cohort-benchmarking

name	cohort-benchmarking
description	Benchmark a client privately against the firm's OWN book of business. At each close or review, compute the client's operating metrics (gross margin, payroll-%-of-revenue, DSO, rent-%, owner-comp ratio), feed each one DE-IDENTIFIED into the vr-cohort MCP under its NAICS peer group keyed by a one-way HASH of the client id, then score the client against that cohort and flag outliers in plain English ('payroll is 3.2 sigma above your 14 retail clients'). Builds a proprietary peer dataset no public benchmark matches, behind a hard PRIVACY WALL: only the abstracted statistical signal crosses clients, never a client's confidential figures. Use during month-end-close, financial-statement-prep, or QBR prep, or when asked 'how does this client compare', 'is X normal for this industry', or 'flag anything off'. Reads metrics from quickbooks/vr-ledger; observations and scoring are GREEN; any peer comparison reaching a client deliverable is RED until reviewed.
version	1.0.0
author	RealDeal CPA
license	MIT
platforms	["linux","macos","windows"]
metadata	{"hermes":{"tags":["Benchmarking","Cohort","Peer-Analysis","NAICS","Outlier-Detection","Privacy-Wall","De-Identification","Advisory"],"related_skills":["month-end-close","financial-statement-prep","client-context","consulting"]}}

Cohort Benchmarking

Every close, every client teaches the firm something — if you capture it. This skill turns the book of business into a private benchmark: each client's operating metrics feed an anonymized peer pool keyed by industry, and each client gets scored against that pool. The firm's edge is a dataset built from its own clients that no purchased benchmark can match — and a privacy wall that keeps it ethical.

The privacy wall (read first, never cross)

Only the abstracted statistical signal crosses clients — a metric value attached to a NAICS code and a hashed id. No client's confidential data ever lands in another client's workspace.

• Never pass a client name, EIN, account number, or any identifier to vr-cohort. The id is always a one-way hash: client_hash = sha256(firm_salt + client_id)[:16]. Keep firm_salt in home/cohort/salt (created once, never committed, never logged).
• The cohort store holds only (naics, metric, value, client_hash, period) — numbers, not books.
• A score you read back is a position ("3.2 sigma high"), not another client's figure. When you report it, describe the cohort in aggregate ("your 14 retail clients"), never name a peer.
• With fewer than the minimum cohort size, do not score — report "insufficient peers (n=4)" rather than a number that could fingerprint one client. Treat n<8 as not-yet-benchmarkable.

The metrics (compute these five each cycle)

Pull real numbers from quickbooks (qb_pnl_report, qb_ar_aging, qb_balance_sheet_report) or vr-ledger (income_statement, balance_sheet) — tools first, memory never. Use Decimal, not floats.

• gross_margin = (revenue - COGS) / revenue
• payroll_pct = total payroll / revenue
• dso = (accounts receivable / revenue) * days-in-period
• rent_pct = rent expense / revenue
• owner_comp_ratio = owner compensation / revenue Express ratios as decimals (0.0–1.0) consistently so cohorts stay comparable. Record the period and the source report in a workpaper note for each figure (see QC below).

The vr-cohort MCP (only these three tools for this server)

1. add_observation(naics, metric, value, client_hash) — record one metric for one client into its NAICS peer pool. Call it once per metric per close. Re-feeding the same client/metric updates that client's contribution (idempotent on client_hash), so re-running a close does not double-count.
2. cohort_stats(naics, metric) — return the peer-pool summary for a metric (n, mean, stdev, median, quartiles). Use it to confirm the cohort is large enough and to describe the peer set.
3. score(naics, metric, value) — position one client's value in its cohort: z-score / sigma, percentile, and the in/out-of-band call. This is what flags outliers.

Argument shape: tools take JSON-string args (e.g. add_observation('{"naics":"448140", "metric":"payroll_pct","value":0.41,"client_hash":"a3f9c1d20b8e4f56"}')) and return concise JSON.

The loop (run at each close / review)

1. Resolve NAICS for the client from client-context; default to the 6-digit code, fall back to the 4- or 3-digit parent if the 6-digit cohort is too thin.
2. Compute the five metrics from the tools. Sanity-check (margins in [-1,1], no divide-by-zero revenue) before you trust them.
3. Hash the client id with the firm salt.
4. add_observation each metric — this client now contributes to the pool.
5. cohort_stats each metric. If n is below the floor, mark "insufficient peers" and skip scoring for that metric.
6. score each metric for this client's value. Anything beyond +/-2 sigma (or outside the inter-quartile band by a wide margin) is an outlier — flag it.
7. Narrate the outliers for the advisory write-up: direction, magnitude in sigma/percentile, and the cohort in aggregate. "Owner comp is 0.9 sigma — normal. Payroll is 3.2 sigma above your 14 retail clients; worth a labor-efficiency look."

Reading the signal (judgment, not just z-scores)

• High sigma is a question, not a verdict — a 3-sigma DSO might be a genuine collections problem or a one-off large receivable. Look before you conclude.
• Direction matters: high margin is usually good, high payroll-% usually isn't; say which.
• A cluster of mild outliers (margin low + payroll high + owner-comp high) tells a story a single metric won't. Read them together.
• Cohorts drift — a metric that was 2 sigma last year may be the new normal as the pool grows. Re-score against the current pool, not a remembered number.

Authority & deception

• Computing metrics, feeding the cohort, and reading scores back internally are GREEN.
• Any peer comparison that reaches a client deliverable (deck, memo, QBR) is RED — route through self-review-qc and Valentino's review/name before it goes out. Never present a benchmark as audited or as a public industry standard when it's the firm's internal pool; label it "firm peer set (n=__)". The no-client-deception line is invariant.
• If a YELLOW judgment call arises (NAICS ambiguous, cohort borderline-thin, a metric that looks wrong), surface it rather than guessing.

Workpaper & QC

For every cycle, leave a note under home/cohort/: per metric, the value, the source tool + report + period, the NAICS used, the cohort n, and the score. Record the salt's existence (never its value). On re-runs, confirm observations are idempotent (no double-count) and that no client identifier ever appears in the cohort store or logs. Cross-link: the numbers → financial-statement-prep and month-end-close; the client/industry facts → client-context; turning a flag into advice → consulting.