Ma Deal Workbench Analysis

M&A Deal Workbench Analysis

Tasks give you a deal id, a running workbench, and input/payloads/answer_template.json. You return one JSON object conforming to that template and nothing else.

The deliverable's name changes between tasks (issue register, deviation matrix, closing package, transition review, committee memo) but the work is always the same: pull the deal's records, compare the current draft against the governing playbook or policy, quantify the gaps in dollars, and classify each one.

Scoring is field-by-field against a reference answer, so a right analysis in the wrong units, casing, or shape loses the point. Precision matters more than prose.

Procedure

1. Read the prompt for the contract it sets

Extract and write down: deal_id, which side you act for, the playbook or policy id, the required units (currency, percent precision, month format, date format), and the list of topics the answer must cover. Prompts state per-task overrides — percent precision differs between tasks, and one task can set a different precision for a single field. The prompt's own wording wins over any default.

2. Read the template before gathering data

answer_template.json is the specification: it fixes the key names, the array sort order, the enum vocabularies, the stable issue/redline id lists, and the null convention. Templates come in two styles:

• mirror — the JSON shape itself, with placeholder values
• descriptive — blocks like required_output_shape, allowed_enums, issue_object_fields that describe the shape in prose

Both are binding. Where a template enumerates possible_issue_ids or stable_issue_ids, those ids are the complete candidate set: work through every one and decide whether the deal data supports it.

3. Pull the deal bundle

python3 skill/scripts/fetch_deal.py <DEAL_ID> -o bundle.json --triage

This fetches every per-deal record plus the governing playbook rules and policy thresholds into one JSON file, and prints a triage report to stderr: stale terms flagged for exclusion, playbook preferred-vs-fallback prose, consent and contract splits, employee totals, exact risk-estimate sums, and a percent-of-headline ready reckoner.

The workbench holds ~85 deals, many of them deliberate near-name clones of the real one. Key every lookup on the exact deal_id; never resolve a deal by project or target name. For cross-table checks use POST /api/query with {"token": "deal-workbench-readonly", "sql": "..."}.

4. Analyse

Work through reference/analysis_rules.md. The four rules that decide most of the score:

1. Drop stale terms. Only staleness_flag = "current" rows are live. Stale rows exist to be excluded — and templates often want their ids listed as exclusions.
2. Read the prose, not just the numbers. A term's numeric_value carries its headline figure, but draft_value text routinely holds a second figure the template asks for separately. Playbook limit_value is the fallback; the preferred position exists only in preferred_position prose, and a fallback's conditional clause can set another issue's required position.
3. Missing terms are findings. When the client's position requires an affirmative provision and the draft is silent, emit the issue with missing_required_term and an empty source_term_ids.
4. Quantify against the stated basis. Dollar figures come from headline_value unless a source names another basis; deltas are positive integers measured in the client's favour.

Direction of every comparison follows client_side: a buyer wants a higher cap and longer survival, a seller the opposite.

5. Validate before returning

python3 skill/scripts/validate_answer.py answer.json input/payloads/answer_template.json \
    --bundle bundle.json --percent-decimals 2

Checks enum conformance, integer-dollar formatting, percent precision, count fields against the arrays they summarise, required top-level keys, and every record id against the bundle so nothing is hallucinated. Pass per-field precision overrides as --percent-decimals '1,fully_diluted_pct=4'. Exit code is non-zero when something is wrong; fix and re-run until clean.

Then re-read the prompt's coverage list and confirm each requested topic appears.

Output discipline

• Emit only the JSON object — no prose, no markdown fence.
• Currency as bare integers. Percent points at the stated precision. Months as integers.
• Enum strings and record ids copied exactly; normalise source High/Medium/Low to the template's HIGH/MEDIUM/LOW.
• Never round a sourced figure to make it look tidy — some stored values end in irregular digits on purpose.
• Follow the template's null convention: emit every field with explicit null when the template lists each field as "... or null"; include only applicable keys when the template's sample object is a union of variant-specific fields.

Files

• reference/workbench_data_model.md — endpoints, response wrappers, table and field reference, and what each decision-carrying field means
• reference/analysis_rules.md — classification, arithmetic, aggregation, risk-rating, prioritisation and output rules
• scripts/fetch_deal.py — deal bundle fetcher + triage report
• scripts/validate_answer.py — template conformance and hallucination checker