Ma Deal Workbench JSON

M&A Deal Workbench → structured JSON

A family of tasks share one shape: you are counsel for one side of one deal, you read that deal's records from a workbench API, you compare the current draft against a playbook or policy, and you emit one JSON object conforming to a supplied answer_template.json with no prose around it.

The deliverable name changes (issue register, deviation matrix, closing package, committee escalation memo, transition review). The method below does not.

The single most important rule

The answer template is the specification. The prompt is context.

Read input/payloads/answer_template.json completely before touching the network. It dictates field names, nesting, enum vocabularies, id vocabularies, units, and rounding. When the prompt and the template disagree about a field, the template wins. When the template fixes a list of stable IDs, that list — not the number of records you find — determines how many objects you emit.

Run this first to extract the machine-checkable contract:

python3 skill/scripts/extract_contract.py <path-to-answer_template.json>

It prints every enum vocabulary, every stable-ID list, every declared unit, and the required field tree. Keep that output in front of you while you build.

Workflow

1. Establish access

Read environment_access.md for the base URL, the allowed endpoint list, and the POST /api/query token. Use only what that file authorizes. Confirm the service is up before planning around it:

curl -s -o /dev/null -w '%{http_code}\n' "$BASE/api/deals"

If POST /api/query is authorized, prefer it — one SQL round trip beats a dozen REST calls, and it is the only practical way to do cross-table checks.

2. Pull the whole deal bundle at once

python3 skill/scripts/wb.py bundle PRJ_XXXX          # every table for one deal
python3 skill/scripts/wb.py sql "SELECT ..."         # read-only SQL
python3 skill/scripts/wb.py get /api/deals/PRJ_XXXX  # raw endpoint

Pull everything up front — terms, playbook/policy, consents, employees, material contracts, regulatory, cap table, diligence findings, benchmarks, risk estimates, notes, documents. Fields you think are irrelevant routinely turn out to be the evidence that a required term is missing.

See reference/workbench_api.md for the endpoint map, JSON wrapper keys, and the full table schema.

3. Scope to exactly one deal

The corpus holds many decoy deals, and project names are deliberately reused across unrelated deal_ids (near-twin names differing by one word or suffix are the norm, and several distinct deals share an identical project_name).

• Filter every table on deal_id = '<the id in the prompt>'. Never match on project_name, target name, or counterparty.
• GET /api/search searches across all deals. Treat its output as leads only; re-fetch anything you intend to use, scoped by deal_id.
• Take client_side, playbook_id, and policy_id from the target deal's own deals row. Do not infer the playbook from the prompt's framing — a deal can carry a counterparty-flavored playbook id, and decoy deals carry the other policy version with different thresholds.
• A deal has a playbook or a policy, rarely both. Compare against whichever the deals row names; if it names neither, the comparison basis must come from the prompt or template.

4. Keep only current draft terms

draft_terms.staleness_flag is current or stale.

• Filter to current for every substantive comparison. Stale rows are planted distractors.
• A category whose only row is stale means the current draft is silent on that category — it is a missing term, not a draft position. Never read a numeric off a stale row.
• Expect at most one current row per (deal_id, category). Two current rows in one category means you mis-scoped the query.

5. Compare against the playbook or policy

Read reference/field_semantics.md before doing this — the numeric columns do not mean what their names suggest. In particular:

• playbook_rules.limit_value is the fallback threshold, not the preferred one. The preferred position exists only as prose in preferred_position and must be parsed out of that sentence.
• Fallbacks are frequently conditional ("fallback N months if escrow is at least X%", "cap may reach N% only for verified …"). The condition is part of the position; check it against the deal's own data and let it drive the final_position / recommended_action enum you pick.
• Polarity flips with client_side. Sellers want caps, escrows, and survival low; buyers want them high. The same draft number is draft_below_playbook for one side and draft_exceeds_playbook for the other. Derive direction from client_side, never from intuition about which number is "worse".
• For policy-driven committee tasks, escalate only categories where policy_thresholds.restricted_flag = 'yes' (equivalently, the row's approval_required names the committee). Categories routed to a lower approver are distractors even when the draft sits near their threshold.

6. Treat absent terms as issues

Most templates enumerate more issues than the deal has draft terms. The gap is intentional: where the template names an issue the current draft does not address, and the surrounding records show the protection is needed, emit it as a missing term.

• source_term_ids (or equivalent) → []
• issue_status → missing_required_term
• recommended_action → add

Conversely, several draft terms may collapse into one template issue — the source-id field is an array precisely so a scope term and its fee term can share an issue. Emit one object per template ID, not one per record.

7. Compute amounts

• Default base is the deal's headline_value. Override it whenever a source row states a different basis — enterprise value, equity value, upfront_cash, fully diluted shares, material contracts. The basis column is authoritative over the prompt's default.
• amount = round(base * percent_points / 100), emitted as an integer.
• Deltas point to the position you are demanding: delta_to_fallback = draft_amount - fallback_amount (keep the sign convention consistent across the document and consistent with the field's name).
• Composite terms hide their second number in prose. numeric_value carries one quantity only; a row reading "escrow equals X% for N months" exposes X in numeric_value and leaves N to be parsed from draft_value.
• Rounding is per-task — 2dp, 1dp, whole points, and 4-decimal holder fractions all appear. Take it from the prompt and template, never from habit.
• cap_table.fully_diluted_pct is a fraction summing to 1.0, not percent points. Multiply consideration by it directly; convert only if the template asks for percent points. Use as_converted_shares, not shares, wherever the template says as-converted.
• Sanity check: upfront_cash + stock_value + milestone_value should reconcile to headline_value. If it does not, re-check that you are on the right deal.

8. Map supporting tables to fields

Aggregate exposure totals include only the components your escalated issues actually rely on. When a template has included_ / excluded_exposure_components fields, it expects you to drop the categories no live issue maps to — and to say so explicitly rather than silently summing all three.

9. Emit and validate

Output one JSON object, nothing else — no markdown fence, no commentary.

python3 skill/scripts/validate_answer.py answer.json <path-to-answer_template.json>

It flags unknown enum values, IDs outside a fixed vocabulary, non-integer currency, template placeholders left unreplaced, and structural drift. Fix everything it reports, then re-read the checklist below.

Pre-submission checklist

• Every record used came from a query filtered on the target deal_id.
• Only staleness_flag = 'current' rows drove numbers; stale-only categories were treated as missing, not as positions.
• Playbook/policy came from the target deal's own deals row.
• Comparison direction matches client_side.
• Preferred values were parsed from prose; limit_value was used as the fallback only.
• Every template stable ID appears exactly once; no invented IDs.
• Every enum value is copied verbatim from the template's vocabulary.
• Currency fields are integers; percents and months match the required precision; dates are YYYY-MM-DD.
• Counts in the summary block agree with the arrays they describe.
• Nulls are used where the template allows them, rather than 0 standing in for "not applicable".
• Output parses as JSON and contains no prose.

Files

• reference/workbench_api.md — endpoints, wrapper keys, table schema
• reference/field_semantics.md — column meanings and the traps in each
• scripts/wb.py — fetch a deal bundle, run read-only SQL, hit raw endpoints
• scripts/extract_contract.py — pull enums, stable IDs, and units from a template
• scripts/validate_answer.py — check an answer against the template contract