IC3 employer gate draft YAML — exact composition, mutation pins, blocked pre-lock

.gitignore

@@ -113,6 +113,9 @@ paper/paper_files/
 scratch/
 *.pkl
 
+# Claude Code local state (agent worktrees are not repo content)
+.claude/
+
 # First-estimates ceremony-owned pre-publication state.
 runs/first_estimates_attempt.claim
 runs/first_estimates_retry.claim

data/external/employer_firm_target_sources.md

@@ -1,6 +1,6 @@
 # Employer-firm target extract provenance
 
-Provenance sidecar for the four committed aggregate extracts that
+Provenance sidecar for the committed aggregate extracts that
 feed the employer-firm extension's calibration targets and gate
 references (E1/E2/E7/E11/E12; issue #192, ADR 0003, workstream B).
 All are small, tidy derivatives of published public aggregate files —
@@ -92,5 +92,133 @@ are keyless and authoritative.
   firm-size-coded contribution as published. 2025Q1 firm-size cells
   are suppressed in this release (status flag 5) and load as NaN.
 - **Unit caveat:** job counts, as for QWI. Origin-x-destination
-  size-ladder flows (E11's J2JOD reference) are a later, separate
-  extract.
+  size-ladder flows are in `j2jod_us_firmsize_od_2015on.csv` (entry
+  6 below).
+
+## 5. `j2j_us_sexage_2015on.csv` — J2J flows by sex x age group
+
+Fetched on **2026-07-17** (extracts 5 and 6 are the second wave,
+flagged in PR #223's method findings; same fetch script).
+
+- **Source URL:** https://lehd.ces.census.gov/data/j2j/R2026Q1/us/j2j/j2j_us_sa_f_gn_ns_oslp_u.csv.gz
+  (release-stamped, not `latest_release`; see the QWI note above)
+- **Release:** R2026Q1, V4.14.0 (`version_j2j.txt`: J2J US
+  2000:2-2025:1, `j2jpu_us_20260312_1118`); sex x age worker detail
+  (`sa`), no firm characteristics (`f`), not seasonally adjusted,
+  state/local/private ownership (`oslp`)
+- **Raw sha256:** `0e043fc8796bd3e11231ff6d174fdfebed926c9d40da4f069a3ad31eed55aba0`
+- **Transformation:** pure row/column filter, no re-aggregation:
+  the all-industry margin (`industry == "00"`) only, the full
+  sex (0/1/2) x age (A00-A08) grid margins included, 2015Q1 onward;
+  12 flow measures with status flags. The full NAICS-sector detail
+  would breach the 1 MB extract cap, so it is not committed; the
+  raw sector file stays available at the pinned URL. Sex and
+  age-group labels joined from the LEHD schema
+  (https://lehd.ces.census.gov/data/schema/latest/label_agegrp.csv,
+  sha256 `eb478c6eda6c12a57609afaf89bbb42dd4d9fb2ee883f6dd0399fb717b27889b`).
+- **Naming caveat:** the age x sex tabulation is LEHD's `sa`
+  crossing; LEHD's `se` crossing is sex x *education* (the gate-E2
+  registration's "se" shorthand refers to sex x age, i.e. `sa`).
+- **Unit caveats:** job counts, not persons (see the QWI entry);
+  ownership is `oslp` here versus `op` for the QWI extract, so
+  levels are not directly comparable across the two. 2025Q1
+  separation-side measures carry status flag -1 (not computable
+  until the next quarter is released) and load as NaN.
+
+## 6. `j2jod_us_firmsize_od_2015on.csv` — J2J flows by origin x destination firm size
+
+- **Source:** LED Extraction Tool query API,
+  https://ledextract.ces.census.gov (POST the pinned JSON request in
+  `scripts/fetch_employer_firm_targets.py` to `/j2j/download`, then
+  GET `/j2j/download.csv?<encoded query>` from the 303 redirect).
+  The LEHD flat J2JOD files (`j2jod_us_d_fs_*`) publish only the
+  one-sided firm-size margins — the full origin x destination cross
+  is not in any flat file — and the Census data API
+  (`api.census.gov`) still requires a key (probed 2026-07-17), so
+  the LED Extraction Tool is the pinned keyless source.
+- **Release: not reported by the tool; inferred.** The extraction
+  tool exposes no release identifier. Its `/j2j/schema` reports
+  `V4.14.0`, which is the *software* version and is **identical for
+  R2026Q1 and R2026Q2** (`version_j2jod.txt`: R2026Q1 =
+  `j2jodpu_us_20260312_1118`, 2000:2-2025:1; R2026Q2 =
+  `j2jodpu_us_20260618_1116`, 2000:2-2025:2), so it cannot
+  distinguish them. An earlier revision of this entry claimed the
+  schema version "matches R2026Q1" and therefore pinned the release;
+  that inference does not hold. What is verified: the values served
+  for 2015Q1-2025Q1 are unchanged across the R2026Q1 → R2026Q2
+  rotation (checked 2026-07-23, all 1,476 rows), so the extract is
+  release-stable over its own window whichever release served it.
+  National, all industries, not seasonally adjusted, ownership A00
+  (state/local government plus private — the tool's reported
+  `ownercode`; equivalent to `oslp`).
+- **Archived response:** `raw/led_j2jod_us_fsfs_2015on.csv.gz`
+  (gzip of the tool's CSV as served 2026-07-23; raw sha256
+  `7afad9f408319c54e7e7d802b068723e346f49c840d4fe861e7e55d9528d3944`).
+  Committed because the tool re-runs the query against whatever
+  release is current, so the 2015Q1-2016Q1 detail window — the only
+  window in which this cross is published at all (next bullet but
+  one) — is not re-fetchable in perpetuity. The archive is the
+  builder's default input; a live re-query is the verification path.
+- **Integrity pin: content, not bytes.** `LED_J2JOD_CONTENT_SHA256`
+  = `c52ec512bc3f478d6426efb7b03cccbe1edc952309214030ed53eb42f7a83354`,
+  taken over the response canonicalised (columns sorted, rows sorted
+  by year/quarter/firmsize_orig/firmsize, fixed float format). The
+  byte digest is recorded but **not** enforced.
+
+  Why: the byte digest originally pinned here
+  (`adbd16e2...`, fetched 2026-07-17) stopped matching on
+  2026-07-23 while every one of the 1,476 rows was unchanged, value
+  for value. The tool had reordered its measure columns —
+  `EE,AQHire,EES,AQHireS,J2J,J2JS` where it previously emitted
+  `EE,AQHire,J2J,EES,AQHireS,J2JS`. A pin that fires on cosmetic
+  reordering trains the maintainer to re-pin on sight, which is
+  exactly how a genuine revision would slip through; so byte drift
+  is now reported and tolerated, and content drift raises.
+- **Transformation:** column subset and sort only, no
+  re-aggregation: the full 6 x 6 firm-size grid (codes 0-5 on both
+  origin and destination sides), 2015Q1-2025Q1 (ordinal quarters
+  8060-8100), six flow measures (EE, AQHire, J2J = EE + AQHire, and
+  their stable variants) with status flags.
+- **Detail-window caveat (important for E11):** the full 5 x 5
+  origin x destination detail is released only for
+  **2015Q1-2016Q1**; from 2016Q2 on every national detail cell
+  carries status flag 11 ("aggregate of cells not released because
+  component cells do not meet publication standards") and loads as
+  NaN. The tool aggregates the state-level OD tabulations to the
+  national level, and a state coverage gap from 2016Q2 blocks the
+  aggregate; the same suppression governs the J2J Explorer, so no
+  keyless public source carries the later cross. The one-sided
+  margins (code 0 on either axis) remain published through 2025Q1.
+  E11's origin x destination shape reference is therefore the
+  2015Q1-2016Q1 window; later quarters constrain the margins only.
+- **Margin caveat (corrected 2026-07-23; the earlier version of this
+  bullet was directionally wrong).** The code-0 margins are the
+  tool's aggregates of the firm-size-coded tabulation (status flag
+  10/12) and they do **not** sit systematically below the flat-file
+  `d_fs` margins. Checked against
+  `j2jod_us_d_fs_gn_ns_oslp_u` (R2026Q1) across all 41 quarters, on
+  the all-demographics / all-industry / all-firm-age national cell:
+
+  | comparison | quarters tool > flat | deviation range | mean |
+  |---|---|---|---|
+  | all-size EE margin | 37 / 41 | −1.00% to +2.02% | +0.75% |
+  | per-size margins (orig and dest, sizes 1-5) | — | −3.20% to +3.67% | — |
+
+  The single 2015Q1 figure previously quoted (3,985,308 here versus
+  3,988,566 in the flat file, −0.08%) is real but unrepresentative,
+  and the explanation attached to it — the flat file's inclusion of
+  public-sector "N" flows — is not what drives the gap: a
+  size-N exclusion would bias the tool's margin *downward* in every
+  quarter, and it is above in 37 of 41. The deviations run both ways
+  and are dominated by independent noise infusion applied to the two
+  tabulations.
+
+  **This matters at floor scale.** E11's post-2016Q1 constraints are
+  margins-only, so a ±2-3% cross-source wobble on the quantity being
+  gated is itself an empirical noise datum for the E11 floor build,
+  not a footnote. Reproduce with
+  `scripts/check_j2jod_margin_agreement.py`.
+- Detail cells that fail publication standards (status flag 11)
+  load as NaN — common in the small-x-large corners.
+- **Unit caveat:** job counts, as for QWI/J2J; firm size is
+  administrative national March employment on both sides.