git:20260924.115e7c0 to git:20260927.84e1b67

40 added, 13 removed. Audit A to A.

---

name: cla_review

description: Review one convertible-loan term sheet from the lender's side, against the SECA CLA model term sheets and in the context of the startup's cap table, existing loans and SHA. Use before SICTIC members lend.

---

# CLA term sheet review

Identify the term sheet, extract its terms, judge them for a lender and show what they do in this company.

## Inputs and outputs

The async `cla_review(dataset_name, *, document=None, ticket=None, fresh=False)`

returns one Markdown report in `list[InsightFile]`, named

`cla-review-<startup>-<document>-<model>.md`, one per reviewed document. JSON

intermediates (`identification`, `extraction`, `assessment`, `conversion`,

- `loan-context`, `sha-context`, `audits/<checklist>`) live under

- `insights/cla-review/` and carry the document identity in their identifier.

+ `loan-context`, `reference`) live under `insights/cla-review/`; the audits

+ live under `insights/batch-audit/` because `batch_audit` owns them. Every

+ intermediate carries the document identity in its identifier.

`document` names the data-room path of the term sheet; omitted, the model

selects the most recent CLA term sheet and executed loans are context, not

candidates. `ticket` is the member's own amount and is part of the

configuration key, so a new ticket replaces the generated result for that

document. `fresh` bypasses generated reuse without deleting files or

- overriding manual inputs.

+ overriding manual inputs; the audits are reused by `batch_audit` on their

+ own configuration key and are not regenerated by `fresh`.

**Slices:** slice 0 (contract, rules, references, fixture), slice 1

- (question 1: identify, extract, company- and lender-angle assessment) and

+ (question 1: identify, extract, company- and lender-angle assessment),

slice 2 (question 2, deterministic part: my conversion, the existing loans,

- 10/20 before and after) are built; the audits against the SECA term sheets,

- the SHA context and the synthesis are listed in the report as pending. The

+ 10/20 before and after) and slice 3 (audits against the selected SECA

+ reference, executability against the SHA and articles, synthesis) are

+ built. The lender-perspective checklists are seeded drafts marked

+ `provisional` in `settings.json` until Enrico has reviewed them. The

slices are defined in

[docs/cla-review-design.md](../../docs/cla-review-design.md#7-slices).

- JSON intermediates per document: `identification`, `extraction`,

- `assessment`, `conversion`, `loan-context`.

Artifact identity is the slugified full relative path of the document, so

equal basenames in different folders stay distinct. The automatic

identification is one dataset-level artifact (`identification-automatic`);

an explicit `document` gets its own (`<document>-identification`), so the

two never reuse each other. The report is

`cla-review-<startup>-<document>-<model>.md` directly under `insights/`.

## Workflow and dependencies

Prepare and synchronize the startup. Identify the term sheet: an explicit

`document` is resolved through the shared document-path resolution at the

configured minimum score; otherwise the model selects the most recent CLA

term sheet or draft from three retrieval queries, executed loans being

context and never candidates, and only the selected path is resolved. An

existing `captable_build` classification, when present, seeds candidates

and exclusions and is part of the identification key.

Extract the selected document with the shared `captable_build` CLA

checklist (`config/captable_build/cla_terms.md`, including its term-sheet

provisions group); every value carries a verbatim quote and every absence a

`missing_terms` entry. Assess it twice: `assess_cla` with the

`captable_build` bands (company angle) and `lib.cla_review.assessment`

with `config/cla_review/settings.json` (lender angle). Every lender rule

carries value, source, status, effective date and an `active` flag; only

`approved` rules may be active and judge, an inactive rule yields an open

question with the observation, a rule whose input is unavailable is not

evaluated and names the input. The term in months is measured from the

run date, which is part of the assessment key.

Each stage is a managed JSON artifact validated on read and write against

`config/cla_review/artifact_schemas.json` or the `captable_build` CLA

schema; a stage reuses its artifact when the configuration and the

upstream artifact are unchanged. Manual files win, `fresh` bypasses

generated reuse only, failures are never saved. `ticket` is part of the

report key so a new ticket regenerates the report for the same document

without regenerating the extraction.

Question 2 reads the consolidated `captable_build` snapshot through

`lib.captable.insights.select_consolidated`; it is never generated here and

the registry declares no prerequisite. Its state is part of the conversion

and loan-context keys: `reusable` (content keyed), `absent` or `stale`.

`lib.cla_review.conversion` resolves and labels every input before any

number (ticket: given, else the aggregate amount divided by the configured

member count as a stated assumption; round size: the qualified-financing

minimum; denominator, interest, currency, nominal value, valuation grid

from the cap or `valuation_grid_absolute`), omits a calculation whose

input is missing with the reason, and prices the ticket at every grid

point and conversion date through `lib.captable.model` next to the existing

loans (`lib.captable.notes`) and the other new lenders. `lib.cla_review.loans`

puts the term sheet next to every executed loan, reads the MFN effect both

ways, checks identical-terms membership and counts the 10/20 rules before

and after N members through `aggregate_clas` on a synthetic executed copy.

+ The SECA reference is the one the term sheet names through its extracted

+ `documentation_form` (`settings.documentation_form_references`, an

+ evidenced choice without a model call); a `bespoke` or `unstated` form

+ ranks both references in one direct model call. The `reference` artifact

+ records which path was taken. Every checklist in

+ `config/cla_review/checklists/` is then audited through

+ [batch audit](../standards_and_architecture/SKILL.md#checklist-audits)

+ with the complete term sheet and the selected reference as the cacheable

+ prefix, and every checklist in `executability_checklists/` with the term

+ sheet and its extracted conversion assumptions (capital sources, consents,

+ pre-emption, share class, accession) plus the SHA, articles and register

+ documents a `captable_build` classification names, so retrieval reaches

+ the company's own documents; the SECA drafting notes are evidence, never

+ instructions. `skill_name` is `cla_review-<document>`, which puts the

+ document identity into the audit identifier and the stored `skill` field.

+ Each audit is validated complete before the report; a failed check is an

+ error, not a finding. The synthesis is one Markdown generation over the

+ audits, both assessments and question 2 under

+ `config/cla_review/summary_instructions.md`, which forbids turning an

+ `open_question` rule, a missing input or an insufficient snapshot into a

+ verdict. The report renders every audit as a table and ends with the

+ synthesis.

## Side effects and failure behavior

Preparation may import, convert and index documents. The review calls models

and saves intermediates and the report; it sends nothing. An absent or stale

cap-table snapshot yields an insufficient-evidence paragraph for question 2

and the run completes; a malformed snapshot, failed path resolution and

- technical failures remain errors. No plausible term sheet, an unresolvable path or an extraction

- failure stops the run without saving a partial artifact. The report aids human

- review and is not legal advice.

+ technical failures remain errors. No plausible term sheet, an unresolvable path, an extraction

+ failure or a failed audit check stops the run without saving a partial report. A

+ run costs one model call per check (41 with the seeded checklists), each

+ carrying the term sheet and the reference in its prefix, plus at most one

+ ranking and one synthesis. The report aids human review and is not legal

+ advice.

## Usage

```bash

conda run -n sictic-env python -m skills.harness '/cla_review "<STARTUP>"'

conda run -n sictic-env python -m skills.harness -- /cla_review "<STARTUP>" --document "legal/cla-term-sheet.pdf" --ticket 25000

conda run -n sictic-env python -m skills.cla_review --startup "<STARTUP>" --document "legal/cla-term-sheet.pdf" --ticket 25000 --fresh

```

The direct CLI uses `--startup`; `--dataset` and `-d` remain compatibility aliases.

## References

- [Implementation](cla_review.py)

- [Design and slices](../../docs/cla-review-design.md)

- - [Configuration, rules and SECA reference term sheets](../../config/cla_review/)

+ - [Configuration, rules, checklists and SECA reference term sheets](../../config/cla_review/)

- [Shared CLA extraction checklist](../../config/captable_build/cla_terms.md)

- [Shared audit contract](../standards_and_architecture/SKILL.md#checklist-audits)