:** Sonnet 4 + original Opus 4 → migrate to Sonnet 4.6 / Opus 4.8.
 
 **Fast mode:** Opus 4.8 fast mode is $10/$50 per MTok (~2.5× speed). Opus 4.7 fast mode is $30/$150; Opus 4.6 fast mode is deprecated (~2026-06).
 
+**Fable 5 maturity caveat (2026-06-10):** routing guidance lives in [`model-routing.md`](../rules/model-routing.md) — as of launch week, Fable 5 is deliberately **not** routed to the forked-reviewer fleet (2× the Opus price on the judgment tier, and day-one forced-tool-protocol reliability was observed to lag Opus 4.8: 28/28 structured-output subagent failures in one session's workflow vs 0 on Opus). Re-evaluate as point releases land.
+
 ## Prior generations
 
 It is fine to mention older versions in **historical** contexts (CHANGELOG entries) or in explicit **"prior generation" / comparison** lines (e.g. "Opus 4.8's `high` does what Opus 4.7's `xhigh` did"). They must **not** be presented as the current / newest / default model.
@@ -30,4 +33,5 @@ The checker allows a line to mention an older version when it carries a marker s
 
 1. Update the table **and** the `<!-- CURRENT: ... -->` marker above, plus the "Last verified" date.
 2. Run `./scripts/check-model-versions.sh` and fix every current-state surface it flags.
-3. Add a "Changed — model refresh" entry to `CHANGELOG.md`; leave historical CHANGELOG entries intact.
+3. **Manually grep for superlatives** — `grep -rniE "newest|most capable" README.md CLAUDE.md guide/ docs/index.html .claude/rules/` — and re-verify each hit. The checker validates *version strings*; a claim like "X is the newest model" is a **semantic** assertion it can only partially catch (it flags `newest`/`most capable` lines that name a non-top tier, but tier-relative phrasings like "the newest Opus" are legitimately allowed). The 2026-06-09 Fable 5 launch made exactly this class of claim false while the gate stayed green.
+4. Add a "Changed — model refresh" entry to `CHANGELOG.md`; leave historical CHANGELOG entries intact.

.claude/references/scheduled-routines.md

@@ -21,18 +21,23 @@ The nightly reproducibility job is the *backstop*. The *immediate* signal is the
 
 ## Setting one up
 
+`/schedule` takes a **natural-language description**, not flags:
+
 ```text
-/schedule create "nightly-repro" --cron "0 6 * * *" \
-  --prompt "Run /audit-reproducibility against the current passport. If any claim FAILs (not EXPLAINED), summarize which tables are affected and push a notification; otherwise exit quietly."
+/schedule nightly at 6am: run /audit-reproducibility against the current passport.
+If any claim FAILs (not EXPLAINED), summarize which tables are affected and notify me;
+otherwise exit quietly.
 ```
 
-`scripts/nightly-repro-check.sh` is a thin local equivalent for users who prefer a machine cron over a Routine (note: a local cron does not survive a closed laptop — prefer `/schedule`).
+A precise cron expression (e.g. `0 6 * * *`) is applied via `/schedule update` *after* the routine exists; manage with `/schedule list` / `update` / `remove`. Two scheduling constraints to design around: the **minimum interval is 1 hour**, and accounts carry a **daily run cap** — so batch checks into one routine rather than many small ones. Routines operate on **committed repos**: anything uncommitted or private-by-design (e.g. a local research vault) is invisible to them.
+
+`scripts/nightly-repro-check.sh` is a thin local equivalent for users who prefer a machine cron over a Routine — and the right tool for uncommitted/private material (note: a local cron does not survive a closed laptop; for committed repos prefer `/schedule`).
 
 ## Guardrails for unattended runs
 
 - **Never point an unattended loop at a submission portal, shared/restricted data, or a co-author's inbox without a human gate.** Routines *propose*; a human *sends*. (`/triage-inbox` never auto-sends; the [`git-guardrails`](../hooks/git-guardrails.py) hook still blocks destructive git even in a routine.)
 - **Bound the cost.** A nightly full-manuscript re-audit is fine; a nightly 7× `/seven-pass-review` is not — cost-pilot first.
-- **MCP may be absent in headless/cron runs** (interactively-authenticated servers like Gmail). Routines that need them must degrade gracefully.
+- **Connectors are INCLUDED by default — least-privilege them.** Cloud Routines run with **all of your claude.ai connectors attached, write access included, and no approval prompts**. An unattended routine that only needs to read your repo should have Gmail/Calendar/Slack *removed from that routine's connector list* before it ever fires — the risk is not a missing connector but a fully-armed one acting without you. (Locally-authenticated MCP servers in your *terminal* sessions are a separate thing and may still be absent in other headless contexts — degrade gracefully either way.)
 
 ## Cross-references
 

.claude/references/v2.0-backlog.md

@@ -36,6 +36,15 @@ The v2.0.0 "modernization" release (2026-06-09) shipped the loop-first / gate-en
 
 ## Still deferred (post-v2.0)
 
+### From the 2026-06-10 currency audit (do-soon tier)
+
+- **Awesome-list distribution** — after topics/homepage/CITATION.cff land: one-line PRs to `hanlulong/awesome-ai-for-economists` (CC0) and the main claude-code awesome list(s). Low effort, high reach for the exact audience.
+- **Zenodo DOI** — connect the GitHub repo to Zenodo (owner's account), then the next tagged release mints a DOI automatically; add the DOI badge to README + the `doi` field to CITATION.cff. The citability flywheel for replication-package acknowledgements.
+- **Hook touchpoints note** for `orchestrator-protocol.md` / `scheduled-routines.md` — optional wiring (reduce-trigger / failure-resilience / per-reviewer logging hook events). *Verify each event name against current docs before documenting* — this audit caught fabricated CLI syntax; don't repeat it with hook names.
+- **Persona-segmented cost table** (grad student vs faculty) — a small "what a month costs" table in the guide's Cost-Conscious Composition section; the model-routing rule now carries a one-line version.
+- **Non-Claude coauthor gate note** — document the path for a coauthor who pulls the repo *without* Claude Code: `./scripts/install-hooks.sh` works for them too (pre-commit is plain bash/python), but skills/hooks guidance assumes the Claude loop; a short "for your non-Claude coauthor" README box would close it.
+- **Audience scope (owner-set, 2026-06-10):** econ + closely related fields. The psychology/sociology/public-health cards below remain deferred indefinitely — do not build without an explicit owner ask.
+
 ### Portfolio hub: package development — Stata & Python (R shipped in v1.10.0)
 
 The R package triad shipped in v1.10.0. Stata and Python remain, to cover the rest of the owner's package portfolio (2 Stata packages, 1 Python package).

.claude/rules/model-routing.md

@@ -39,6 +39,17 @@ Model tier is the second cost lever; **effort is the first.** Every model runs a
 
 Set per skill/agent with the `effort:` frontmatter field. Several skills ship at `effort: high` for genuinely hard gates (e.g. `/seven-pass-review`, `/simulation-study`, `/r-package-check`). Match effort to the cognitive demand the same way you match model tier — and tune effort before you swap models.
 
+## Where Fable 5 fits — and where it does not
+
+**Fable 5** (GA 2026-06-09) is the most capable model in Claude Code — and this rule deliberately does **not** route any of the template's fleet to it. Two verified reasons:
+
+1. **Cost discipline.** Fable 5 is $10/$50 per MTok vs Opus 4.8's $5/$25 — a flat 2× on exactly the judgment tier the 70/20/10 split exists to guard. The referee/editor/verifier agents are bounded, single-sitting tasks; Fable's premium is priced for *long-horizon, larger-than-one-sitting* autonomous work, which the fleet is not.
+2. **Protocol maturity.** In one launch-week session's fan-out, Fable 5 subagents failed the forced structured-output tool protocol 28/28 times vs 0 observed failures on Opus 4.8 — a single-session signal, not a benchmark, but exactly the failure mode that matters here: in a fan-out fleet, a silent tool-protocol failure means a review lens returns *nothing*. (Same logic as the "don't push Opus down a tier" anti-pattern: a too-immature judge is as bad as a too-cheap one.)
+
+**Where Fable 5 *is* the right call:** your own interactive sessions on the hardest long-horizon work — a multi-day refactor, a deep research synthesis you'll steer by hand — where you are in the loop to catch a protocol hiccup and the task actually exploits the model's horizon. Select it per-session (`/model fable`); leave the fleet's `model:` pins alone. Re-evaluate at Fable point releases (the protocol gap is the kind of thing that gets fixed); when it does, the high-judgment tier is the natural first candidate.
+
+**Cost reality check (grad-student budgets):** a full `/review-paper --peer` runs a meaningful fraction of a dollar-denominated token budget at Opus prices; doubling the judgment tier doubles that line item with no quality evidence yet. When cost-constrained, drop *effort* first (the first lever, above), then tier — never the reverse.
+
 ## Why this matters
 
 Cost reduction on routed skills is typically **50–80%** with no quality loss on the mechanical tier. The cache-TTL change (5-min default in 2026; Claude subscriptions get 1-hour automatically, API keys opt in) made multi-turn pipelines on API keys materially more expensive; per-agent routing recovers that lost ground without sacrificing the high-judgment lens where it matters.

.claude/scripts/statusline.sh

@@ -2,7 +2,7 @@
 # Claude Code status line: shows permission mode, model, and git branch.
 #
 # Claude Code pipes a JSON session snapshot to stdin. Relevant keys:
-#   .model.display_name      e.g. "Opus 4.x"
+#   .model.display_name      e.g. "Opus <version>"
 #   .permission_mode         e.g. "bypassPermissions" | "plan" | "acceptEdits" | "default"
 #   .workspace.current_dir   absolute path of the cwd
 #

.claude/skills/submission-disclosures/SKILL.md

@@ -0,0 +1,76 @@
+---
+name: submission-disclosures
+description: Generate the submission-time disclosure block for a manuscript — the AI-use disclosure statement matched to the target journal's policy, CRediT author-contribution roles, conflict-of-interest statement, and data-availability statement. Use when the user says "AI disclosure", "disclosure statement", "do I need to disclose Claude", "CRediT roles", "conflict of interest statement", "data availability statement", or is preparing a submission package. NOT statistical-disclosure screening of restricted-data outputs — that is /disclosure-check.
+argument-hint: "[manuscript path] [journal short-name, e.g. AER] [--no-ai | --statements-only]"
+allowed-tools: ["Read", "Grep", "Glob", "Write", "WebSearch", "WebFetch"]
+effort: medium
+---
+
+# /submission-disclosures — The Submission-Time Disclosure Block
+
+Draft the four statements journals now require (or strongly expect) at submission, in one pass: **AI-use disclosure**, **CRediT contributor roles**, **conflict-of-interest**, and **data availability**. Journals tightened AI-use policies through 2025–2026; an undisclosed-AI finding at a top journal is now a research-integrity problem, not a formatting one — so the statement should be drafted *deliberately*, not improvised in the submission portal at midnight.
+
+**This skill is about the author's disclosures TO the journal.** It is unrelated to [`/disclosure-check`](../disclosure-check/SKILL.md), which screens restricted-data *outputs* for statistical-disclosure risk (small cells, PII). Same word, different worlds.
+
+## When to use
+
+- Preparing a submission or resubmission package and the portal asks for AI-use / COI / data-availability statements.
+- A revise-and-resubmit at a journal that adopted an AI policy since the original submission.
+- A coauthor asks "do we need to say we used Claude/Copilot/ChatGPT on this?"
+
+## Phases
+
+### Phase 1 — Resolve the journal's actual policy
+
+1. If a journal short-name is given, read its profile in [`journal-profiles.md`](../../references/journal-profiles.md) (top-5 econ + AEA-imprint policy notes + poli-sci top-3).
+2. **Verify the current policy on the journal's own site** (`WebSearch`/`WebFetch`: "<journal> artificial intelligence policy authors", the journal's submission guidelines page). Policies moved fast in 2025–2026; a cached or remembered policy is not good enough for a submission. Record the URL and retrieval date in the output.
+3. If no explicit AI policy exists, default to the strictest common denominator (disclose tools, scope of use, and human responsibility) — over-disclosure is free; under-disclosure is not.
+
+### Phase 2 — Inventory what was actually used
+
+Interview briefly (or infer from the repo when evident — e.g. `quality_reports/`, session logs, a CLAUDE.md):
+
+- **Which tools** (Claude Code, Copilot, ChatGPT, Grammarly-class) and **for what**: writing/editing prose, code authoring, code review, literature search, data analysis, translation.
+- **What stayed human**: research design, identification choices, interpretation, final verification of every number and citation (tie to the repo's own verification story — `/audit-reproducibility`, `/verify-claims` — when true, *say so*: "all AI-assisted numbers were independently verified against code" is a strength, not a confession).
+- **What AI was NOT used for** when the journal cares (e.g., most policies bar AI as a listed author and bar undisclosed AI-generated images/data).
+
+### Phase 3 — Draft the four statements
+
+Write `quality_reports/submission_disclosures_[manuscript-slug].md` containing:
+
+1. **AI-use disclosure** — journal-matched wording: tools + versions, scope of use, the affirmation that authors take full responsibility for all content and verified all AI-assisted output. Honest and specific; never boilerplate that overclaims ("no AI was used") when the repo's own logs say otherwise.
+2. **CRediT roles** — the 14-role taxonomy mapped to each author (interview for the mapping; flag roles no author holds).
+3. **Conflict-of-interest** — funding sources, paid/unpaid positions, data-provider relationships (IRB/data-use agreements often constrain what must be stated; cross-ref [`confidential-data.md`](../../rules/confidential-data.md)).
+4. **Data availability** — aligned with the replication deposit: openICPSR/DCAS language when the target is an AEA-imprint journal (delegate the deposit itself to [`/replication-package`](../replication-package/SKILL.md); restricted-data access language per [`confidential-data.md`](../../rules/confidential-data.md)).
+
+With `--statements-only`, emit the statements to chat without writing the file. With `--no-ai`, skip statement 1 (the user asserts no AI assistance — note in chat that the repo's own session logs may contradict this, if they visibly do).
+
+### Phase 4 — Parity check against the manuscript
+
+Grep the manuscript for an existing acknowledgments/disclosure section; flag contradictions (e.g., the paper thanks "research assistance" that the COI omits, or an existing AI statement that the new one contradicts). Do not silently overwrite — surface the diff.
+
+## Exit behavior
+
+- **Statements drafted, policy verified:** write the file, print the four statements + the policy URL/date, and remind the user the statements are drafts for *author* review — sign-off is theirs.
+- **Journal policy unverifiable** (site unreachable, no policy found): emit the strict-default statements, clearly marked "default wording — verify against the journal's current author guidelines before submission."
+- **Inventory contradicts `--no-ai`:** stop and surface the contradiction; never produce a false "no AI" statement.
+
+## Flags
+
+- `--no-ai` — Skip the AI-use statement (user asserts none was used). The skill still warns if repo evidence visibly contradicts the assertion.
+- `--statements-only` — Print the statements to chat; write no file.
+
+## Cross-references
+
+- [`.claude/skills/disclosure-check/SKILL.md`](../disclosure-check/SKILL.md) — statistical-disclosure screening of restricted-data outputs (the other "disclosure"; unrelated).
+- [`.claude/skills/replication-package/SKILL.md`](../replication-package/SKILL.md) — the deposit the data-availability statement must match.
+- [`.claude/references/journal-profiles.md`](../../references/journal-profiles.md) — per-journal calibration, incl. the AEA DCAS policy note.
+- [`.claude/rules/confidential-data.md`](../../rules/confidential-data.md) — restricted-data constraints on what the statements can say.
+- [`.claude/skills/humanize/SKILL.md`](../humanize/SKILL.md) — detecting AI-voice in prose; disclosure and voice are separate obligations.
+
+## What this skill does NOT do
+
+- **Screen outputs for statistical disclosure risk** — that is [`/disclosure-check`](../disclosure-check/SKILL.md).
+- **Build the replication deposit** — that is [`/replication-package`](../replication-package/SKILL.md); this skill only writes the statement that points at it.
+- **Decide your ethics.** It drafts honest statements from what you report and what the repo shows; whether a use *needed* disclosing under a vague policy is the author's call — the skill defaults to disclosure when in doubt.
+- **Submit anything.** Statements go in the user's submission package by the user's hand.

.github/CODE_OF_CONDUCT.md

@@ -0,0 +1,7 @@
+# Code of Conduct
+
+This project follows the [Contributor Covenant v2.1](https://www.contributor-covenant.org/version/2/1/code_of_conduct/).
+
+**The short version:** be kind, be constructive, assume good faith. Critique ideas and code as hard as you like — the whole template is built on adversarial review — but never people. Harassment, discrimination, and personal attacks are not tolerated.
+
+**Enforcement:** report conduct issues privately to the maintainer (see the email on [psantanna.com](https://psantanna.com)). Reports are handled confidentially. The maintainer may edit, remove, or reject contributions and ban contributors for behavior that violates this code.