Ma Deal Workbench

M&A deal workbench → structured JSON deliverable

These tasks all have the same shape: a prompt names one deal, points at a read-only M&A workbench, and requires only JSON conforming to an answer_template.json. Output is graded by field-level exact comparison, so a defensible-but-invented value scores the same as a wrong one. Everything you emit must be traceable to a record you actually read, or to arithmetic on those records.

The one rule that matters most

Derive, don't infer. The workbench is the whole world. Do not add issues, IDs, amounts, or classifications from general M&A practice. If a record does not say it, it is not in the answer.

Its corollary, filter, don't pad, is the single highest-value habit: a list field named "required", "blocking", or "closing condition" holds only the records whose own flag says they qualify. Listing every record and tagging each with a type is measurably worse than listing only the qualifying ones — non-qualifying records belong in the template's non-blocking / excluded / notice-only field, if it has one, and nowhere otherwise.

Workflow

1. Read the answer template before touching the API

The template is the contract and the strongest hint about the expected answer. Extract:

• Required top-level keys and their exact spelling. Emit exactly these.
• allowed_enums — every enum-typed value must be a verbatim member. Watch case: workbench records store "High"/"yes", templates usually want "HIGH"/true.
• Pinned ID lists (possible_issue_ids, stable_issue_ids, stable_redline_ids, an issue_id "one of:" list). When the template pins IDs, that list is the expected membership — produce every ID, no more and no fewer. When it does not, include an entry only where a record evidences it.
• Enums that only make sense nested (fee_model, tax_allocation_method, forum, condition_type) — these are telling you the vocabulary for the free-form *_normalized / must_have_terms objects. Use those exact strings there.
• Units — integer dollars, percent-point decimal places, integer months. The prompt sometimes overrides the template; the prompt wins.
• Named final positions. A final_position / recommended_action enum whose values map 1:1 onto the issue IDs is a decoded answer key — pair them by name and let the wording drive the row. These names are often self-describing, encoding the target number, the fallback, and the condition attached to it; read them as instructions.

2. Pull the deal bundle

scripts/fetch_deal.py <base_url> <deal_id> <out_dir> fetches the deal record, follows its links map, and pulls the governing playbook or policy named on the deal. It discovers routes from the deal record, so it does not depend on a hardcoded route list.

The workbench holds dozens of decoy deals whose project names differ from the target's by a syllable or a suffix, with cloned supporting rows carrying similar-looking amounts. Filter on the exact deal_id every time, including in SQL. Never let a similarly named project supply a number.

3. Establish the governing standard

The deal record names exactly one:

Parse the numeric preferred/fallback out of the prose positions — limit_value alone is ambiguous (it is sometimes the preferred bound, sometimes the fallback).

4. Select the rows

Use only draft_terms with staleness_flag == "current". Then, per template shape:

• Issue register / deviation matrix — one row per template-pinned issue ID. Map each to the current draft term(s) sharing its category; a pinned ID with no matching current term is missing_required_term with empty source_term_ids.
• Committee escalation — include a term only if all hold: current, its category has a governing threshold, the draft breaches it, and the threshold is in scope for the requested body (restricted_flag == "yes" / approval_required matches). Stale, in-policy, and lower-approval rows are deliberate distractors — put them in the template's excluded_* fields, not in the escalation list.

reference/classification_rules.md has the status / risk / action decision tables and the record-flag filters (consents, material contracts, employees).

5. Compute the numbers

• Base = the deal's headline_value unless the term or rule basis explicitly names another basis. headline_value also equals upfront_cash + stock_value in these deals; a draft term that states its own dollar figure confirms the base — check it.
• amount = round(base * percent / 100), emitted as an integer.
• Deltas run in the direction that harms your client (see polarity table in reference/classification_rules.md). Report the gap to the fallback when the template asks for one gap, and both fallback and preferred gaps when it asks for two.
• Exposure ranges come from risk_estimates rows, matched by category. Count each category once in an aggregate even when several issues cite it, and list which categories you included/excluded if the template has fields for that.
• Every total must equal the sum of the components you actually emitted.

6. Validate, then emit

Run scripts/validate_answer.py <candidate.json> <answer_template.json>. It checks required keys, enum membership, integer-dollar fields, percent rounding, and pinned-ID coverage. Emit raw JSON only — no prose, no code fences, no trailing commentary.

Pitfalls

• Padding a required/blocking list with non-qualifying records. The costliest recurring error. required_for_closing == "no" and consent_required == "notice only" are never blockers.
• Inventing IDs. Prefer stable workbench IDs (CNS_*, MAT_*, EMP_*, TERM_*, RSK_*, FND_*). Only synthesize one where the template explicitly permits it (e.g. "synthetic regulatory id").
• Free-styling risk ratings. Inherit risk_default from the governing rule, or the record's own risk_rating, before making a judgment call. Normalize to the enum's case.
• Ignoring client_side. It flips draft_exceeds_playbook vs draft_below_playbook and flips which party a term protects.
• Double-counting an exposure category across several issues in an aggregate total.
• Trusting limit_value instead of the preferred/fallback prose.
• Wrong percent precision. Two decimals, one decimal, and four-decimal holder percentages all appear; re-read the prompt for each task.
• Adding keys the template does not list is harmless; omitting listed keys is not.