Repository navigation
Adjudicate the UK spine grain at the ingest/calibration seam #688
Description
Joint adjudication between workstream E (#145) and workstream D (#623) of the UK-migration master epic #665. Needs resolution before the SPI-channel child of WS-E lands its strata declarations, and before the swap-acceptance child.
The mismatch
The two frozen parity instruments (uk/efrs_parity_reference.json, the 131-column input-mass register) were extracted from the incumbent's enhanced_frs_2023_24.h5 — the pre-clone eFRS: 53,508 households carrying national grossing weights. But the national staging build's --input-h5 is the certified populace_uk_2023.h5 — the post-clone artifact: 535,080 households (10 OA clones each, geography-assigned, at a tenth of the weight), and some gate parameters are tuned to that grain (the zero-weight strata caps of 100,000 = 10,000 SPI-synthetic × 10 clones).
WS-E's port naturally reproduces the pre-clone spine — that is what the incumbent's own build produces before its clone-and-assign step, and the only grain where the frozen references compare apples-to-apples. Decided on the WS-E side: the pre-clone spine is E's canonical product and parity runs there.
To adjudicate jointly with WS-D
- Staging input contract. Either (a) keep it unchanged and run clone-and-assign (today D-owned tooling) on the microcosm spine before
--input-h5— pipeline order: E spine → clone → staging gates → calibration; or (b) teach the staging build to accept a pre-clone spine and move cloning after staging — E spine → staging gates → clone → calibration. Option (b) is architecturally cleaner (a national staging artifact has no intrinsic need of clones; clones exist for geography) but reorders D's seam and forces re-deriving every gate parameter tuned at the post-clone grain. - Gate parameters. Strata caps, weight-ratio and ESS thresholds, and the input-mass reference identity are re-derived at whichever grain the staging build consumes.
- Clone count as a declaration. The 10-clone factor is currently an environment default in the incumbent and a CLI knob in microcosm; wherever cloning ends up, the factor should be a declared build parameter recorded in the release manifest.
Acceptance
A recorded decision (option a or b), the gate-parameter re-derivation plan at the chosen grain, and the swap-acceptance child unblocked.
Activity
juaristi22 commentedon Aug 22, 2026
Adjudication recorded (María, 2026-08-21) — the staging-input half, national default only: the national calibration's input contract is a pre-clone 2024-25 microcosm spine, via #743's declared non-certified staging-candidate posture. Verified basis: the incumbent's published
enhanced_frs_2024_25.h5is itself pre-clone (52,846 households = (16,288 raw + 10,000 SPI)×2 + 270 CGT donors); itsclone_and_assign×10 feeds only the local-weights product, whose constituency matrix collapses into the national weight (final_weights.sum(axis=0)). Pre-clone candidate vs pre-clone incumbent is apples-to-apples; the grain-dependentuk_zero_weight_stratacap re-derives at 10,000 as a signed measurement when the held run executes (fraction/ratio fences are grain-robust and do not move; #686 owns arming). One signed method difference carries into the score receipt: incumbent national weights = collapsed local solve, ours = direct national solve under doctrine.Still open here: the clone-count-as-declaration question and the local-lane half. Runs are held per the #623 narrowing (WS-E E8+E10, or an explicitly named base).
juaristi22 commentedon Aug 29, 2026
The pilot's measured grain (the evidence this adjudication was waiting for) — and the clone-count half answered. #761's Round A ran on spine-i (
0cce0399…) at the merged #795+#801 main; receipts inexperiments/761-uk-rowwise-pilot-receipts.mdon PR #815.Pre-clone grain, measured: 52,846 households = (16,288 raw FRS + 10,000 SPI-synthetic) × 2 + 270 CGT band donors, carrying 16,288 distinct
source_household_id(the SPI channel's 10,000 sources nest inside the FRS set). Post-clone at K=1: row counts unchanged, ids untouched, weights ÷1, exactly one appended mass record — mass conserved withabs_delta: 0.0. Thehousehold_id mod 10^8rule from the pool era does not apply to this spine and is now structurally refused on inputs carrying the explicit lineage columns; distinct-source accounting reads the columns directly.Item 3 (clone count as a declaration) is answered: K is a declared build parameter recorded in the rowwise manifest (
parameters.n_clones) and scanned across candidates in the dry-run plan (candidates.clone_counts); it is neither an environment default nor an undeclared CLI habit. Receipts: build manifest shab1a1294afeecff6e…, Logbook genesis rowa4d2c0b2b356493c…on the ratifieduk/localchain.The local-lane K itself adjudicates on #761 from the measured support (recommendation there: K=4; ruling waits for the spine-j re-run per María's 2026-08-28 ruling).