# HIRE-1 — headcount as a new data dimension: does hiring growth (a documented anomaly) survive on this apparatus, and is revenue-per-employee worth a look?

status: open · raised 2026-09-04 (owner instinct, interactive session: "information is valuable, some more than others — we should be in pursuit of data") · class: research / NEW DATA (backfillable lane — register the test, then acquire) · judgement: owner-raised, panel/triage to rank · effort: ~0.5 pd coverage probe, then ~2 pd acquisition + one pre-registered run · horizon: none (the history exists in old filings; nothing decays)

---

**Plain-language summary for an owner reading one paragraph.** Nothing in
this system has ever looked at how many people a company employs. That number
is genuinely new information relative to everything we hold (prices, volumes,
options, borrow flags, short flow, and the profit/asset fundamentals), and
one version of it has published evidence behind it: firms that grow headcount
fast tend to earn LOWER returns afterwards (Belo, Lin & Bazdresch 2014, the
"hiring rate" anomaly — a cousin of the asset-growth effect, which was tested
here as a tilt and rejected). The owner's original framing, revenue per
employee, is a productivity LEVEL; it sits in the quality family this shop
has already killed, so it rides as a secondary check, not the lead. This row
registers the exam before anyone touches the data: what to acquire, how to
date it, the one statistic, and the number that kills it. The honest prior is
a NO — large-cap anomalies published in 2014 usually decay — and a cheap,
permanent no is a win under the charter.

## Prior art (checked 2026-09-04 — C0)

- `grep -ri "employee|headcount|hiring"` over RESEARCH.md, research/, and the
  owner-side memory: **zero hits** — never considered.
- Rejected relatives (do not relitigate): `asset_growth_tilt`, `quality_tilt`,
  `quality_filter`, `accruals_tilt`, value sleeve (memory/fundamentals-pit,
  memory/strategic-posture). Those were TILTS on the momentum book from
  balance-sheet data. This row is a different input (labour) and, for the
  lead hypothesis, a different mechanism (real-investment/over-expansion).
- The two "next dataset" deep-research passes (2026-06-09: 33 candidates,
  none cleared; 2026-07-25: short-sale constraints → shortability capture)
  did not list headcount. New candidate, not a re-proposal.

## Mechanism (why it could carry information the price has not absorbed)

Hiring is a real investment decision that shows up in the 10-K once a year,
weeks after the fiscal year end, in prose — not in the structured financial
statements. Aggressive hirers are, on average, over-extrapolating growth; the
2014 result is that their subsequent returns are lower, with an effect size
comparable to the investment/asset-growth anomaly. Revenue per employee is a
productivity level: informationally denser about the business, but slow,
public, and in the quality family the market prices well in large caps.

## Data plan — owned, point-in-time, backfillable

1. **Coverage probe first (~0.5 pd, no trial cost).** SEC XBRL `dei:EntityNumberOfEmployees`
   where filers tag it; count Russell 1000 PiT-member × fiscal-year cells
   covered 2006–2025. If ≥ 70% → proceed on XBRL alone. If not → text
   extraction from the 10-K itself (Item 1 "Employees" / "Human Capital"
   paragraph, EDGAR full-text) for the gaps. Record the coverage number in
   this row before any statistic is computed.
2. **Dating rule — the shop's cardinal one:** a value is known at
   `filed_date + 1 day`, never at fiscal year end (memory/fundamentals-pit).
   The hiring rate at date t uses the two most recent FILED values.
3. **Storage:** `data/fundamentals/headcount.parquet` (gitignored research
   tier, next to `quarterly_fundamentals.parquet`; one row per filing:
   symbol, period_end, filed_date, employees, source ∈ {xbrl, text}).
   Backfillable, so NOT a CAPTURES.md stream; no moat claim.
4. **Universe / prices:** Russell 1000 PiT membership (owned) on the golden
   store (survivorship-free). No new dependency in the live path.

## The pinned statistics (the whole family — nothing else counts)

- **S1 — PRIMARY: hiring-rate rank-IC.** Per calendar month-end m from
  2006-01 to 2024-12: `h_i(m)` = (EMP_latest − EMP_prior) / EMP_prior using
  values filed before m; Spearman rank correlation across names with valid
  h between `−h_i(m)` and the subsequent 12-month total return. Report the
  mean monthly IC, its Newey-West t (lag 11), and the fraction of months
  with the literature's sign (negative hiring → positive IC of −h).
- **S2 — SECONDARY (the owner's framing): revenue-per-employee rank-IC**,
  same construction, revenue from the owned XBRL `Revenues` tag where present
  (added to `fundamentals.py`'s field map only for this test). Reported, not
  gating.
- **Null (the C2 convention):** 200 random-permutation nulls of h across
  names within each month; S1 is quoted as a percentile of that null.

## Decision rule and kill criterion — pre-registered, negative-controlled

- **Stage 1 (one registry row, spent only when computed):** S1 passes iff
  mean IC > the null's 95th percentile AND the sign matches the literature
  AND the effect holds in both halves (2006–2015, 2016–2024). **Any leg
  failing = KILLED**, recorded in this row and the public kill list; S2 is
  reported alongside regardless and cannot rescue a failed S1.
- **Stage 2 (only after a Stage 1 pass, second registry row):** the
  operational form is a **long-only EXCLUSION lens** on the flagship — drop
  the top hiring-rate decile from the monthly top-50 selection (the same
  shape as the shortability exclusion idea; annual data updates monthly as
  filings arrive, so it fits the monthly cadence) — judged by the full SHIP
  gate: walk-forward verdict BETTER, standard AND survivorship-free CPCV with
  PBO ≤ baseline and OOS Sharpe > baseline, corr robust to n_groups, no
  window regressing ≥ 0.05, DSR reported. Absolute Sharpes carry the
  timing-luck spread; PBO/corr carry their null percentiles.
- **Negative controls before either stage:** the IC harness must return a
  null-percentile ≈ 50 on a shuffled h (else the harness is broken); the
  exclusion lens with a RANDOM decile excluded must not clear the SHIP gate
  (else the gate is broken).
- **Vacuity / expiry:** if the coverage probe returns < 50% of cells, this
  row closes as "not acquirable at retail cost" without a trial spent. If not
  acquired within two quarters of approval, it lapses and is re-proposed
  with a fresh coverage probe.

## Effort, cost, and what this is NOT

~0.5 pd probe; ~2 pd acquisition if text extraction is needed; one CPCV-class
run per stage. Trial cost: 0 rows until Stage 1 computes, then 1; Stage 2 adds
1. It is NOT a moat capture (backfillable), NOT a live change (nothing touches
sizing or a pinned rule), and NOT a promise: the modal outcome is a recorded
NO, which is the cheap permanent closure the charter counts at full value.

*Route note: raised by the owner interactively; it enters `open/` like any
panel row and moves to `approved/` only by the owner's hand. Computation
starts only after that move — registration before computation.*

---

## Triage note — 2026-09-07 (annotation only; the body above is untouched)

**Rescuing two review findings that currently exist only on a rolling marker.**
Panel run 19 (2026-09-04) performed its change-review on this row and recorded
the verdict *"prior-art claims TRUE, S2 quality-adjacency honestly handled,
path correct"* — plus **two sharpness notes addressed to triage**, which live
ONLY in the `beacon/panel` `last-run.txt` marker. That file is overwritten by
the next panel run; nothing in `research/` carries them. Verified this run: a
grep of `research/` for "block-permute", "anti-conservative", "Newey" and
"both halves" returns only this row's own §"The pinned statistics" text.
Reproduced here verbatim so approval does not silently lose them:

> two sharpness notes for triage: (1) define "holds in both halves";
> (2) within-month permutation null anti-conservative for autocorrelated
> monthly ICs — gate on NW t or block-permute. Pin before Stage 1 computes.

**Triage does not act on either.** Both would change this row's pinned
statistics, which is substance — and pinning a statistic after it was
proposed, by the party that ranks it, is exactly the shape this shop's
registration-before-computation rule exists to prevent. **They are the
owner's to pin, and the pin must land before Stage 1 computes, not after.**
Note that (2) is not a nit: an anti-conservative null makes S1's "> the null's
95th percentile" leg easier to pass than the row's decision rule intends, so
it biases toward a false YES on the one gate that decides whether a second
registry row is spent.

**Ranking (asked for by the row's own status line).** Not in this week's
decide-three, and the reason is not merit: it carries no date and nothing
about it decays — the filings are in EDGAR and will be there in a year —
while all three items ranked above it have dates inside eight days or are
live safety wedges. It leads the Waiting list instead, and it is named there
as the cleanest available test of whether the approval gate exists at all:
this is the owner's own idea, the first frontier-lens proposal ever to reach
`queue/open/`, and it has sat unapproved for four days while eleven other
items were built by hand around it.