replicate-paper — locked reproduction, then a fenced, disclosed in-design search

What this skill carries — grep references/ for any subject the names below miss: !d=${CLAUDE_SKILL_DIR}; command -v skill-toc >/dev/null 2>&1 && exec skill-toc "$d"; s=$HOME/.claude/skills/plugin-utils/bin/skill-toc; [ -x "$s" ] && exec "$s" "$d"; echo "(skill-toc unavailable: references and scripts are NOT listed here — install the plugin-utils plugin, or start a new session so its bin/ reaches PATH)"

Templates: ${CLAUDE_SKILL_DIR}/templates/check.sh, ${CLAUDE_SKILL_DIR}/templates/PROMPT.md, ${CLAUDE_SKILL_DIR}/templates/ledger.csv. Worked example (ADKO not reproduced, Antoniou fragile hit): ${CLAUDE_SKILL_DIR}/references/worked-example.md. Scoreboard rows and the appendix paragraph: ${CLAUDE_SKILL_DIR}/references/closing-and-disclosure.md.

A replication is a locked reproduction of the paper's printed cells on the paper's own sample. When it fails, the user may authorise an in-design search (the AGK rule): deviations only inside the paper's own design, every specification priced and ledgered, the count disclosed beside any hit. A hit found that way is a searched specification, never a locked result.

fetch paper ─► cite cells & definitions (page/line) ─► lock predictions + bar (user) ─► md5 the lock
   ─► reproduce on the paper's own window ──pass──► extension rows ─► done
                     │ fail
                     ▼
   AskUserQuestion: in-design search? ──no──► record the non-replication ─► done
                     │ yes
                     ▼
   scaffold grind_<id>/ (check.sh, frozen.md5, ledger, PROMPT.md) ─► check.sh at baseline exits 1
   ─► grind.sh run --gate <leads_exhausted key> --wait-alert 1
        ├─ check exits 0 ─► verify hit vs printed table ─► user: accept as fragile? ─► robustness rows ─► disclose
        │                         └─ second gated cell owed? ─► phase 2 (PROMPT2, new journal)
        └─ leads_exhausted filed ─► grind.sh stop ─► record NOT reproduced, with the count

<EXTREMELY-IMPORTANT> ## IRON LAW: NO ESTIMATE BEFORE THE LOCK, AND NO BAR CHANGE AFTER AN ESTIMATE

The bar — which cells, sign, |t|, CI, window — is written, user-approved and md5-frozen before the first coefficient is computed. After that, only the user changes it, and their words go verbatim into check.sh's header. A bar chosen after seeing estimates is a specification search reported as a replication: it hands the user a result that the next referee takes apart, which is the opposite of help. If an estimate makes you want to move the bar, that wish is the finding — report it. </EXTREMELY-IMPORTANT>

1. Obtain the paper and cite it

• Get the version of record through the fetch-paper skill; pdftotext -layout it to a .txt beside the PDF so every cite is a line number. Only a working paper on disk → say so in check.sh's header ("WP values may differ"). Prevents: a build from memory of the paper, and a fence nobody can adjudicate.
• Every cell, filter, variable definition, window and winsorizing rule the build uses is cited by page or .txt line. Prevents: an "allowed reading" that the paper never offered.

2. Lock predictions and the pass bar

• Write prediction_locked.md (bar cell(s), published sign, p threshold, window(s), class/measure rule, pre-specified robustness, any predictions) and get it approved through AskUserQuestion. md5 it into frozen.md5 before estimating. Prevents: the bar drifting toward the estimate.
• Derive the published SE: se = |b|/|t|; when only p is printed, se = |b|/z with z = Φ⁻¹(1 − p/2); 95% CI = b ± 1.96·se. State the derivation and the rounding of the printed coefficient in the header. Stars only → the CI is undetermined; ask the user, do not invent one.
• Gate only cells significant in print. A published null is scored by the matched-null rule (ours insignificant at 5% passes), never as a hit. Prevents: AGK 2019 Table 4 — a "done" inside a null's CI, rejected by the user because a null's CI admits almost anything.
• Joint rule over every printed-significant cell the user gates. One step must carry them all IN; hits on different steps never combine. Ask which cells are gated before the loop starts. Prevents: CHL g_34, DROA IN with BHAR OUT on the same row.

3. Reproduce on the paper's own sample and window first

• The locked reproduction runs on the paper's sample and window before any extension; extension windows are separate rows with their own predictions. Data that do not reach the paper's window → the first job is the pull, not a different gated window. Prevents: gating a window the paper never estimated; the Antoniou hit came from restoring 2003-2016, not from any construction switch.
• WRDS through the wrds skill: SAS on the SGE grid via qsub only, never Postgres; a one-year test job, inspected, then one pull per source over the whole window, split by year locally. Prevents: 18 serial per-year queries plus count probes, which the user stopped by hand.

4. The in-design grind

Only after the user picks "in-design search" through AskUserQuestion. One loop per paper, under scratch/grind_<id>/, ledger at data/output/<id>_grind.csv.

Piece	Rule	Prevents
check.sh (from the template)	Exit code is the verdict: 0 hit, 1 not yet, 2 integrity. Per cell: published sign, |t| ≥ 1.96 on our coef/se, coef inside the published CI. Joint over the gated cells. Prints BEST/MOST_CELLS_IN progress lines. Run it at baseline; it must exit 1 on the committed run, not 0.	a loop that grades itself; a check that passes on nothing
Integrity → exit 2	frozen files changed, ledger header changed, a (step, cell, window) priced twice, non-numeric coef/se, a gated cell that is a published null	silent edits to the locked build; re-pricing until it hits
frozen.md5	md5 of the committed run's outputs, scripts, tests, note, the lock file and check.sh itself	an iteration editing the baseline or the gate
Ledger	Append-only CSV, step,cell,window,class_rule,coef,se,t,p,n,first_stage_F,switches,sample_n; g_NN_<slug>; a corrected build gets a new number	a deleted row the count no longer discloses
Declared switches	Every switch vs the committed run, ;-joined key=value, declared in the log before estimating, with the fence reason and the paper's page/line	a hit whose deviations nobody can list
inference_only=1	Set in switches for a step that changes only SEs/clustering; check.sh reads the field, never the slug; such a row is priced but never ranks or hits	clustering passed off as a replication
No threshold tuning	Median splits, winsor percentiles, bandwidths, size cuts stay at the paper's values; an alternative enters only as a named reading of the text, before its row is priced; N and treated shares are reported, never targeted	tuning a cutoff on the outcome
Fence	Allowed: choices the paper leaves open or defensible readings of its text. Forbidden: controls/estimators the paper lacks, selecting on the outcome or residuals, raw-value edits, windows outside the paper's	a "replication" of a model the paper did not run
Rejected steps	A step the user later rules out of fence stays in the ledger and the count, and is listed verbatim in check.sh's REJECTED set so it cannot hit	BY17 g_13, controls in a column that prints none
User decisions	Quoted verbatim, dated, in check.sh's header	an operator's paraphrase widening the bar
PROMPT.md (from the template)	One g_ row per iteration; ranked leads, each citing the paper's page/line and the code line it would change; stack only switches named before pricing	an iteration re-diagnosing from scratch, or stacking on t

Launch through the grind skill, detached, with the gate on the FILED key:

G=/abs/scratch/grind_<id>; J=$G/<id>.jsonl
setsid nohup bash ${CLAUDE_PLUGIN_ROOT}/skills/grind/scripts/grind.sh run \
  --journal "$J" --prompt-file "$G/PROMPT.md" --check "bash $G/check.sh" \
  --gate "jq -R 'fromjson? | select((.kind==\"floor\" or .kind==\"note\") and .key==\"leads_exhausted\")' $J | grep -q . && { echo 'leads_exhausted filed: operator stop due'; exit 1; } || exit 0" \
  --wait-alert 1 --stall-after 8 --max-iters 40 >"$G/grind.log" 2>&1 </dev/null &

The first wait after the floor is filed alerts you; answer it with grind.sh stop, never by waiting for the stall count. Prevents: 3-8 empty iterations after leads_exhausted (BKL 10 of 20); and a grep -c leads_exhausted that counted prose mentions as filings.

Phase 2 re-arm — a first hit accepted, another gated cell owed, or the user adds a gate: copy PROMPT.md to PROMPT1.md; write PROMPT2.md = a # PHASE 2 header (what phase 1 found, next step number, the new ranked leads, what not to re-price) + PROMPT1.md; a new check2.sh when the bar changes (user's words verbatim); launch with a new journal <id>_p2.jsonl, same ledger. The phase-1 journal, PROMPT1.md and the superseded check are retired: add them to frozen.md5, never resume them, one live journal per loop. Prevents: phase-1 floors and the stall count carrying into a different objective.

5. Closing

1. Verify a hit against the printed table yourself — re-read the published cells, rerun the hit step from its script, confirm check.sh exit 0 and the ledger row match. Prevents: relaying an iteration's claim.
2. Ask the user (AskUserQuestion) to accept it as fragile, or reject it — fragile hits are the norm. Record the answer verbatim. Prevents: a search hit promoted to a result by the operator.
3. Price robustness rows the user asks for on the hit build (another window, alternative construction, singletons): disclosed, never counted as a second hit even when check.sh prints HIT.
4. Disclose every searched step: the g_ count, pulls, rejected and inference-only rows, the switches the hit stacks, and what it does not survive.
5. Scoreboard: hit rows are appended after the main rows, tagged in-design search hit:, with lock_status searched and n_specs_searched; the locked verdicts and the main tally do not move. Details in references/closing-and-disclosure.md. Prevents: a search silently rewriting the bar's pass count.
6. Not reproduced is a result: record the best row per cell, why the fence was exhausted, and the count.

6. Composition — use these, do not restate them

Need	Skill
the PDF	fetch-paper
the loop, journal, floors, stop/status	grind
SAS/SGE pulls, CCM/link tables	wrds
delegating a heavy pricing or build step	farm-out — the result-deciding fix as its own short row, cleanup as separate rows
the estimation code itself	ds conventions

Red flags — STOP

About to	Why wrong	Do instead
Compute a coefficient with no prediction_locked.md in frozen.md5	every later bar is post hoc	lock, get approval, md5, then estimate
Gate a cell the paper prints as insignificant	a null's CI admits almost anything	matched-null rule; gate printed-significant cells
Count two cells IN on different steps as a joint hit	the paper's result is one coding	joint rule: one step, all gated cells
Edit check.sh, a threshold or the window after seeing estimates	the gate moves to the estimate	ask the user; new check2.sh, their words verbatim
Re-price a ledger row or delete one	the disclosed count lies	new step number stating the correction
Name a row _cl/_hc and let the slug decide	the check never sees it	inference_only=1 in switches
Gate the loop on a prose grep for leads_exhausted	it matches notes about the key	jq on .key=="leads_exhausted"
Let the loop run after leads_exhausted is filed	each pass is an empty model call	grind.sh stop on the first alert
Arm a session hold beside the grind on the same objective	each waits on the other	one driver: the grind
Loop a WRDS pull by year, or query Postgres	the user stopped exactly this	one SGE pull per source, one-year test first
Report a hit without the g_ count and the stacked switches	a searched row read as a replication	disclose the count beside the hit
Move a main-row verdict because a search hit landed	the locked bar stops meaning anything	append a tagged hit row; main tally unchanged