executive-comp-benchmarking

Example: Executive Compensation Benchmarking

The second Executive Compensation arm agent: a dark, board-ready dashboard that positions the subject's Named Executive Officers' pay against the approved peer group's real, SEC-disclosed proxy pay — element by element, as a percentile of the peer distribution versus the committee's target-percentile policy — and stops at a human approval gate. It never recommends pay.

It reads the shared benchmarking engine over the committed real peer proxy dataset (each figure from a company's latest DEF 14A Summary Compensation Table; the subject is the same synthetic Acme Corp the rest of the portfolio uses), and leads with the honest headline — including where the subject is behind target.

Synthetic subject, real peer pay. Acme (the subject) is synthetic; the peer figures are actual, as-disclosed SCT proxy pay for real public companies — an illustrative, dated snapshot (provenance + per-company SEC sources: governance/proxy-comp-data.md), not investment/comp advice; verify against each filing for current actuals. Peer figures are actual pay, not target opportunity — individual executive names are not stored (the dataset is role-based).

What it shows (the way a committee reads a proxy)

For each benchmarked role (CEO / CFO / COO / CLO) and each pay element — base salary, annual cash incentive, total cash, LTI / equity, total direct comp — the dashboard shows:

• where the subject sits as a percentile of the peer distribution (mid-rank), on a 0–100 track with the committee's target band shaded;
• the peer P25 / median / P75 for that element;
• a below / on-target / above call, with the gap in percentile points.

The honest headline for Acme: cash is competitive, but long-term equity sits below target across roles — 6 of 20 positions land below the committee's band, all in LTI and total direct comp. A thin role (CHRO, only two peers disclose it) is suppressed, not given a spurious percentile.

The methodology (defensible by construction)

• Actual, not target. Peer figures are as-disclosed SCT pay (equity at grant-date fair value). Positioning actual pay against a target-percentile policy is the standard proxy read; the dashboard says so and never implies the peer numbers are targets.
• One incumbent per company per role, medians not means. A transition-year outgoing officer is dropped; a single founder mega-grant can't distort the market read.
• Suppress thin roles. A role below the engine's minimum peer count is shown suppressed, never positioned off two data points.
• US SCT only in the distribution. Two peers are foreign private issuers (they file a 20-F / furnish a 6-K circular, not a US Summary Compensation Table, and their grant-date equity is not cleanly comparable); they are excluded from the percentile distribution and surfaced as a caveated count — the math runs on the 14 US-SCT peers.
• Row-level provenance. Every peer figure in proxy_comp.csv carries its SEC filing URL, form type, currency, extraction date, and a per-row basis caveat — so any number traces back to the exact filing.

Run it

cd examples/executive-comp-benchmarking
python3 run.py                                              # draft only
python3 run.py --publish                                    # refused: needs a named committee approver
python3 run.py --publish --approved-by "Compensation Committee Chair"

Test it

python3 evals/test_benchmarking_agent.py
python3 ../../tools/verify_evidence.py --verify-sources output/*.evidence.json

Sample output

• Committee dashboard (HTML)
• Committee digest
• Dashboard evidence graph
• Digest evidence graph

Each evidence graph makes nine material committee claims machine-verifiable. The local proxy snapshot, synthetic subject, all 16 underlying public-company SEC disclosures, positioning transformation, target-band assumptions, exclusions, suppression policy, checks, and caveats are first-class nodes; no network call is needed to re-hash the committed inputs.

The .html + .md are the deterministic artifacts (CI byte-diffs them against a fresh run.py). The .png is an illustrative browser render committed for the README preview — it is not byte-gated (screenshots are browser-dependent), so treat it as a snapshot, not the source of truth.

What it demonstrates

• Honest positioning: leads with where the subject is behind target (equity), not only where it is competitive — the read a committee actually needs.
• Actual-vs-target discipline: peer pay is labelled as as-disclosed SCT actuals, positioned against a target-percentile policy — the seam a comp professional probes first, made explicit.
• Presentation + governance only: every percentile, quartile and call comes from the shared benchmarking engine; the agent does no math, suppresses thin roles, fails closed, and stops at a human approval gate.
• Real peers, synthetic subject: the peer distribution is real public-company proxy pay (dated snapshot, per-company SEC sources); only Acme is synthetic, and no individual name is stored.

Related

• executive-comp-peer-builder — builds the peer group this agent benchmarks against (run it first; this agent positions pay once the group is approved).
• skills/sec-comp-research — a portable skill that pulls the same kind of real proxy data from SEC EDGAR for any tickers you supply.