# WP29 — Descriptive summary of the mandate corpus

Coding date 2026-10-29. Inputs: `research/corpus/` (13 files, hashes verified against `corpus_manifest.csv`), `research/corpus_automated.csv` (script), `research/corpus_coding.csv` (analyst). Codebook: `research/corpus_coding_protocol.md`.

**Denominator.** Eleven distinct CLAUDE.md mandates (WP15, WP18–WP26, WP28). `WP15_PROMPT_PROMPT_INITIAL_CLAUDE.md` is byte-identical to `WP15_CLAUDE.md` (same SHA-256) and is counted once. The WP27 design brief is a different genre (a post-implementation description of a simulator, not an instruction file) and is excluded from every "of 11" count; WP27's CLAUDE.md was not retrievable (HTTP 404) and nothing is inferred about it. Coding was a single automated + LLM-analyst pass with no human second coder; no intercoder reliability is claimed. The corpus is small and researcher-selected: the counts motivate design patterns for WP29 and do not estimate prevalence.

**Important limit.** These documents are instructions. They say what the author asked Claude Code to do. They do **not** establish how the earlier working papers were actually executed, whether the instructions were followed, or whether the papers were produced without human assistance (several mandates in fact anticipate human steps: expert raters, human review of ambiguous studies, credential provision, confirmation of funding lines).

## 1. Length and form

| Statistic (11 mandates) | Value |
|---|---|
| Words — min / median / max / mean | 1 092 (WP19) / 1 966 (WP22) / 8 809 (WP28) / 2 879 |
| Sorted word counts | 1 092, 1 114, 1 132, 1 358, 1 671, 1 966, 2 240, 2 285, 2 314, 7 685, 8 809 |
| Markdown headings — min / median / max | 1 (WP21) / 19 (WP25, WP26) / 151 (WP15) |
| Code blocks | 0 in six mandates; 1 in WP18, WP20, WP25; 12 in WP15; 27 in WP28 |
| Markdown tables | 0 except WP28 (2) |
| Language | English body in all; WP18 and WP20 carry a French prefatory paragraph (heuristic label `mixed`) |

Two mandates (WP15, WP28) are an order of magnitude longer than the other nine, which cluster between 1 100 and 2 300 words. Length tracks the amount of engineering prescription (directory trees, manifests, CLI commands, rubrics), not the number of substantive research rules.

## 2. Feature counts

Values are the analyst codes from `corpus_coding.csv` (first token of each cell). "partial" is defined in the codebook as present but indirect, conditional or without an operational mechanism.

| Feature | yes | partial | no | Note |
|---|---|---|---|---|
| Explicit research question (verbatim) | 10 of 11 | — | 1 (WP22 states objectives, not a question) | |
| Explicit estimand or hypotheses | 7 | 4 | 0 | H1–H6 lists (WP25), theorem queues (WP24, WP28), pre-specified primary outcome (WP26), core-estimand section (WP22) |
| Assumptions stated in the mandate | 2 (WP24, WP28) | 8 | 1 (WP23) | most mandates *require* assumptions to accompany every result without stating them |
| Falsification / null-result conditions | 6 | 4 | 1 (WP21) | "attempt to disprove its own central claim" (WP15); "A negative finding is publishable" (WP26) |
| Literature / novelty requirement | 11 | 0 | 0 | universal; usually a "novelty gate" plus a verified literature matrix |
| Data-access controls | 7 | 4 | 0 | credentials, scraping, privacy, licensing rules; `partial` = theory papers with only "do not pretend access" |
| Temporal-information controls | 6 | 1 | 4 | universal among the data-based mandates (WP15, WP20, WP22, WP23, WP25, WP26); not applicable to the pure-theory ones |
| Baseline / benchmark requirement | 8 | 3 | 0 | `partial` = baseline scenario or benchmark model rather than comparison methods |
| Phases declared | 10 of 11 | — | 1 (WP23) | 5 to 13 phases; median 7 |
| Dependency / gate structure | 10 explicit | 1 implicit (WP21) | 0 none | but only 2 (WP15, WP28) tie a gate to an executable test |
| Stopping rules | 2 (WP15, WP28) | 6 | 3 | `partial` = "if blocked, document and deliver a partial result" |
| Recovery / resume instructions | 2 (WP15, WP25) | 6 | 3 | `partial` = progress logs, caching, retry limits |
| Provenance requirements | 9 | 2 | 0 | manifests with hashes in WP15, WP28; prompt/model/seed logging in WP25, WP26 |
| Review / self-audit stage | 10 | 1 (WP23) | 0 | results audits, adversarial referee reports (WP24, WP28), red-team agent (WP26) |
| AI-use / disclosure statement | 1 (WP28) | 0 | 10 | |
| Budget / resource limits | 1 (WP26) | 3 | 7 | WP26 asks for API-cost tiers; others estimate or profile without a cap |
| Integrity prohibitions ("never invent …") | 11 | 0 | 0 | universal, usually in the opening paragraph |
| Pre-registration mentioned (automated keyword) | 4 (WP19, WP21, WP25, WP26) | — | 7 | none specifies a registry or timestamp mechanism |
| Random seeds mentioned (automated keyword) | 11 | — | 0 | |
| Statements marked as added/reconstructed | 6 | — | 5 | explicit "added" sections in WP24, WP26; reconstruction notes in WP23; prefatory repository-mapping notes in WP18, WP20, WP25 |

## 3. Recurring design patterns (what the mandates consistently require)

1. **Role framing + integrity preamble.** Every mandate opens by casting Claude Code as a multidisciplinary "research team" and immediately prohibits fabrication of data, results, citations, proofs or significance (11 of 11). The prohibition is the single most stable element of the corpus.
2. **Ordered phases ending in "only then write the paper".** Ten of eleven declare 5–13 phases; nine make manuscript drafting conditional on validated results ("Never write a result in the paper before it has been generated and verified from a saved machine-readable output", WP15). Gates are named in ten mandates, most explicitly as numbered acceptance gates (WP20: Gates 1–7; WP25: six named gates; WP28: Gates 0–12).
3. **Verified-literature / novelty gate before contribution claims** (11 of 11), typically with a literature matrix file and a "do not call it novel unless …" clause.
4. **Audit and self-review as a phase** (10 of 11): a results audit mapping every claim to a file (WP18, WP19), an adversarial referee report (WP24), three simulated referees with a verdict vocabulary (WP28), or a red-team agent inside the architecture (WP26).
5. **Traceability of every number** (9 of 11 "yes"): "Every paper number must map to a result file" (WP18), results registries feeding LaTeX macros (WP15), run manifests with hashes (WP28), versioned outputs (WP26).
6. **Information-set discipline wherever data exist** (6 of 6 data-based mandates): look-ahead, as-of-date, vintage and chronological-split rules.
7. **Named baselines** (8 of 11), from "random walk, no change" (WP23) to cost-matched single-agent controls (WP26).
8. **Honest-failure fallbacks**: "If blocked, document the issue honestly and deliver a scoped partial result" (WP18); "A dead end with a mathematical explanation is better than a fabricated breakthrough" (WP28).

## 4. What the mandates consistently lack

1. **Externally enforced or executable gates.** Gates are declarative. Only WP15 ("The release script must fail if …") and WP28 ("completing one phase's acceptance tests before starting the next") attach a gate to something that can fail mechanically; none specifies who or what checks a gate other than the executing agent itself.
2. **Independent verification.** "Independent audit" (WP19), "independently checked" (WP20), "independently verified" (WP21), "hostile-referee critique" (WP25) and "three independent review perspectives" (WP28) are all to be performed by the same agent; WP28 is the only mandate that names this tension ("do not have the same model merely declare its own work accepted").
3. **Operational pre-registration.** Four mandates ask for pre-registration; none specifies a registry, a timestamp, a hash or a freeze mechanism that would make it verifiable.
4. **Stopping rules.** Only WP15 (a dedicated "Stop conditions" section) and WP28 (stop only on a real blocker; mandatory final verdict) define when to stop; most others say "continue with feasible work".
5. **Recovery / resumption.** Only WP15 (resumable, idempotent, checkpointed jobs) and WP25 (checkpointing, manifests) require it; the rest rely on progress logs.
6. **Budgets.** Seven of eleven set no resource limit; only WP26 asks for an API-cost budget, and no mandate states a numeric token, dollar or compute cap.
7. **AI-use disclosure.** Ten of eleven say nothing about disclosing AI involvement in the paper; WP28 is the exception because it inherits a prior "GPT-6.1" co-author credit that it must reconcile with journal policy.
8. **Quantified success criteria.** Targets are phrased as "substantially improve" (WP23), "publication-quality", "publishable contribution", "the standard expected in a top finance journal" (WP15) — none is checkable from the document. Page, figure, table and reference counts are given as ranges with anti-padding escape clauses (WP18, WP19, WP20, WP24, WP28).
9. **Authorial provenance of the instruction itself.** WP23 states that part of its text was reconstructed by an assistant after truncation; WP24 and WP26 carry sections "added" on 2026-10-27; WP18, WP20 and WP25 carry prefatory repository-mapping notes. The mandate as retrieved is therefore not always the mandate as originally issued.

## 5. Verbatim excerpts illustrating the patterns

1. Integrity preamble — WP15_CLAUDE.md: "Never fabricate data, results, citations, institutional details, variable definitions, statistical significance, sample sizes, or conclusions. Never write a result in the paper before it has been generated and verified from a saved machine-readable output."
2. Phase gate before drafting — WP18_CLAUDE.md: "**Gate:** do not start manuscript drafting until the baseline model is validated, claimed GE equilibria clear markets, principal simulations and exhibits reproduce, bibliography is verified, and theoretical claims have proofs or appropriate qualifications. If blocked, document the issue honestly and deliver a scoped partial result."
3. Executable release gate — WP15_CLAUDE.md: "The release script must fail if: a table references a missing result; a figure is missing source data; LaTeX does not compile; citations are unresolved; labels are duplicated; page count exceeds the current target; anonymous PDF contains author-identifying text; manuscript claims disagree with the results registry; data revision is not pinned; tests fail."
4. Novelty gate — WP28_CLAUDE.md: "Do not claim that distributed gradients, one-step Newton, split-sample jackknife, Richardson extrapolation, `K = o(N^{7/8})`, or `K = o(N^{11/12})` are intrinsically novel. Investigate who has already obtained each idea or bound." and "**Gate 1:** no unqualified novelty language is allowed until the closest competing theorems have been inspected and cited."
5. Temporal-information control — WP22_CLAUDE.md: "Never use post-transaction information or future data in real-time estimates." / WP26_CLAUDE.md: "Comparable retrieval must be strictly as-of-date; no same-property future transfers, post-sale listings, or near-duplicate records leaking across partitions."
6. Pre-specified primary outcome — WP26_CLAUDE.md: "Pre-register primary endpoints and hypotheses before running final test evaluations. Primary outcome: out-of-time absolute log-price error on arm's-length residential transactions."
7. Falsification and null results — WP15_CLAUDE.md: "The project must attempt to disprove its own central claim." / WP28_CLAUDE.md: "A dead end with a mathematical explanation is better than a fabricated breakthrough. A strong negative result or counterexample may itself become the contribution."
8. Self-review by the same model, acknowledged — WP28_CLAUDE.md: "Simulate at least three independent review perspectives; log objections and responses … Record unresolved issues candidly; do not have the same model merely declare its own work accepted."
9. Instruction provenance — WP23_CLAUDE.md: "[Sections 11–15 and the remainder of the protocol were reconstructed by the author's assistant because the pasted protocol was truncated.]" / WP26_CLAUDE.md: "## Repository conventions (added by the implementation, 2026-10-27)".

## 6. Implications for the WP29 protocol design (motivation only)

The corpus shows a stable *rhetorical* protocol — role, integrity rules, phases, gates, audits, traceability — and an equally stable absence of *mechanisms*: gates are not executable, audits are not independent, pre-registration is not timestamped, budgets and stopping rules are rare, and the instruction text itself is sometimes amended after the fact. WP29's protocol should therefore turn the recurring declarative elements into verifiable ones (hashed and frozen mandates, machine-checked gates, external evaluators, explicit budgets and stop conditions, recorded human interventions) rather than add more prose rules. Nothing in this corpus demonstrates that the earlier papers were produced without human assistance; that question requires run-level evidence, not instruction files.