Correct the amterr lab's layer-3 claims and make it rerunnable

paper/FACTS.md

@@ -130,7 +130,7 @@ Supersession notes (feature-round regeneration, 2026-08-09): E2's stage AUCs/mag
 | # | Fact | Source |
 |---|---|---|
 | K1 | Cause shares, any-presence over each case's nine `AGENCY` slots, share of official error dollars (sum HWGT × `AMTERR`; FY file pools 12 monthly samples so pooled dollars are annual): Colorado strict computation 7.4% ($6.9M/yr), broad rules-engine set 10.6% ($10.0M/yr), software-only 6.3% ($5.9M/yr); national strict 3.5% ($232M/yr), broad 19.3% ($1.27B/yr), software-only 2.5% ($163M/yr), of $6.59B/yr official error dollars. All 53 jurisdictions in the artifact; classes overlap by construction; accounting conventions, not causal estimates of engine-preventable error | analysis/cause_shares.json (deterministic; conventions and code-to-class map serialized in-file) |
-| K2 | The lab snapshot's "10 cases, $3.28M, 3.3%" broad-residual line (ANALYSIS.md L62-67) does not reproduce in dollars: the 10-case count reproduces from the SAV cause slots, but HWGT × AMTERR gives $4.47M and HWGT × engine-gap gives $2.91M (4.5% of the 283-case replay dollars). The manuscript does not quote the $3.28M figure | analysis/cause_shares.json colorado_replay_reconciliation |
+| K2 | The lab snapshot's "10 cases, $3.28M, 3.3%" broad-residual line (ANALYSIS.md L62-67 at 755a7f3; corrected in the lab on 2026-10-03, with every lab figure recomputed in paper/snapshot/labs/amterr/claims_audit.json) does not reproduce in dollars: the 10-case count reproduces from the SAV cause slots, but HWGT × AMTERR gives $4.47M (4.5% of the 283-case replay dollars) and HWGT × engine-gap gives $2.91M. The manuscript does not quote the $3.28M figure | analysis/cause_shares.json colorado_replay_reconciliation |
 | K3 | Layer-2 finding-nature tabulation, committed and test-locked: cases partition by whether populated findings are computational (NATURE in {36,42,43,54,64,65,75,79,80,98,123} or ELEMENT 520; deduction natures {52,53,56,57} only when the paired AGENCY cause is in layer 1's broad computing-apparatus set {10,17,19,20,21,22}). Colorado deviation convention reproduces the archived table exactly: 13/19/13/260 cases, $3.66M/$7.47M/$4.68M/$96.8M, 3.3/6.6/4.2/86.0% of $112.6M. National deviation: 4.6/11.1/6.5/77.7% of $8.133B. The artifact also serializes a wider agency-responsibility convention (system-side = all 14 agency-cause codes) with materially larger shares (CO system-caused 31.1% deviation) and both official-error-denominator variants | analysis/cause_shares.json (schema v2); analysis/NATURE_REPORT.md (techdoc line citations) |
 
 ## L. Engine-comparison artifact and the simulator's verification view (added after revision 5)

paper/snapshot/labs/amterr/README.md

@@ -0,0 +1,140 @@
+# amterr lab: rerunning it
+
+This directory holds the Colorado FY2024 error-case lab described in
+`ANALYSIS.md`. It has three stages:
+
+1. `reconstruct_co_fy2024.R` (R) reconstructs each error case's pre-correction
+   inputs and writes `co_fy2024_reconstruction.csv` and
+   `fy2024_reconstruction_national.csv`.
+2. `amterr_replay.py` (Python, inside an axiom-oracles checkout) runs the Axiom
+   rules engine on those inputs and writes `amterr_replay_results.json`.
+3. `audit_claims.py` (Python, this repository) recomputes every figure in
+   `ANALYSIS.md` into `claims_audit.json`. It also regenerates
+   `native_decomposition.json` and `../phase_a_classification.json` from the
+   May 2026 posting and checks them against the committed copies.
+
+Stage 3 needs only this repository and the QC files. Stages 1 and 2 need the
+pinned external checkouts below.
+
+## Pins
+
+The lab ran on 2026-07-11; the replay output is stamped 23:36Z. The commits
+below were the main-branch heads at that time. On 2026-10-03, reruns at these
+pins reproduced `co_fy2024_reconstruction.csv`,
+`fy2024_reconstruction_national.csv` and `amterr_replay_results.json` byte
+for byte.
+
+| Component | Pin |
+|---|---|
+| giannella/snap_qc | `741e10bf75c36a9ce8ac20f922b5eec3d4539dc2` |
+| snap_qc `qc_data/qc_pub_fy2024.sav` | sha256 `ab6420fa359ab9bcc280a21b9ba7b11172c79c6f7a661718bc6e318f97723fbb` (carries the May 2026 posting's weights) |
+| TheAxiomFoundation/axiom-oracles | `d34aa6fa04287387e0dab6912d128ef746c7b6b2` (merge of #268) |
+| TheAxiomFoundation/rulespec-us | `b53ce208771085030939db4b9691762506b6bca2` (#826) |
+| TheAxiomFoundation/axiom-rules-engine | `de0efdc73b469132ee268e1c832e8f7148b91431` (#102) |
+| Engine release binary | sha256 `bb8ec23689697a5417b74c38196c0488e002e4ee6fe3b33faabb39005e6e5eee`, the 2026-08-06 build of that commit recorded in `../../cert/CERT_REPORT.md` (built there with `cargo build --release --offline`) |
+| FY2024 QC CSV, May 2026 posting | zip `https://snapqcdata.net/sites/default/files/2026-05/qcfy2024_csv.zip` (sha256 `0f3230a4318307d3088382546095eebfde03e781da6f65c9eac7f077bd4263f4`); member `qc_pub_fy2024.csv` sha256 `45193eb7370463ab3067d71da23a580fec34a5460341e4e750dda0be061e1aa9` |
+| FY2024 QC CSV, August 2026 posting | zip `https://snapqcdata.net/sites/default/files/2026-08/qcfy2024_csv.zip` (sha256 `b8b29b8593f78aa51c48332c47d2d92fa5bbecf5346570acb45e26f2d9ebd2b5`); member sha256 `e871a8e9caca0be72e2003b09bdf71e1d020984b52289d2b74c4c6b88c4f793b` |
+| R | 4.3.0 with haven 2.5.5, dplyr 1.1.2, tidyr 1.3.0 |
+
+The binary that ran on 2026-07-11 was not recorded. The 2026-10-03 rerun used
+the 2026-08-06 build of the same commit, attested in `CERT_REPORT.md`. Its
+output matched byte for byte under CPython 3.14 (free-threaded) and 3.13.
+
+The two postings differ only in HWGT, FYWGT, HWGT_OLD and FYWGT_OLD. The
+replay reads weights only for its `weight` field. Run against the August
+posting, it changes that field in 90 of 283 rows and nothing else.
+
+## Commands
+
+Set these once (example paths):
+
+```bash
+export REPO=$PWD                                  # this repository's root
+export LAB=$REPO/paper/snapshot/labs/amterr
+export WORK=$HOME/amterr-rerun                    # any scratch directory
+mkdir -p $WORK
+```
+
+### Stage 3: audit (this repository only)
+
+Download the two zips named in the pins table into `$WORK/may2026/` and
+`$WORK/aug2026/`. The axiom-oracles loader sends a `Referer` of
+`https://snapqcdata.net/datafiles` with its request; the same page lists the
+files for a browser download. USDA replaces postings, so the May 2026 zip may stop
+being served; the hashes identify which posting a file is. Then unzip and
+check the CSV hashes against the table:
+
+```bash
+unzip -o -d $WORK/may2026 $WORK/may2026/qcfy2024_csv.zip
+unzip -o -d $WORK/aug2026 $WORK/aug2026/qcfy2024_csv.zip
+shasum -a 256 $WORK/may2026/qc_pub_fy2024.csv $WORK/aug2026/qc_pub_fy2024.csv
+```
+
+```bash
+SNAP_QC_CSV_MAY2026=$WORK/may2026/qc_pub_fy2024.csv \
+SNAP_QC_CSV_AUG2026=$WORK/aug2026/qc_pub_fy2024.csv \
+uv run --frozen --extra analysis python $LAB/audit_claims.py --check
+```
+
+`--check` exits 1 if the regenerated audit differs from `claims_audit.json`;
+drop it to rewrite the file. The script refuses a CSV whose hash does not
+match its posting. `tests/test_amterr_lab.py` runs the same regeneration
+when both CSVs are present at the default paths (`~/.cache/axiom-oracles/`)
+or at the paths in those variables, and skips it otherwise.
+
+### Stage 1: reconstruction (R, about 70 s)
+
+```bash
+git clone https://github.com/giannella/snap_qc $WORK/snap_qc
+git -C $WORK/snap_qc checkout 741e10bf75c36a9ce8ac20f922b5eec3d4539dc2
+shasum -a 256 $WORK/snap_qc/qc_data/qc_pub_fy2024.sav
+mkdir -p $WORK/recon
+SNAP_QC_REPO=$WORK/snap_qc AMTERR_LAB_DIR=$WORK/recon Rscript $LAB/reconstruct_co_fy2024.R
+cmp $WORK/recon/co_fy2024_reconstruction.csv $LAB/co_fy2024_reconstruction.csv
+cmp $WORK/recon/fy2024_reconstruction_national.csv $LAB/fy2024_reconstruction_national.csv
+```
+
+The script prints `wrote 283 CO error rows; 15902 national error rows` and
+`CO error cases within $5: 246 of 283`. Without `AMTERR_LAB_DIR` it writes
+next to itself.
+
+### Stage 2: engine replay
+
+Build the engine at its pin, then check out the harness and the rules. The
+2026-10-03 rerun used the attested binary in the table; a fresh build on
+another machine may differ in sha256, so compare the replay output, which is
+what the pins have to reproduce.
+
+```bash
+git clone https://github.com/TheAxiomFoundation/axiom-rules-engine $WORK/engine
+git -C $WORK/engine checkout de0efdc73b469132ee268e1c832e8f7148b91431
+(cd $WORK/engine && cargo build --release)
+export BIN=$WORK/engine/target/release/axiom-rules-engine
+
+git clone https://github.com/TheAxiomFoundation/axiom-oracles $WORK/axiom-oracles
+git -C $WORK/axiom-oracles checkout d34aa6fa04287387e0dab6912d128ef746c7b6b2
+git clone https://github.com/TheAxiomFoundation/rulespec-us $WORK/rulespec-us
+git -C $WORK/rulespec-us checkout b53ce208771085030939db4b9691762506b6bca2
+mkdir -p $WORK/replay && cp $LAB/co_fy2024_reconstruction.csv $WORK/replay/
+```
+
+The harness's QC loader hash-checks only a zip it downloads. A CSV it finds
+in `AXIOM_SNAP_QC_DATA_DIR` or `~/.cache/axiom-oracles/snap-qc/` is used
+unchecked, so point it at a directory holding the May 2026 posting's CSV and
+check that file's hash first:
+
+```bash
+shasum -a 256 $WORK/may2026/qc_pub_fy2024.csv
+cd $WORK/axiom-oracles && uv sync --frozen
+cd $WORK/axiom-oracles && \
+  AXIOM_SNAP_QC_DATA_DIR=$WORK/may2026 \
+  AXIOM_SNAP_QC_RULESPEC_ROOT=$WORK/rulespec-us \
+  AXIOM_SNAP_QC_AXIOM_BINARY=$BIN \
+  AMTERR_LAB_DIR=$WORK/replay \
+  uv run --frozen python $LAB/amterr_replay.py
+cmp $WORK/replay/amterr_replay_results.json $LAB/amterr_replay_results.json
+```
+
+The script prints `engine(original) vs RAWBEN <=$5: 246/283 (86.9%)` and
+`solver within $5: 246/283; engine agrees (<=$5) on 246 of those`. Without
+`AMTERR_LAB_DIR` it reads and writes next to itself.

paper/snapshot/labs/amterr/amterr_replay.py

@@ -6,16 +6,17 @@
 the ORIGINAL values — run the engine, and compare the allotment to RAWBEN
 (the benefit the agency actually issued).
 
-Interpretation:
-  match  -> the issuance is explained by correct arithmetic on the original
-            facts: an input/information error, faithfully propagated.
-  miss   -> no plausible original value + correct math reproduces the
-            issuance: computation-side error (or reconstruction failure —
-            correctednotes/at_max separate those).
-
-Run from the axiom-oracles worktree:
-  uv run python /Users/maxghenis/.cache/axiom-oracles/amterr-lab/amterr_replay.py
-with AXIOM_SNAP_QC_RULESPEC_ROOT / AXIOM_SNAP_QC_AXIOM_BINARY set.
+Reading the result:
+  match  -> the engine on the solver's inputs is within $5 of RAWBEN.
+  miss   -> it is not.
+ANALYSIS.md (layer 3) says what each does and does not show.
+
+Run from an axiom-oracles checkout at the commit pinned in README.md:
+  AXIOM_SNAP_QC_RULESPEC_ROOT=<rulespec-us checkout> \
+  AXIOM_SNAP_QC_AXIOM_BINARY=<axiom-rules-engine release binary> \
+  uv run python <this directory>/amterr_replay.py
+AMTERR_LAB_DIR (default: this file's directory) holds the input
+co_fy2024_reconstruction.csv and receives amterr_replay_results.json.
 """
 
 from __future__ import annotations
@@ -49,7 +50,7 @@
 )
 from axiom_oracles.populations.snap_qc import load_qc_units
 
-LAB = Path("/Users/maxghenis/.cache/axiom-oracles/amterr-lab")
+LAB = Path(os.environ.get("AMTERR_LAB_DIR", Path(__file__).resolve().parent)).expanduser()
 
 
 def _num(v):