Connect · ingestion spine
Backfill and replay
A source corrects itself, a connector has a defect, a batch arrives late. Any of those means a window has to be read again — and reading it again is the single most reliable way to create a second version of the truth. Every replay on this page names its window, carries an idempotence key, holds records the spine has already seen, and is written to the evidence ledger as a replay rather than as fresh activity.
Health
What has been replayed, and what it created
The number that matters is not how much was replayed. It is how many duplicates the replay created.
The idempotence key is the contract primary key joined to the replay window. When a replayed record resolves to a key the spine already holds, it is held rather than inserted, and the hold is counted. That is the mechanism, and across every completed replay it has not yet let a duplicate through. It has been exercised against modeled extracts only — the same mechanism against a real source with real key collisions is untested.
The replay register
Every window that has been asked for, whether or not it ran. A refused replay stays on the register with the reason it was refused.
| Ref | Feed | Window | Why | Requested | Approved by | Status | Replayed | Held | Dupes | Ran against |
|---|---|---|---|---|---|---|---|---|---|---|
| RPL-001 | Group ERP — Financials Journal lines | Jul 15, 2026 → Jul 18, 2026 | Late arriving Three days of postings from two European company codes arrived after the extract window had closed. | D. Halvorsen Jul 19, 09:00 | D. Halvorsen | Completed | 41,820 | 41,820 | 0 | modeled-extract |
| RPL-002 | Invoice Capture and AP Automation Supplier invoices | Jul 22, 2026 → Jul 23, 2026 | Connector defect The capture confidence field was truncated to two decimals for one day, pushing borderline invoices into the exception queue. | M. Okonjo Jul 24, 09:00 | M. Okonjo | Completed | 9,140 | 8,871 | 0 | modeled-extract |
| RPL-003 | Source-to-Pay Suite Purchase requisitions | Jul 27, 2026 → Aug 3, 2026 | Contract change The supplier tax identifier changed type at the source. The week was re-read against the amended mapping to see what the old mapping had dropped. | K. Boateng Aug 4, 09:00 | K. Boateng | Completed | 26,410 | 25,902 | 0 | sandbox |
| RPL-004 | Group ERP — Materials and Inventory Material movements | Jul 30, 2026 → Jul 31, 2026 | Source correction A goods receipt posting run was reversed and reposted at the source after a plant-level error. | R. Nakamura Aug 1, 09:00 | R. Nakamura | Completed | 14,205 | 13,996 | 0 | modeled-extract |
| RPL-005 | Contract Lifecycle Management Contracts | Jul 4, 2026 → Jul 18, 2026 | Connector defect Renewal notice period arrived in months on part of the estate where the contract declares days. Two weeks were re-read to size the problem. | V. Marchetti Aug 6, 09:00 | V. Marchetti | Completed | 3,180 | 3,037 | 0 | sandbox |
| RPL-006 | Core HR and Payroll Worker records | Jun 19, 2026 → Jul 3, 2026 | Scheduled backfill Two payroll periods were loaded to give the cost-per-transaction figure a denominator that spans a full quarter. | S. Bergqvist Jul 5, 09:00 | S. Bergqvist | Completed | 18,600 | 18,600 | 0 | modeled-extract |
| RPL-007 | Customer Relationship Management Opportunities | Aug 2, 2026 → Aug 3, 2026 | Late arriving One region ran an offline close and the day arrived twenty-two hours after the window. | A. Whitfield Aug 4, 09:00 | A. Whitfield | Completed | 7,310 | 7,310 | 0 | modeled-extract |
| RPL-008 | IT Service Management Service requests | Aug 7, 2026 → Aug 8, 2026 | Connector defect Resolution timestamps arrived with a local offset instead of UTC for one day, distorting the service level clock. | J. Petrakis Aug 9, 09:00 | J. Petrakis | Completed | 22,940 | 22,111 | 0 | modeled-extract |
| RPL-009 | Indirect Tax Engine Tax determinations | Jul 11, 2026 → Jul 17, 2026 | Source correction A rate table was corrected retroactively for two jurisdictions. | F. Aubert Aug 10, 09:00 | F. Aubert | Completed | 5,620 | 5,473 | 0 | modeled-extract |
| RPL-010 | Treasury and Cash Management Cash positions | Aug 11, 2026 → Aug 12, 2026 | Late arriving A value date restatement arrived from one banking partner after the position had been struck. | L. Duarte Aug 12, 09:00 | L. Duarte | Completed | 1,240 | 1,198 | 0 | sandbox |
| RPL-011 | Enterprise Data Warehouse Conformed fact rows | Aug 13, 2026 → Aug 14, 2026 | Scheduled backfill A subject area was regrained from daily to hourly and the prior day was re-read at the new grain. | N. Ferreira Aug 14, 09:00 | N. Ferreira | Completed | 88,400 | 88,400 | 0 | modeled-extract |
| RPL-012 | Supply and Demand Planning Supply plan lines | Aug 14, 2026 → Aug 15, 2026 | Source correction A published scenario was superseded within the day and the earlier version had already been read. | R. Nakamura Aug 15, 09:00 | R. Nakamura | Completed | 6,120 | 5,901 | 0 | modeled-extract |
| RPL-013 | Contact Center Platform Contact interactions | Aug 15, 2026 → Aug 16, 2026 | Connector defect Queue identifiers were sent unmapped for part of a shift and landed as other. | T. Abara Aug 16, 09:00 | T. Abara | Running | 4,870 | 4,870 | 0 | modeled-extract |
| RPL-014 | Planning and Consolidation Plan version lines | Aug 9, 2026 → Aug 16, 2026 | Scheduled backfill A full week is being re-read so the planning feed and the ledger cover the same period end to end. | D. Halvorsen Aug 16, 09:00 | D. Halvorsen | Approved, queued | 0 | 0 | 0 | modeled-extract |
| RPL-015 | Configure, Price and Quote Quote lines | Aug 12, 2026 → Aug 17, 2026 | Late arriving The quote feed has been outside its freshness commitment for six days and the missing window has to be re-read before the pricing figures can be trusted. | A. Whitfield Aug 17, 09:00 | Not approved | Awaiting approval | 0 | 0 | 0 | modeled-extract |
| RPL-016 | Transportation Management Shipments | Jul 19, 2026 → Aug 17, 2026 | Late arriving The transportation feed is more than a day beyond its commitment and a month of shipment events has not been read. | G. Mwangi Aug 17, 09:00 | Not approved | Awaiting approval | 0 | 0 | 0 | modeled-extract |
| RPL-017 | Configuration Management Database Configuration items | Jul 29, 2026 → Aug 17, 2026 | Connector defect The configuration feed has not moved in two and a half days against a one-day commitment. | J. Petrakis Aug 16, 09:00 | Not approved | Refused | 0 | 0 | 0 | modeled-extract |
| RPL-018 | Digital Analytics Session events | Aug 4, 2026 → Aug 17, 2026 | Late arriving The digital feed is more than two days behind a four-hour commitment. | C. Delgado Aug 15, 09:00 | Not approved | Refused | 0 | 0 | 0 | modeled-extract |
Why a window gets read again
An exhaustive partition of all 18 replays by the reason recorded on the request.
The idempotence key on each replay
The key is the contract primary key joined to the declared window. Two requests for the same window against the same contract resolve to the same key, so the second one holds rather than re-reads.
What a replay does to the evidence ledger
12 replays are written to the ledger with an evidence sequence. The statement on each row is what the ledger records — a replay never appears as fresh activity, because a corrected month would otherwise read as a busy month.
Actions
What is waiting on a person
A replay re-reads a window that has already been read. The only thing that makes it safe is idempotence, so every replay is measured on what it did not duplicate.
Approve or refuse a replay request
Requested and waiting. A replay is not started without a named approver.
Close out a refused replay
Refused for a reason that still stands. The underlying window is still unread.
Run an approved window
Approved and queued. Nothing has been re-read yet.
Watch a replay in flight
Running now. Nothing is committed until it completes.
Operations
What this desk is allowed to start
A surface that only reports is not operable. This is the work this page can set in motion, and the bound it runs into.
Trigger and bound
This desk can request a window, record an approval, re-read the window against a modeled extract or a sandbox, hold every record whose idempotence key has already been applied, and write the result to the evidence spine as a replay rather than as fresh activity. It cannot approve its own request, replay against a customer instance, or overwrite a record a person has since corrected.
Live observability
What the record shows right now
What each of the 249,855 re-read records did on arrival. The two states are exhaustive: a record was either recognized as already applied and held, or it carried a change and was written. 0 duplicates were created, which is the number the whole mechanism exists to keep at zero.
Current distribution
249,855 records
Is policy and strategy coming to fruition
Whether the written intent is holding here
12 replays have completed and produced 0 duplicates.
Holding on the record
The written intent is that a window can be re-read at any time without corrupting what is already there, and that the re-read is visible as a replay rather than passing as fresh activity. Measured against that, the mechanism holds: 12 of 18 replays have completed, 247,489 of the 249,855 re-read records were recognized as already applied and held, 2,366 carried a genuine correction, and 0 duplicates were created. The limit is what they ran against: 15 against a modeled extract and 3 against a sandbox, and none against a customer instance. Idempotence is proven on this estate and unproven anywhere else.
A contract, a validation rule, a freshness commitment and a replay path are real mechanisms and they are what this wave builds. What they are not is a connection. No connector on this estate is attached to a customer instance, so every batch, every replay and every write-back attempt recorded here ran against a modeled extract or a sandbox, and the registry keeps saying modeled.
Every replay recorded here ran against a modeled extract or a sandbox copy — the ran-against column says which, on every row. The idempotence mechanism has therefore never been tested against a real source producing real key collisions, late-arriving corrections and partial windows at the same time. 0 duplicates created is a true statement about what has been run; it is not yet a claim about what happens in production.
6 of the 18 windows carry no evidence sequence. Those are the ones that have not completed — requested, approved and waiting, running, or refused. A refused replay stays on the register because the fact that somebody asked for a window and was told no is itself worth keeping.