Repository navigation
Conversation
The fixture registered its two recipient kernels into the issued financial run's own registry, which the run seals at issuance and re-checks in check_atomic_survey_financial_run, so every test errored with FINANCIAL_KERNEL_REGISTRY_CHANGED. Production (graph_survey_puf55) composes a private KernelRegistry over the run's kernels; the fixture now does the same. The last-owner mutation test also accepts the node-population refusal that now fires first, as its siblings do. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…time extended_status_host recorded its student _MEMBER_PINS patch on the outer MonkeyPatch context after the generator had already patched the same attribute, so on exit it restored the generator-era pins over the value the module's first fixture installed, and that fixture's retained student receipts failed STUDENT_CONTENT_CHANGED at its own teardown (attributed by pytest to the last test, tax42). The patch now lives in its own context nested inside the generator's lifetime, so undo order is LIFO. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
One rest job and both wheels 2/4 jobs hit GitHub's six-hour limit at 76-79% of their shards (the alphabetically late survey-graph modules cluster there); the partition is by file count, so a few heavy modules dominate a job. Six and eight shards spread the heavy modules 13 and 9-10 per job instead of up to 23, and every sharded pytest run now prints --durations=25 so the next imbalance is visible in the job log. tools/ci_test_groups.py --verify and its regression test pass. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…erified-lanes-20260910 # Conflicts: # PROGRESS.md
…erified-lanes-20260910 # Conflicts: # packages/microcosm-build/src/microcosm/build/uk_runtime/__init__.py
…erified-lanes-20260910 # Conflicts: # packages/microcosm-build/tests/test_spec_engine_country_bundles.py
Main's #896 resolves the local git code pin before any I/O in run_uk_calibration, so four tests that previously refused or read back before the pin now reach git rev-parse. In the installed-wheels CI job the package lives under /tmp/wheels-venv, where no repository exists, and test_readback_failure_updates_build_record_and_failed_attempt failed with "Could not resolve the local git code pin". Give those four tests the existing invented_code_pin fixture, as their siblings already have; the unresolved-pin refusal test keeps its own failing pin. Verified: the file passes 23/23 in the repository and 23/23 with a git that exits 128 on PATH, which reproduces the wheels job condition. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Ship the native property and tax graph without the child-property completion, keep the native CPU cap and remove the repeated-verification cost instead of buying it, and make a from-scratch build plus certification of the default the merge gate for this integration and a standing check on main. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Records the lane's base, goal, evidence base and the two local constraints: root out.md belongs to another lane, and the v5 cold run holds the machine. Co-Authored-By: Claude Opus <noreply@anthropic.com>
_update_series boxed every value of a plain float, integer or boolean column into a Python object and framed them one at a time; on pandas 3.0.3 a plain column is a NumpyExtensionArray with no _data/_mask, so the fast masked path never applied to it. _object_stream now builds the identical framed bytes -- the same 'object' dtype header, the same shape header, the same per-value length prefixes and payloads -- with vectorised numpy, and returns None for every column kind it cannot reproduce exactly, including a Categorical whose values are an integer-coded view. No digest value moves. test_graph_executor_series_stream compares the live helper against a verbatim copy of the pre-change body at the byte level over every column kind, the float specials, non-canonical NaN payloads, both int64 endpoints and a 200-frame random sweep. Measured 0.47 s -> 0.09 s on a 6,928 x 240 frame. Co-Authored-By: Claude Opus <noreply@anthropic.com>
docs/us-native-verification-once.md states, per mechanism, the guarantee today, the change, why the guarantee survives and the new cost class -- and the two residuals the epoch's unconditional final re-validation exists to close. Co-Authored-By: Claude Opus <noreply@anthropic.com>
_records read the 2.4 GB of ACS person CSV inside csv_pus.zip one Python byte at a time; with peek and columns.lines it was 57-65% of the admission phase in every measured native run. It now finds terminators with bytes.find, counts quote parity with bytes.count, and keeps each terminator cursor until the scan passes it -- so an absent carriage return costs one search per 1 MiB block rather than one per record, which is where the bulk of the win is. The two ceilings are charged by _fence. A record no longer than the smaller of the live ceilings cannot have violated either, because the token counter resets at every record start; anything longer is replayed byte by byte, so the refusal code and the byte it fires on stay the fence's own -- including the LF that closes a CRLF record, which the fence charges against the record ceiling only. test_us_acs_record_fence_scan compares the shipped fence against a verbatim copy of the byte loop: every boundary case, every block-boundary case, every ceiling on both sides, 4,500 randomised strings, 300 row-shaped inputs, and all 9,331 strings up to five bytes over the branching alphabet under each of five ceiling settings -- 46,655 exhaustive comparisons. Measured on the real staged archive: 10.8 MB/s -> 522.8 MB/s, digests identical. Re-pins acs_native_coverage_binding._ACCEPTED for the changed module: 475aa795c8a5b49a0dd3405a0877866dddd03012a1f2b9743447fab1fe85bcff -> 9ec68721a4cf480ef412c51ab354db9000eb7a7989e35e6e09b1574d88e8e49f (shasum -a 256 .../acs_person_coverage_authentication.py, after formatting) The module's graph_implementation_inventory contract is unchanged: the rewrite adds no _RESOURCE_CALLS name and no non-stdlib import. Co-Authored-By: Claude Opus <noreply@anthropic.com>
executor.py re-derived source_content_key for every source a cold-executed node declared, and keys._directory_identity read_bytes()es every file of the tree: O(executed source-declaring nodes x source bytes), which on the native graph is the 3.48 GiB source tree re-read per node. run_graph now carries a _SourceIdentities cache in its own locals, populated by the run-start pass as it reads. A cached key is reused only while the path's stat signature -- (dev, ino, size, mtime_ns, ctime_ns) for a file, and for a directory its own identity plus the relative name, type and identity of every entry _directory_identity would walk -- is identical to the signature taken both immediately before and immediately after the read that produced the key. A derivation whose two signatures disagree is not cached at all. source_content_key itself stays pure and uncached: three key and codec tests call it twice on one path with the bytes changed in between. Before the manifest is built, every source is re-derived in full with the cache bypassed, written inline rather than through _source_paths_and_keys because a build test counts calls to that code object. That closes the two cases a stat signature cannot decide, and gives the executor a guarantee it never had: a source changed during a node that declares none is now caught. RunManifest gains source_identities, attached exactly like populations and mass_ledgers -- repr=False, compare=False, outside content_addressed, outside to_json, outside every node receipt and cache record. test_graph_executor_source_identity pins the contract that had no test at all: a source rewritten by its own node refuses and writes nothing to the store, a file added to or removed from a directory source refuses, a same-length rewrite with mtime restored refuses, a change during a source-free node refuses at run end, and an unchanged source is read twice per run however many nodes declare it. The graph suite falls from 108 s to 72 s. Also records the real-archive record-fence parity receipt: both ACS person members, 2.4 GB, 3,422,890 records, identical digests, 226.9 s -> 4.6 s. Co-Authored-By: Claude Opus <noreply@anthropic.com>
Every borrow of AuthenticatedSurveyPopulationPreparation re-ran the whole authentication: both source catalogues, the ACS native coverage binding (eight full zip reads), the nested ASEC native population (seven files), the ten-file roster re-hash and every pure seal -- once per executed node through the population observer, and again inside every kernel that borrows it. The ASEC native capsule did the same on every .frame access. survey_population_preparation.verification_epoch() makes that happen once. Inside an epoch each borrow still pays a cheap tier in full -- the live authority and attached owner payloads, _encode(_producer()), and _file_stats over the whole roster -- and the expensive tier is skipped only while a signature is unchanged: the stat identity of every path the four foreign verifications and _source_files re-read, the identity and length of every borrowed payload, and a read-free witness of every live Frame's storage (buffer address, shape, strides, dtype and writeable flag per column, index and weight). A signature that moved is a memo miss, not a refusal: the complete validation runs and raises exactly what it would have raised. Leaving the epoch, at every nesting level, re-validates every memoised capsule in full with the memo bypassed, so a change no signature can see -- an in-place write into a live buffer, or a value written into a frozen plan row -- still makes the run refuse before the caller receives anything. Outside an epoch nothing is memoised and every existing borrow, refusal and trace-based mutation test keeps its exact meaning; _validate and _validate_state are untouched. The three foreign owners gain only a _verified_source_stats() helper naming the paths their own verification re-reads. Their inventory contracts do not move. The two capsule modules' unbound_uses digests do, because the new functions use already-imported names in new scopes; both are re-derived through graph_implementation._dependency_contract and recorded in PROGRESS.md. Co-Authored-By: Claude Opus <noreply@anthropic.com>
run_atomic_survey_population and run_atomic_survey_financial now wrap their whole body in survey_population_preparation.verification_epoch(). Every native capsule the run borrows is authenticated in full once, reused while its sources and live storage are provably unchanged, and re-authenticated in full before the run returns. The financial run's prefix opens a nested epoch; each closes with its own full re-validation. Both commits are an indentation plus one 'with' line. Proven mechanically: the statement lists of both function bodies parse to identical ASTs before and after (50 statements and 102 statements), the argument lists are unchanged, the docstrings stay outside the wrapper, and both files were ruff-format clean at the base so every hunk lies inside the wrapped function. Keeping the body in its own function matters: a build test asserts caller.f_code is runner.run_atomic_survey_population.__code__. The epoch's exit closes the nested capsule's epoch first, so this owner's final validation reaches the nested population's complete file check rather than its memo, and collects both refusals so neither can be lost to the other in a finally block. A foreign owner's error raised by a deferred borrow is translated exactly as _checked translates it, instead of escaping raw. test_us_native_verify_once_epoch: 17 tests. Validated once and reused (one _source_files pass for six borrows, five memo hits); the producer encoding never memoised; every on-disk mutation -- append, truncate, same-length rewrite with mtime restored, a file added to a source directory, a touch -- refusing at the borrow that follows it AND again when the epoch declines to close over it; a signature miss outside the roster re-running without refusing; a change no signature can see refusing at the close; a failing body finalizing nothing; and nested epochs finalizing at every level. Also adopts the change in docs/validation-cost-and-borrow-boundaries.md, which recorded this optimization as proposed rather than adopted, and corrects the codecs docstring that called the executor's check post-run. Co-Authored-By: Claude Opus <noreply@anthropic.com>
Reduce the two published CPS ASEC disability income slots to one non-workers-compensation leaf by borrowing the retirement-detail owner's already-qualified DIS literals and the retired eCPS arithmetic. No raw member is read, no source authority is issued and nothing is completed. The archived parameter dictionary and derive_us_disability_benefits_from_asec are reused rather than restated, and the workers' compensation exclusion is bound to the printed meaning of code 1 rather than to the bare integer. A slot resolves only through a closed vocabulary; NIU, missing, under-15 and contradictory slots stay unknown instead of becoming an observed zero, and a "yes" receipt with no populated source stays unknown too. The archived arithmetic's reading is recorded beside the qualified leaf and never adopted. Conditional allocation provenance is evaluated against each flag's own printed universe and never qualifies a receipt; topcode flags describe only the dollars the leaf admits. The thin attachment copies one qualified row to every clone of its original source person by identity join, takes no draw, mutates no receiving cell, and leaves rows on an arm this source never observes explicitly unknown rather than zero. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Independent review found four defects in the adapter and four claims the code did not support. A workers' compensation slot whose amount literal was missing was being admitted as a known zero, so the module contradicted its own rule that a missing cell stays unknown. The exclusion now applies only to the two statuses a coherent readable slot can carry. The implementation fence bound the detail owner's seal but not its qualifier or literal projection, so a swapped qualifier could have handed over doctored amounts, and it listed constants by hand, so REPORTING_AGE and others could be retuned after import without refusal. It now collects every module-level constant and binds the borrowed callables this module actually depends on, including the routing owner's receipt, literal and code-frame helpers. The routing owner's family allocation reading is universe-blind by its own charter, so a flag read outside its printed universe was standing as this family's allocation answer beside a per-flag label that said the opposite. The two scopes are now reported under separate names. Tests cover each fix, prove the archived function itself is called with its archived parameters rather than merely reproduced, mutate the borrowed table for real between the two seals, and check that every attached cell is the source person's own qualified cell. The module docstring and the doc drop the overbroad "reads no raw member" and "produced a zero for every row" claims. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A second independent review confirmed the first round's four fixes and found three more, all of which predate those fixes. The constant sweep stored another module's live mapping by reference, so retuning the archived parameter dictionary in place mutated the snapshot with it and passed its own equality check. Constants are now detached before they are sealed. The implementation digest was taken at call time, so a file edited after import left the loaded functions passing the fence while the recorded `implementation_sha256` named bytes that never ran. The digest is now pinned at import and compared, so such an edit refuses. The fence bound the detail owner's qualifier but not its member capture or amount comparison. The owner cross-checks the retained DIS amounts against the money owner, but `DIS_YN`, `DIS_SC1/2` and the allocation and topcode literals reach this leaf from the capture alone, so a replaced capture could have turned a workers' compensation slot into an admitted one. The capture and comparison path, the routing owner's capture and digest helpers, and the archived function's own input guard are now bound too. The doc says plainly that this list is enumerated, not a transitive closure, and names the capture boundary a consumer should know about. Topcoding reads as "some admitted slot is topcoded", so one readable flag of 1 now settles it rather than being erased by a second admitted slot whose flag cannot be read. The evidence no longer claims `raw_member_read: False` from the public qualifier, which does cause a read through its owner; it records `adds_raw_member_reader` and names the capture owner instead. The previous commit message's "the borrowed callables this module actually depends on" overstated an enumerated list, and the first commit's "No raw member is read" was true only of the pure entry points. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…laims A third independent review confirmed the second round's fixes and found one more defect: the receiver supplies its person table through its own callable, which runs after the attachment's entry check, so a receiver that retuned a module constant between the two could have had its own name used for the canonical leaf. The attachment now rechecks the implementation before returning. The implementation pin attests this file as read at the end of its own import; it cannot attest the window between the loader's read and that one, and the evidence scope and the doc now say so rather than claiming "the bytes as imported". Three descriptions were more general than the code: the owner cross-checks only the two retained DIS amounts against the money owner, not the receipt, source and flag literals; the archived arithmetic read whatever a never-asked row's literals held, usually but not always zero; and a receipt answer alone does not settle a slot, since a "no" or NIU answer beside a nonzero amount or a populated source code is a contradiction rather than a nonreceipt. New tests close the gaps the review named: the published source-code table is pinned by value instead of read back from the owner, the age-15 boundary and receipt-universe drift are exercised, the second slot excludes on the same terms as the first, a contradictory out-of-universe row is shown carrying a positive archived reading that is still not adopted, a definite allocation is shown outranking an unresolved flag where the owner's universe-blind reading does not, clone transport is shown following identity rather than row order, and a hostile receiver is shown being refused. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A fourth independent review found that binding a class object leaves its initializer swappable: the identity survives, so the implementation fence passed while the constructor this module calls after its last check could be replaced by whoever supplied the receiver, whose own table() runs inside the attachment. The fence now seals each returned type's own methods, the attachment is constructed before the terminal checks, and both results are required to have kept the payload they were handed. The archived replay runs over the four fields as the owner parsed them, so it is evaluable only where each read inside its own printed domain, while the retired function accepted any finite number. A source code outside the published table is therefore a row the old pipeline would have summed and this replay reports as unevaluable. The docstring, the doc and a new test say so; the previous "whatever it read is recorded" wording did not. Two further descriptions are narrowed to what the code establishes: the routing owner answers an unresolved flag literal before a nonzero flag outside its universe, so its universe-blind reading is not unconditionally a publisher allocation; and "usually zero" claimed a frequency no static code or invented fixture establishes. The third commit's claim that the receiver fence was closed was true only of the path its test exercised. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`native_survey_handoff.REQUIRED_RELEASE_EVIDENCE` lists `national_and_cd_target_fit`, and nothing produced it. The canonical two-artifact yardstick, `tools/score_us_release_head_to_head.py`, emitted `family`, `entity`, `value_basis` and `period` per fiscal row and no geographic identifier at all, so a national-versus-congressional-district reading of a head-to-head could not be computed from the scorecard even after a real run — while the native measurement kernel that will produce the candidate is already geography-scoped by construction (`graph_fiscal_measurement._LEVELS`, `FiscalGeographyScope`). Add `us_runtime/target_geography_view`: a pure classification of one compiled target row into the national/state/congressional_district view its declared evidence names. It prefers explicit declared level evidence, in the order the compiler produces it (the calibration hierarchy's geography tier, then `ledger_geography_level`, then `geography_scope`), renames ledger's `country` to `national` exactly as `fiscal_targets.py:2816-2826` already does, and reuses the shared `microcosm.data.us_critical_targets.is_congressional_district_target` rather than restating CD membership. A row no declared evidence places is reported `unresolved` and kept out of the three views: labelling it national would make the national view look better than the artifact is. A disagreement between the shared CD classifier and an explicit declared level is counted and reported, never silently resolved. Wire it through the scorecard: per-row `geography_level`/`geography_id` and their provenance, a `by_geography_level` rollup and a `geography_view_resolution` receipt on each artifact, a comparative `by_geography_level` block on the head-to-head, and Markdown sections for both. Per-view contributions are shares of the same `relative_error_loss` aggregate, so they sum to it exactly; the tests pin that, the counter reconciliation, the refusal routes, and equality with the frozen kernel's scope vocabulary. Schema version 3 -> 4. No gate, threshold, tolerance, band or verdict is added. The scorecard stays evidence for the owner's flip decision; this only decomposes that evidence. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`experiments/replacement_scorecard/incumbent_48b9d479.json` is an incumbent-only run with `comparison: null`, which is the shape published before a candidate exists. Pin that path: the per-artifact view rollup and its unresolved count must render there too, and the comparative section must not, or the `national_and_cd_target_fit` evidence would only appear once a candidate is available. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Root reviewed e45793e and did not adopt it: `target_geography_view.us_target_geography_view` chose a geography level from one piece of evidence and then searched an ordered metadata key list (`ledger_geography_id`, `congressional_district_geoid`, `state_fips`) that consulted neither the level nor the declaration that named it. A row whose hierarchy sat below the advertised views would have reported that sub-state id as a state's, and a district row without its own geoid would have reported the parent `state_fips` the compiler derives from that geoid (`fiscal_targets.py:2825-2830`). That search was unreachable for every spec the US compiler can currently emit: each reference carries a hierarchy seed, `_calibration_hierarchy` refuses an empty or non-single-valued fact geography (`ledger_targets.py:934-952`), and `HierarchyNode` refuses an empty id, so every compiled row resolves through `hierarchy_geography` with a bound id. No scorecard's counts change. This is a contract repair that closes the hole before a producer reaches it, and the commit claims nothing more. An identifier is now read only from a declaration whose own declared level is the row's resolved level -- hierarchy geography, then the ledger copy, then the key that level owns -- and a row with none keeps an empty identifier with `geography_id_source` `none` rather than borrowing one. Every declaration at the level stays visible in `declared_geography_ids`. Two encodings exist and are kept apart rather than reconciled. The prefixed census GEOID is one ledger field read twice (`hierarchy.geography.id` and `metadata['ledger_geography_id']`), so those two are comparable and `geography_id_declarations_conflict` asserts they agree -- a defensive invariant expected to read zero, labelled as such in the artifact note. The bare `state_fips` / `congressional_district_geoid` is a prefix-stripped restatement, never compared against a prefixed id and never a conflict; it is also not interchangeable, because both CD vintage prefixes strip to the same four digits. `"0400000US"` is an unnamed literal in four production modules with no shared constant, so no conversion is attempted here. Because of that, the rollup counts distinct *canonical* ids and reports `rows_with_bare_geography_id` and `rows_without_geography_id` beside them, with per-view `geography_id_source_counts`: mixing the two spellings in one set would count one area twice, which is the same class of wrong count the level binding exists to prevent. Markdown states the new counts in both the per-artifact and comparison sections. Schema version 4 -> 5. No gate, threshold, tolerance, band or verdict is added, and the scoring path, the loss definition and the per-target error rows are untouched. TESTS NOT EXECUTED. Free disk measured 8.54 GiB falling to 7.76 GiB during this work, below the operator's 10 GiB floor, and root notices -180/-184 hold new test runs while it is; root's continuity instruction for this job is code and static review only. Eight new tests and two fixture updates are authored and reviewed but unrun; the bounded no-engine harness is staged at 893/native-comparison-geography-id-binding-20260920/run_tests.py with the exact invocation recorded in the lane report. Static checks that were run: `ast.parse` on all three edited modules, repo-wide `ruff check` (passes), `ruff format --check` on the touched files (clean), and `tools/ci_test_groups.py --verify` (verification=ok). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The inline comment on `_CANONICAL_GEOGRAPHY_ID_SOURCES` called a difference between the two copies "a real conflict" without saying why. Name the upstream refusal (`ledger_targets.py:934-952`) that makes it an invariant violation rather than an encoding difference, matching the property docstring and the artifact note. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`_declared_geography_ids` returns `(source, identifier)` pairs, and the binding unpacked the first pair as `geography_id, geography_id_source`, so every resolved row reported its identifier as the source name and its source as the identifier: a California state row came out with `geography_id="state_fips"` and `geography_id_source="06"`. The empty branch was reversed the same way, so an unbound row reported `geography_id="none"` rather than `""` -- which would also have put the string "none" into a view's distinct-area set. Caught by the bounded no-engine suite, which had been held while free disk sat below the 10 GiB floor and ran once that cleared (21.45 GiB free): nine tests failed on the swap, including the pre-existing `test_view_resolution_routes_and_refusals`, whose geography_id assertions predate this branch's identifier work. After the fix the whole file is green under the same envelope: 54 tests, 0 failures, 0 errors, 4 pre-existing `requires_us` skips, 24.45 sampled CPU seconds, 39.71 wall seconds, 460,275,712 bytes peak RSS, and zero guard denials for country imports, network, child processes, external data and external writes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Source-qualified SS report completion for selected original ASEC and ACS people: full-original ASEC DESIGN donors, real categorical fit/probability artifacts, reason-constrained completion, and a no-fit empty-recipient path. No canonical beneficiary writes and no release qualification. The probability and normal-branch report nodes join the donor version: the graph compiler makes a FILTER depend on every member of its base version, so a fit consumer on the full-source version cycled through the donor FILTER (graph-successor3). Successor4 passed 6/6 under the admitted harness; the full graph and source files pass 32/32. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The keep-all and placement nodes named every entity ID and person membership column in their Slices. Those columns arrive in the executor's structural view and have no owner in the compiled declaration, so the first genuine Stage B run refused at compile: placement_receiving read person.person_id, which no node owns in the canonical state version. Declare only the qualified person values and candidate outputs, and omit an entity with neither. The host compile test now owns no structural column in its fixture CREATE, matching the real host, and asserts the slices name none. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The genuine Stage B successor3 run reached the enrichment compile and refused with CURRENT_SURVEY_AMOUNTS_PREFIX_DECLARATION_OR_KEY: 113 parent nodes kept their declarations and kernel hashes but changed keys, starting at survey_puf55.receiving. compile_graph makes a structural node depend on every non-structural member of its base version, and that FILTER's base is the financial version, so the arm-zero recipient projection, matrices and fixed-input nodes declared on the financial version joined the authenticated parent's predecessor closure. The original arm now opens survey_puf55.original_source_version, a keep-all FILTER off the financial version, and declares its three source nodes there. Arm-one declarations are byte-identical. The host reconstructs the version through the binding, and a new test compiles a parent FILTER with and without the original arm and asserts its predecessors do not change (and that the old binding did change them). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The genuine Stage B successor4 run passed the prefix invariant and executed the enrichment graph to the terminal observation, where the original placement cut refused PUF55_ORIGINAL_PLACEMENT_ARM_ONE_OUTPUT_OWNER. A replay diagnostic showed every arm-one output column at the terminal owned by survey_geography.canonical_state_version: population.patch gives a structural version ownership of every carried cell, so the receiving owner can never equal survey_puf55.attach once a FILTER sits between arm one and the terminal. candidate_outputs now accepts the arm-one attach node or the receiving version as the owner and, when the version re-owned the cells, requires the carried values to equal the arm-one values (ARM_ONE_OUTPUT_CARRIED). Any other owner still refuses. A new test covers the accepted re-owned shape, a changed carried value, and a later owner. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
dataclasses.replace shares the receiving frame between the reowned inputs, so the value-mutation case must run after the later-owner case. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…acement artifact The genuine Stage B successor6 run executed the original-arm enrichment cold, ran the required replay, and refused PUF55_ORIGINAL_PLACEMENT_PLACEMENT_ARTIFACTS in the host's independent reconstruction: the persisted placement document differed from the recomputed one in exactly one field, the third input_population_stamps entry (the receiving terminal). _stamp is an in-process mutation seal over replay-variant state and is already verified before and after the document is built; persisting it made the artifact non-reproducible under replay. The document now records the three input population versions instead. A new test permutes the receiving table columns (changing the seal, not the values) and asserts byte-identical artifacts. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
placement_result returns a columns/document pair, not a KernelResult. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…er lane attempt) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nalization Split clone-index-2 whole-household copies off the receiving population by clone index (never by source ID and role), run the shared arm-zero checks on the exact two-clone core, and carry every tail cell unchanged. Two-clone placement documents are byte-identical. Finalization declares its non-fixed allocation reads from the checked arm-one population. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
run_us_survey_enrichment(original_finalization=<exact policy name>) now replaces the conservative original-arm placement by the whole-arm finalization, on top of an explicit original_application_seed. The attach node keeps its ID and keep-all version, declares all 43 non-fixed outputs plus explicit age/basis reads, and has its own kernel reference and artifact type. Omitting the option keeps every existing declaration, kernel and receipt field unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Record the per-target chain-prefix rule, the maintained allocation and its three stricter arm-zero refusals, the explicit numerical policy, the host option and declarations, and the native clone-index audit against the (source ID, role) keying fixed on main by #992. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The graph declarations and the host terminal observer now check axes, ownership and candidates on the receiving population with own-tail copies split off by clone index, so a tail in the observed terminal no longer changes the declared nodes or trips the axis check before the pure result can carry it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
REVIEW-1: the with_tail fixture copied whole clone-one rows, so a tail copy's candidate cells equalled its twin's and the tail == incumbent assertions could not catch a carry keyed on source ID and role instead of the clone index (the #992 failure mode). Tail copies now carry producer-like capital-gains values, distinct floats and flipped booleans; the existing assertions stay. New tests exercise the TAIL_CARRIED refusal (a carried tail cell tampered after the carry) and the ARM_ZERO_AXIS guard. REVIEW-1's twin-fill mutant now fails 3 tail tests (it passed all 10 before); 177 finalization, placement and host tests pass. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Contributor
Author
|
Closing per Max's ruling d1161 (2026-10-15): the graph-native US line won't continue as a separate PR line. The US moves onto the graph via the shared national chain instead (the 2026-10-09 consolidation review). Nothing is lost: branch
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Draft. This is one lane of the graph-native US restart (#956). It finishes the named PUF-arm work in the 2026-09-20 publication forecast: mixed-known-member units, finalization, pruning and whole-host attachment. It also checks the native PUF arm against the capital-gains own-tail clone problem that #992 fixed on main.
Base. The branch starts from
native-release-integration-20260919at47960af43, which is not onmain. The PR againstmaintherefore shows the whole native line: 515 commits ahead and 237 behind. This lane's work is only the top five commits,47960af43..b51cc9beb: 8 source and test files, 2 docs and 2 changelog fragments. Review that range. Retarget or rebase once the native line lands.What changed
Whole-arm finalization is new and opt-in. It lives in
puf55_original_finalization.py.tax_unit_knownholds, which means every member is qualified. I read this ingraph_legacy_apply_observedandpuf55_survey_observed. So a unit gets target t only if every fixed input earlier in the chain is known for all its members. The fixed inputs sit at chain positions 0–7, 10, 11, 28 and 31, and both route profiles share one target order. Fixed values are never rewritten, and unknown members stay unresolved._write_person_tax_unit_totalsand_write_person_tax_unit_boolean_counts, using the maintained bases, pass order and age-15 earnings universe. Arm zero is stricter than arm one in three cases. These units stay unresolved:puf_maskrows, so none of its donor alignments has an arm-zero precedent. This policy applies none of them: no clipping (a negative nonnegative draw stays unresolved), no donor snapping, no tail cap, no donor positive-rate pruning and no signed-mass alignment. The policy is an exact node parameter.Host wiring. Pass
run_us_survey_enrichment(..., original_application_seed=..., original_finalization="microcosm.us.puf55-original-finalization.v1").us.survey_puf55.original_finalization_attach@1) and artifact type, declares reads ofageand the present bases taken from arm one, and consumes the same 55-step edges through the same strict merger.finalization_policyonly when the option is set.Own-tail clone. #992 fixed stages on main that keyed copies by source ID and role. On the native line:
graph_combined_clone,puf55_survey_recipients._alignedandpuf55_route_finalization._receiving_axiskey copies by source ID and clone index. They refuse clone indices outside {0, 1}.graph_*orpuf55_*module.native-agi-tail-*branches.The original-arm placement and finalization read the late terminal, so a tail could come before them. Both now split tail copies off by clone index, never by source ID and role:
Details are in
docs/us-puf55-original-finalization.md.Verification
All data is invented. The 55-target envelopes are synthetic and built from a real three-target template. Nothing here is evidence of 55 fitted models, a genuine financial owner, a whole-host run or a calibrated dataset.
test_us_puf55_original_finalization.py: 54 passed. It covers:test_us_puf55_original_placement.pyandtest_us_puf55_original_host.py(116 passed). With_application,_observed_recipientsandtest_us_graph_spm_enrichment_options.py, the PUF55 original-arm set is 254 passed. The PUF55 subset oftest_us_spine_blindness.pypassed 26.ruff checkis clean, andtools/ci_test_groups.py --verifyis ok with the new file in the US groups.test_us_graph_us_survey_enrichment.pyerrors in fixture setup insiderun_atomic_survey_financial. The error isMoneyRefusalError: STUDENT_SOURCE_SIZEfromasec_student_controls._snapshot, which surfaces asDEMOGRAPHIC_SOURCE_CONTRACT. No file on that path differs from47960af43. I did not rerun it on the base tree.Still open
🤖 Generated with Claude Code