The 5 LPA-review agents
/fundadmin:fund review-lpa reads the LPA and classifies the fund. The terms agent runs first and builds an inventory of the provisions. Risk, compliance and obligations then work from the source and that inventory, and may run in parallel. The recommendations agent runs last, once all four have returned, and is not scored. The draft report is for professional review; nothing in it is approved until a person signs it off.
The weights
This chart shows methodology v2's four scored dimensions, with risk in teal. This scoring chart is separate from the cyan and violet F used as the FundAdmin AI brand mark.
- Terms & Provisions · 25%
fundadmin:fund-terms - Risk Assessment · 31.25%
fundadmin:fund-risks - Regulatory Compliance · 25%
fundadmin:fund-compliance - Obligations & Timeline · 18.75%
fundadmin:fund-obligations - Recommendations (excluded) · not scored
fundadmin:fund-recommendations
| Agent | Dimension | Weight | What it does |
|---|---|---|---|
fundadmin:fund-terms | Terms & Provisions | 25% | Extract and categorize every defined term, fee, waterfall, governance, reporting, LP-protection, capital, transfer and confidentiality provision from an LPA or fund document; its output feeds the risk, compliance, obligations and recommendations agents |
fundadmin:fund-risks | Risk Assessment | 31.25% | Score each LPA provision 1-10 for GP-favorability, LP exposure and enforceability, and rank the highest-risk provisions from the document's own terms, citing a comparison only from an identified, dated source; consumes the terms agent output |
fundadmin:fund-compliance | Regulatory Compliance | 25% | Compare an LPA or fund document with SEC, ERISA, FATCA/CRS, AML/KYC, AIFMD and SFDR requirements and flag provisions that are missing, unclear or inconsistent with a cited requirement; it does not certify compliance. Consumes the terms agent output |
fundadmin:fund-obligations | Obligations & Timeline | 18.75% | Map every obligation, capital activity trigger, notice period, cure period, reporting deadline, key person trigger and LPAC consent requirement in an LPA into an operational timeline with financial exposure; consumes the terms agent output |
fundadmin:fund-recommendations | Recommendations (excluded) | 0% | Consumes the terms, risk, compliance and obligations outputs and drafts priority-ranked negotiation points, alternative language and action items for qualified human review, citing a comparison only from an identified, dated source or an ILPA published position |
The score
Methodology v2 uses raw weights 20 / 25 / 20 / 15 over a denominator of 80, which normalise to 25% / 31.25% / 25% / 18.75%. Risk is rated 1 to 10 (higher is worse) and converted to 0 to 100; the other dimensions are rated 0 to 100 directly. Recommendations carry no weight.
- The result is a heuristic document favourability score, 0 to 100. The legacy name "LPA Safety Score" survives in some field names; it is not a safety verdict, a legal compliance certification or an investment recommendation.
- If any scored dimension is unknown, no aggregate is issued: the score is reported as unavailable rather than filled in.
- Coverage confidence (how much of the document the evidence covers) is reported separately from favourability.
- There are no letter grades. Nothing in the score implies acceptance.
The weighting is the author's framework and is described in each agent's prompt and in the scoring methodology; it is a way to rank provisions for review, not a legal opinion.
What to expect
- A full review runs all five agents. Duration depends on the host, the document and the tools available; no timing is guaranteed.
- Before the agents run, the review asks four regulatory facts it cannot read from the LPA: the adviser's registration status, the offering route, the investor mix and the marketing jurisdictions. Without them, or in an unattended run, the compliance dimension is reported as unavailable and no aggregate score is issued. Even with them, a compliance point is assessed only where a registered authority is on file; otherwise it stays UNKNOWN.
- The report opens with Key Issues and Asks: the most material issues ranked, each with its source locator, direct effect and a proposed stance (push, trade or accept). A stance is a negotiation position, never an investment or compliance verdict.
- The report,
LPA-REVIEW-[fund]-[date].md, is written into an isolated run with an evidence sidecar (see output files). A report with missing checks is markedPARTIALorBLOCKED. - On a host without subagents, the same five prompts run inline in the same order: terms, then risk, compliance and obligations, then recommendations.
- The agents are only used by
review-lpa; the other commands are single skills.
