[MaxGhenis]

Part of #1074. This branch started from #1345's head (ea4606d6b9), and #1345 merged into main on November 4, so this PR now targets main.

Re-pins the US Chronicle feed so it carries Federal Student Aid's loan portfolio tables, which #1357 binds as opt-in calibration targets. Every row of #1345's feed stays the same, the re-pin adds 461 ed_fsa rows, and the default compile's 33,250 targets don't change.

The Chronicle cut

41b2fb2b864a28c0eda03322b77a97ef147dfa3c, on Chronicle branch feed/us-a9026cc-ed-fsa-1103, tag microcosm-us-feed-1074-v1. It cherry-picks chronicle#347's commit 4448a7c96 onto #1345's pin a9026cc8f (tag microcosm-us-feed-1343-v1) and changes nothing else. Chronicle main doesn't contain the cut. Chronicle#347's commit also brings a Census ACS S1501 state package, which the scope leaves out. A separate session cut the commit and built the feed (Chronicle branch hub-notes/ed-fsa-feed-cut-1103 at bf3456f9, notes/ed-fsa-feed-cut-1103/). This PR uses that session's feed and scope, after these checks:

Check	Result
Feed sha256 (gunzipped)	1d45a89c28177a95371b4cc4068d3657c7bf12bd8edad7c82fe25bf453f9ea07
Lines	40,642
schema_version	chronicle.consumer_fact.v3 on all 40,642 rows
Scope sha256	73a39c166a8f78c8a8a1ae711515e5d2e6b76a74952e67d7783cf974a4615633
Feed with the 66 new pairs' rows dropped	40,181 lines, sha256 4d89e4a498ffd73a407821d9ce2c45837072901e0a29acf3206160d1fafd7c14 (#1345's feed)

What changes

• us/chronicle_feed_scope.json: the notes' scope, byte for byte. It has 706 pairs: Re-pin the US Chronicle feed so three dimension values carry one label each #1345's 640 plus 66 ed_fsa pairs, all at build year 2026, over 72 package runs. No microcosm test needed a different order or format.

• us/chronicle_feed.json:

  • source_commit;
  • the build text;
  • fact_row_count 40,642;
  • facts_sha256 1d45a89c…;
  • scope_sha256 73a39c16….

  consumer_fact_schema_versions stays ["chronicle.consumer_fact.v3"] and consumer_fact_schema_sha256 stays bdb51e2a… (see below).

• target_parity_manifest.json and target_parity_feed_families.json, regenerated:

  • they name the new feed (consumer_facts_us_41b2fb2.jsonl) and its digest;
  • the feed now has 93 families, 11 of them new ed_fsa.* families;
  • tools/build_us_target_parity_manifest.py declares those 11 off_by_default reviewed exclusions, under one fence for the ed_fsa namespace;
  • the files count 32 compiled families and 64 reviewed exclusions, up from 53;
  • the label variants don't change. Judgment call 1 says why these files change.

• Tests:

  • test_us_chronicle_feed.py: _CHRONICLE_COMMIT, _FACTS_SHA256, the row count, the scope's pair and run counts, and the exclusion count. _SCHEMA_SHA256 and the v3 schema-versions assertion stay the same and still hold;
  • test_release_target_parity.py: the exclusion count, plus a test that the manifest lists all 11 ed_fsa families as off_by_default reviewed exclusions behind the microcosm#1074 fence;
  • test_us_esi_premiums.py: the report chain, below.

• experiments/us-feed-ed-fsa/: compare.json, plus added_rows.json and the script that writes it (list_added_rows.py).

• docs/us-chronicle-feed-repin.md: a new section and the pin table. Changelog fragment: changelog.d/1074-us-feed-fsa-loan-portfolio.added.md.

Compare

experiments/us-feed-agi-age-labels/compare_feeds.py (#1345's script; follow-up 4 corrects its docstring) ran with #1345's feed as the old feed and the notes' feed as the new one, and wrote experiments/us-feed-ed-fsa/compare.json. The script compiled both feeds at 2024, as the release does.

• Rows kept: all 40,181. compare.json lists no keys_only_in_old, counts 0 changed_rows and lists no changed_fields, so every old row keeps its key, values and labels.

• Rows added: 461, all from source ed_fsa. They match the ed_fsa rows of the notes' ed_fsa_rows.jsonl one for one:

  Package	Pairs	Rows
  dl_portfolio_by_repayment_plan_fy2026_q3	36	227
  portfolio_by_age_fy2026_q2	4	24
  portfolio_by_debt_size_fy2026_q2	4	36
  portfolio_by_loan_type_fy2026_q3	12	42
  portfolio_by_location_fy2026_q2	1	108
  portfolio_summary_fy2026_q3	9	24

• Targets:

  • 33,250 on both feeds, in the same order, with no changed field;
  • registry version b8d571e91c22 on both;
  • national_state surface: 6,102 targets, 3aa2a0830e7d, on both;
  • no compiled label variants on either.

• Feed label variants: age_20 and age_21, as before.

• Memory: the compile step finished without running out of memory, and the whole script took about 2 minutes.

Why the schema versions stay v3

• The stamp. Every row of the feed, old and new, carries chronicle.consumer_fact.v3. At the cut, the exporter writes every row's stamp from one constant, CONSUMER_FACT_SCHEMA_VERSION. That constant takes its value from the last entry of CONSUMER_FACT_ROW_CONTRACTS (chronicle/consumer_contract.py:35, chronicle/epoch.py:171-176), and the tuple ends at v3. Both files match byte for byte at a9026cc and 41b2fb2b; I read them through raw.githubusercontent.com.
• The schema digest. Chronicle computes consumer_fact_schema_sha256 as the sha256 of policyengine_chronicle/schemas/consumer_fact.v3.schema.json (policyengine_chronicle/schema.py:59-70). That file hashes to bdb51e2a… at both commits, so the digest doesn't change from Re-pin the US Chronicle feed so three dimension values carry one label each #1345's.
• What microcosm does with the stamp, on this branch:
  • _load_fact_rows carries a row's stamp without gating on it (ledger_artifact.py:464-471);
  • fact_schema_versions only reports the stamps (:207-219);
  • the US pin loader checks only that consumer_fact_schema_versions holds a list of distinct, non-empty strings (us_runtime/chronicle_feed.py:136-143);
  • the release tool loads the bare feed with expected_facts_sha256, which checks only the file's hash;
  • validate_chronicle_feed_pin (ledger_artifact.py:257-291) would compare the list with the rows' stamps (:276, :287). It runs only on the transport kernels and the BE and NZ reference tools, never on the US release path.
• Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's code doesn't read the stamp. I grepped fsa_student_loan_targets.py, ledger_targets.py, target_aging.py, Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's diff to fiscal_targets.py and the release tool, and its FSA test file, all at 0ff293b1a. The grep finds schema_version only in the FSA receipt's own version (fsa_student_loan_targets.py:441) and a target profile's (ledger_targets.py:756-765).

So the pin lists only v3. #1357's body expected v4 FSA rows and a pin listing both versions; that would need Chronicle's 0ec21e5 in the cut, which would restamp every row v4.

The report chain test

test_the_committed_feed_pin_banks_the_anchor follows the committed re-pin reports from #454's feed to the pin. Before this PR it refused any report that added rows (keys_only_in_new == [], old_rows == new_rows), so this re-pin would have failed it. This PR moves the per-report check into a helper, _follow_later_repin_report:

• Rows a report may add. A report adds rows only when its entry in _LATER_ROW_KEEPING_REPIN_REPORTS names a committed listing of the added rows (added_rows.json), one source and one record-set namespace ("ed_fsa." here). The listing must name the same keys the report adds, each from that source, outside the B4923C anchor record sets and in a record set whose id starts with the namespace. new_rows must equal old_rows plus the number of added keys.
• The other checks stay: a report must start at the previous feed, keep every key, change only labels and record_set_spec_hash, and leave the anchor record sets alone.
• added_rows.json. compare.json doesn't record where an added row comes from, so list_added_rows.py reads the added keys back out of the new feed, with their source.source_name, record set and period.
• Tests in each direction. One test accepts rows added from the declared source. Seventeen cases refuse, each at the check it names:
  • rows added without a declaration;
  • a row from another source;
  • a listing that misses a key, and one that names a key the report doesn't add;
  • a row in an anchor record set;
  • a listing with no source, and one with no namespace;
  • a row outside the namespace, and one whose record set holds ed_fsa. after its first component;
  • a listing of another feed, and one that repeats a key;
  • dropped rows, a wrong new row count, a value change, a wrong base, an old feed of another size, and a label change in an anchor record set.

Default compile

On this branch, which doesn't have #1357's code, the default compile turns no ed_fsa fact into a target. I ran:

• compile_us_fiscal_target_registry(age_targets=True, packaged CD crosswalk) and apply_us_medicaid_enrollment_substitutions on the new feed;
• the release tool's own checks, called from the tool's module: _check_committed_us_ledger_feed_pin, _check_us_esi_anchor_against_feed (no mismatch), assert_target_parity_manifest_current(registry=…) and us_release_target_parity_gate.

It gives 33,250 targets in 32 families, none of them from ed_fsa, and the parity checks pass against the regenerated resources. The pinned-feed parity tests (test_committed_artifacts_match_regeneration, test_gate_passes_on_real_compiled_registry) ran rather than skipped and pass. So the compiled families stay at 32.

Checking against #1357 (scratch branch, not pushed)

I merged #1357's head, 0ff293b1a335a29d46f1598f2039beffa242e632, into a local scratch branch off this PR.

• Merge. It conflicts in seven files:

  • docs/evidence/spec-engine/us-f0-coverage.json;
  • spec_engine/field_usage.py and spec_engine/inventory_coverage.py;
  • test_spec_engine_coverage_tool.py, test_spec_engine_field_usage.py and test_us_multispine_pool_tool.py;
  • tools/spec_engine_coverage.py.

  git merge-tree gives the same seven conflicts for Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357 against main alone, and against Re-pin the US Chronicle feed so three dimension values carry one label each #1345, so they come from Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's stack against main, not from this re-pin. Neither this PR nor Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's FSA code touches any of the seven. In the scratch branch I resolved all seven to our side.

• Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's FSA tests. I ran its FSA test file, test_us_fsa_student_loan_targets.py: 25 passed. The selector -k 'fsa or student_loan or ed_fsa' matches no test names in Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's other edited test files, because its edits sit inside existing tests. So I ran those files whole in scratch:

  • test_provider_labels 4 passed;
  • test_variable_labels 3 passed;
  • test_us_spine_blindness 504 passed;
  • test_us_fiscal_targets 174 passed;
  • from test_us_fiscal_refresh_builder, test_restated_concept_rules_all_have_a_comparison: 1 passed.

  This PR's three changed test files also pass in scratch: 28, 39 and 159.

• Compile with student_loan_targets=True against this pin and the notes' feed, which I loaded with the pin's digest and checked with _check_committed_us_ledger_feed_pin:

  • 33,382 targets, 132 FSA: 67 balance (65 total_balance, 1 parent_plus_balance, 1 balance_excluding_parent_plus) and 65 borrowers;

  • the FSA receipt reads status bound, reference period 2026-03, over 461 FSA facts;

  • its dispositions match Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's table, and the compile binds the other 136 facts:

    Disposition	This feed	Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's table
    diagnostic	251	251
    other_quarter	58	58
    excluded	8	8
    allocated	4	4
    out_of_scope	4	4

  • _without_opt_in_student_loan_targets takes the registry back to 33,250 targets, and assert_target_parity_manifest_current and the parity gate pass on it;

  • with the flag off, the scratch compile also gives 33,250 targets and no FSA target.

• Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's fixture (us_fsa_portfolio_chronicle_347_facts.jsonl.gz, sha256 9faed0e3…, 244 rows, all v4):

  • none of its rows matches a notes row byte for byte;
  • all 244 share an aggregate_fact_key with a notes row, differ from it only in schema_version, and match it byte for byte once the notes row reads v4.

  In scratch I restamped the fixture's 244 rows v3, and they then match the notes' rows byte for byte. On the restamped fixture:

  • 25 passed, as with v4, so no result changes;
  • the FSA receipt from the v3 fixture equals the one from the v4 fixture, field for field;
  • the 132 targets and the controls from the fixture equal those from the full v3 feed.

Tests at c624e6f (with the feed in place)

I put the feed at ~/PolicyEngine/_buildh-runtime/inputs/consumer_facts_us_41b2fb2.jsonl, where tools/build_us_target_parity_manifest.py and the pinned-feed fixtures look for it, and pointed MICROCOSM_US_CHRONICLE_FACTS at it. Each file ran in its own process (-n 4, -p no:cacheprovider). Only test_us_fiscal_refresh_builder.py skipped any tests, and its 6 skips need the policyengine-us extra (:55), not the feed:

File	Passed
test_us_chronicle_feed.py	28
test_release_target_parity.py	39
test_us_esi_premiums.py	159
test_us_fiscal_targets.py	174
test_us_soi_congressional_district_vintage.py	51
test_us_soi_state_agi_bands.py	29
test_us_medicare_irmaa_holdout.py	24
test_us_acs_local_sparse_cd.py	28
test_us_release_gate_preflight.py	84
test_us_spec_bundle.py	4
test_us_acs_local_release_tool.py	215
test_us_fiscal_refresh_builder.py	346, and 6 skipped

ruff check . passes and tools/ci_test_plan.py verify reports ok. No test I ran says it peaks near 15 GB. CI covers the full suites.

Merge order

1. Re-pin the US Chronicle feed so three dimension values carry one label each #1345 merged into main on November 4.
2. This PR now targets main.
3. Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357. Its targets bind only in a Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357 build that passes --scf-education-debt (student_loan_targets=True); a default build compiles none of them.

#1357 has to resolve the seven spec-engine conflicts against main whatever the order. This PR already takes the step #1357's body lists for the re-pin: the generator declares the FSA families reviewed exclusions, and _without_opt_in_student_loan_targets keeps them off the parity contract.

Judgment calls

1. I regenerated the parity resources. Since no ed_fsa fact compiles by default, the resources could have stayed the same, but then CI would fail:

   • test_parity_resources_and_generator_restate_the_pin (test_us_chronicle_feed.py:120-132) requires both files to restate the pin's facts_sha256;
   • the generator refuses a feed family that neither compiles nor has a fence (_exclusion_for), so it can't run on this feed without one.

   No family's status changed: 32 compiled before and after. The generator classifies the 11 new families off_by_default, the classification TargetFamily documents. This PR takes the third re-pin step in Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's body. Without the fence this PR can't pass CI, so the fence can't wait for Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357.

2. The ed_fsa fence's origin says this re-pin did not search the retired us-data pipeline for a student loan target.

3. DEFAULT_FEED_NAME reads consumer_facts_us_41b2fb2.jsonl, following Re-pin the US Chronicle feed so three dimension values carry one label each #1345's naming.

4. The chain test. I left the logic of Re-pin the US Chronicle feed so three dimension values carry one label each #1345's compare_feeds.py alone; follow-up 4 corrects only its docstring. The added rows' sources come from a separate listing, added_rows.json.

5. The pin's build text still says chronicle build-consumer-artifact refuses the bare feed "at this commit". I didn't run that command at 41b2fb2b. The v3 schema file hasn't changed, and the feed holds the same 994 rows the doc gives as the reason (the concept_alignment rows with no authority: 515 from cms_medicaid, 468 from census_acs, 11 from jct).

6. What the doc takes from the cut session's notes. It attributes three things to them, which I did not rerun: the acbedb5 export, 0ec21e5, and the 217 FSA rows the fixture lacks.

7. Branch. No microcosm branch us-feed-ed-fsa or us-feed-ed-fsa-1103 existed, so I created us-feed-ed-fsa-1103.

Review follow-ups

An independent review of c624e6f approved this PR with four follow-ups. Commits 395d1cc..8982b8f make them:

1. The chain test ties each listed row to a record-set namespace. Each declaration in _LATER_ROW_KEEPING_REPIN_REPORTS now names the namespace its added rows must sit in ("ed_fsa." here). _follow_later_repin_report refuses a listed row whose record_set_id doesn't start with it, after the anchor check. Before this, a listing could label an irs_soi row ed_fsa and pass. 395d1cc commits that case failing, and af45400 makes it pass. A no-namespace case covers a declaration that leaves the namespace out. 8353e34 adds cases for two checks that no case reached: a listing of another feed, and a listing that repeats a key. Both already refused.
2. The parity test's docstring says which families bind. Only four of the 11 ed_fsa families bind under Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's opt-in: location, age, debt size and loan type. The portfolio summary and the six repayment-plan and IDR tables stay diagnostic in every build, and so do the loan-type table's recipient counts. The single namespace fence stays. Giving the seven diagnostic families their own exclusions waits for Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357, which owns their reasons.
3. The doc says how many facts bind. Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's student_loan_targets=True binds 136 of the 461 ed_fsa facts. The doc's paragraph on the anchor test now names the namespace too.
4. compare_feeds.py's docstring says what the script does. It reports keys_only_in_old and keys_only_in_new; it doesn't require both feeds to hold the same keys.

With the feed in place at 8982b8f, test_us_esi_premiums passes 163 (159 before, plus four new cases), test_release_target_parity 39 and test_us_chronicle_feed 28. The follow-ups don't touch the pin, the scope, the feed, the parity resources or the changelog fragment.

A second independent review approved 8982b8f and found two code changes that left the later-report tests passing: matching the namespace anywhere in a record set id instead of at its start, and moving the namespace check ahead of the anchor check. A mutation run over every check in _follow_later_repin_report found six more:

• moving the namespace check ahead of the source check;
• deleting the old-row-count check, the changed-field anchor check, the declared-source check, the source check or the added-row anchor check.

1f25420 gives every check in the helper a name in its assertion message and has each refusing case name the check that must refuse it, through pytest.raises(..., match=...). It adds three cases: a record set that holds ed_fsa. after its first component, an old feed of another size, and a label change in an anchor record set. The helper's logic stays the same, so the new cases pass at once. Each of the eight changes now fails at least one case, under pytest's assertion rewriting and with --assert=plain. 212dd72 rewords the doc's sentence on the 136 facts.

A third independent review approved 212dd72 and found two more gaps in the cases. Each anchor case holds one record set, so a check that read only the first would still refuse it. Each refusing count exceeds the expected count and each refusing digest sorts after the expected one, so a check that compared with <= would still refuse them all. The helper reads every record set and compares with ==; #1513 adds cases for both gaps.

---

🤖 Generated with Claude Code

[MaxGhenis]

Ready for hub review

Head: c624e6fe5038ee2db19ee3a19773a92b570ae945 (two commits on #1345's ea4606d6b9).

Inputs. All checks passed: the five notes files fetched at bf3456f9. The gunzipped feed is 1d45a89c…, 40,642 lines, every row chronicle.consumer_fact.v3. The scope is 73a39c16…. With the 66 new pairs dropped, the feed is 40,181 lines, 4d89e4a4…, which is #1345's feed.

The re-pin.

• The scope is copied byte for byte; no test needed a different order or format.
• The pin changes source_commit to 41b2fb2b, the build text, fact_row_count to 40,642, facts_sha256 and scope_sha256.
• consumer_fact_schema_versions stays ["chronicle.consumer_fact.v3"].
• consumer_fact_schema_sha256 is the sha256 of Chronicle's policyengine_chronicle/schemas/consumer_fact.v3.schema.json (policyengine_chronicle/schema.py:59-70). That file is bdb51e2a… at both 41b2fb2b and a9026cc, so the digest doesn't change from Re-pin the US Chronicle feed so three dimension values carry one label each #1345's.
• In the test, _CHRONICLE_COMMIT and _FACTS_SHA256 change and _SCHEMA_SHA256 doesn't. The v3 assertion still holds.
• I added a doc section and a changelog fragment.

Stamp readers on #1345's branch. All five checks read as the brief says:

• ledger_artifact.py:464-471 carries the stamp without gating on it;
• :207-219 reports it;
• us_runtime/chronicle_feed.py:136-143 checks only the list's form;
• validate_chronicle_feed_pin (:257-291, :276, :287) runs only in the transport kernels and the BE and NZ tools.

I found no other code that reads the stamp. #1357's FSA code has no read either: its only schema_version hits are its receipt's own version and a target profile's.

Compare. experiments/us-feed-ed-fsa/compare.json:

• 40,181 rows kept, 0 changed;
• 461 added, all source ed_fsa: 227, 24, 36, 42, 108 and 24 by package;
• 33,250 targets with 0 changed, the same registry version (b8d571e91c22) and the same national_state surface (6,102 targets, 3aa2a0830e7d);
• no out-of-memory error.

Report chain test. It used to refuse added rows. It now accepts rows added from one declared source, checked against a committed listing (added_rows.json), and nothing else. There is one accepting test and ten refusing cases.

Default compile. No ed_fsa fact compiles: 33,250 targets in 32 families. The release tool's pin, ESI-anchor and parity checks pass.

#1357 scratch merge (head 0ff293b1a, never pushed):

• Seven spec-engine files conflict. git merge-tree shows the same seven conflicts merging Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357 into main alone, so they come from Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's stack. I resolved them to ours in scratch.
• test_us_fsa_student_loan_targets.py passed 25. -k 'fsa or student_loan or ed_fsa' matches no test names in Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's other edited test files, so I ran those whole: 4, 3, 504, 174, and 1 in test_us_fiscal_refresh_builder. This PR's own test files also pass there.
• student_loan_targets=True compiles 132 FSA targets: 67 balance and 65 borrowers. The receipt's dispositions are 251 diagnostic, 58 other_quarter, 8 excluded, 4 allocated, 4 out_of_scope and 136 bound, from 461 facts, which matches Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357's table. Parity passes with the opt-in filter.
• The fixture: 0 of its 244 rows match the notes' rows byte for byte, and all 244 match once restamped v4.
• I restamped the fixture v3 in scratch. Then 25 passed, the same as before, and the receipt is identical. No result changed.

Tests on this branch, with the feed at ~/PolicyEngine/_buildh-runtime/inputs/consumer_facts_us_41b2fb2.jsonl and MICROCOSM_US_CHRONICLE_FACTS set:

File	Passed
test_us_chronicle_feed	28
test_release_target_parity	39
test_us_esi_premiums	159
test_us_fiscal_targets	174
test_us_soi_congressional_district_vintage	51
test_us_soi_state_agi_bands	29
test_us_medicare_irmaa_holdout	24
test_us_acs_local_sparse_cd	28
test_us_release_gate_preflight	84
test_us_spec_bundle	4
test_us_acs_local_release_tool	215
test_us_fiscal_refresh_builder	346

The only skips are 6 in test_us_fiscal_refresh_builder, which need the policyengine-us extra. ruff check . and ci_test_plan.py verify both pass. No test I ran says it peaks near 15 GB. GitHub CI hasn't run, because this PR's base isn't main.

Judgment calls. The body has the full list:

1. I regenerated the parity resources, against the brief's "stay as they are". test_us_chronicle_feed.py:120-132 requires them to restate the pin's digest, and the generator halts on an unfenced family. The 11 ed_fsa families are off_by_default exclusions. 32 families stay compiled, and exclusions go from 53 to 64.
2. The ed_fsa fence says this re-pin did not search us-data for a student loan target.
3. DEFAULT_FEED_NAME is consumer_facts_us_41b2fb2.jsonl.
4. I left Re-pin the US Chronicle feed so three dimension values carry one label each #1345's compare_feeds.py unchanged and added list_added_rows.py to record the added rows' sources.
5. The pin's "bare feed: build-consumer-artifact refuses it" wording is kept but wasn't rerun at 41b2fb2b. The same 994 rows without authority are present.
6. The doc credits the notes for the acbedb5 export and 0ec21e5 findings.
7. The branch is us-feed-ed-fsa-1103. Neither name the brief mentions existed.

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review

Head: 8982b8fcc438687aa9eb3d8ded5f638e6e11f725. It adds four commits to c624e6f as a fast-forward. The branch head was still c624e6f before I pushed.

Setup. I fetched the feed from Chronicle bf3456f9 (notes/ed-fsa-feed-cut-1103/consumer_facts.jsonl.gz) and gunzipped it to ~/PolicyEngine/_buildh-runtime/inputs/consumer_facts_us_41b2fb2.jsonl. It is sha256 1d45a89c…f9ea07, with 40,642 lines. MICROCOSM_US_CHRONICLE_FACTS pointed at it.

1. The listing's record sets (finding 1): done.

• 395d1cc adds the outside-namespace case to test_a_later_report_refuses_rows_added_outside_its_declaration. In it, the listing labels a row ed_fsa but its record set is irs_soi.ty2023.table_1_4. The case failed with DID NOT RAISE (1 failed, 6 passed).

• af45400 gives each declaration a third element, the namespace ("ed_fsa." for the FSA re-pin; the type is at test_us_esi_premiums.py:1114 and the entry at :1123). The chain test unpacks it (:1255) and passes it through (:1264). _follow_later_repin_report takes declared_namespace and asserts that every listed record_set_id starts with it (:1178-1183). That assert comes after the anchor check (:1175-1177).

• 8353e34 adds listing-other-feed and listing-repeats-key (test_a_later_report_refuses_a_listing_it_cannot_follow). Both refused before this change, so they have no failing commit.

• Where each case refuses at the head. I ran a scratch harness that calls the helper directly:

  Case	Assert
  undeclared-additions	:1165
  no-source	:1167
  no-namespace	:1168
  listing-other-feed	:1169
  listing-repeats-key	:1171
  listing-short, listing-extra	:1172
  other-source	:1174
  anchor-record-set	:1175
  outside-namespace	:1183
  wrong-base	:1157
  dropped-row	:1159
  value-change	:1161
  row-count	:1184

• The review's own script on the committed files. I reran the review's scripts/chain_mutations.py against copies of the committed files in scratch:

  • Its case add-non-ed_fsa-row (listing claims ed_fsa, pin row count +1) passed at c624e6f. It now refuses at :1183.
  • The unmutated chain passes.
  • Every other mutation still refuses at the same check as before.
  • Its old two-element declarations now fail the three-element unpack with a ValueError (:1255), because the declaration's shape changed.

2. The parity test's docstring (finding 2): done (8982b8f). test_release_target_parity.py:346-357 (the docstring of the test at :345) now says that only four of the 11 families bind under #1357's opt-in: location, age, debt size and loan type (FSA_BOUND_TABLES, fsa_student_loan_targets.py:153 at 0ff293b). The portfolio summary and the six plan and IDR tables stay diagnostic in every build (FSA_DIAGNOSTIC_TABLES, :168-199), and so do the loan-type recipient counts (:29-31). I checked those three citations at 0ff293b. I also checked that the manifest's 11 ed_fsa families are exactly those 4 + 7 tables. The one namespace fence stays, and the test's asserts are unchanged.

3. The doc (finding 3): done (8982b8f). docs/us-chronicle-feed-repin.md:60-62 now says #1357's opt-in "binds any of them, 136 of the 461". The 136 comes from the review's step 7 log (logs/step7_opt_in_compile.log:10). I didn't rerun #1357's compile.

4. compare_feeds.py's docstring (finding 4): done (8982b8f). :7-9 now says that the script reports the row counts, keys_only_in_old and keys_only_in_new. That matches compare_rows (:145-146). docs/us-chronicle-feed-repin.md:440 is untouched.

5. The runs. Each file ran in its own process, with -n 4 -p no:cacheprovider and the feed in place:

File (packages/microcosm-build/tests/engine_free/…)	c624e6f	8982b8f
us/test_us_esi_premiums.py	159 passed	163 passed
shared/test_release_target_parity.py	39 passed	39 passed
us/test_us_chronicle_feed.py	28 passed	28 passed

The four new tests are outside-namespace, no-namespace, listing-other-feed and listing-repeats-key. Nothing skipped or failed. At 8982b8f:

• ruff check .: all checks passed;
• ruff format --check on the three changed .py files: already formatted;
• python tools/ci_test_plan.py verify: verification=ok.

6. The PR body and the changelog. I added a "Review follow-ups" section at the end of the body and changed nothing else in it. The changelog fragment says nothing these fixes change, so it is unchanged.

Judgment calls

1. A namespace-declared assert and a no-namespace case. I added an assert that the namespace is declared (:1168) and a case for it. Without the assert, a declaration that leaves the namespace out would fail with a TypeError from startswith(None) rather than an assert. The case can't fail before the fix, because the parameter didn't exist yet, so it comes in the fix commit.
2. listing-extra now uses ed_fsa.a and ed_fsa.b. Its record sets were a and b. With the new assert, those would also sit outside the namespace. They now sit inside it, so the case refuses only because the keys differ. It refuses at :1172 either way.
3. The anchor-test paragraph in the doc. I also updated docs/us-chronicle-feed-repin.md:89-96, which describes what the chain test checks, so that it names the namespace. The brief didn't list it, but the fix changes what that paragraph states.
4. Where the two coverage cases live. They are in a new parametrized test that mutates the listing, because their mutations don't fit the existing test's row-tuple parameters.
5. The pin, the scope, the feed, the parity resources and the changelog fragment are unchanged. So are the manifest's ed_fsa classifications; splitting the seven diagnostic families waits for Calibrate federal student loan balances to Federal Student Aid's portfolio (#1074) #1357.

---

Generated by Claude Code

[MaxGhenis]

I merged 212dd72 into main as merge commit ad67d0b. All 11 CI jobs passed at that head (run 41855432535), and GitHub read the branch as mergeable and clean. Three independent reviews approved it: one of c624e6f, and two GPT-6.1 Sol reviews, of 8982b8f and 212dd72. Each of 18 changes to _follow_later_repin_report fails at least one later-report case at 212dd72, under pytest's assertion rewriting and with --assert=plain. The third review found six more that pass, and #1513 adds cases for them.