MEMORY KEPT
Inventory investigations with memory
Find the next
bin to count.
Countback stores earlier stock counts in Sibyl Memory. It uses those counts to select the next bin.
Real Sibyl app. No wallet needed.
Free hosting may take a minute to wake.
01 SKU03–06 bins08 moves max.
THE INTERACTIVE PROOFSYNTHETIC REPLAY
Same stock.
Different next count.
Loading the planner…
MEMORY REMOVED
Remaining states
Challenge it: choose another hidden world
Evaluator quantities are revealed one requested bin at a time. They are never planner inputs.
01 / THE MEMORY
Use earlier counts.
Select the next bin.
Countback can use a count after stock moves. The count record includes its position between movement windows.
Record the evidence
Save an earlier count and its verified position between movement windows. Sibyl retains the observation and its history.
Close the session
After the process stops, a new controller reads the stored evidence. It calculates which stock quantities remain possible.
Ask for the next count
The earlier observation changes the selected bin. New counts reduce the possible results. Conflicting evidence stops the proposal.
THE DELETION TEST
Keep the solver. Keep the stock. Remove the earlier observation.
B → Athe next selected bin changes in the disclosed example
Read the fresh-process proof ↗The example takes one additional count with history and two without. Equivalent JSON or a complete external archive restores the advantage. The dependence is on retained evidence, not exclusive Sibyl mathematics.
02 / THE VALUE
Compare count time
with the added work.
One fewer count helps only if it saves more work than it adds. Include preparation, history collection, freeze control, and review.
FOR INVENTORY-CONTROL LEADS
A contained discrepancy.
An investigation to resume.
Operators already report costly recount and reconciliation work. Countback uses existing evidence to select the next check.
Read the customer evidence & validation plan ↗Public operator pain is documented. Countback has no completed customer pilot or measured warehouse savings.
Your break-even check
MINUTESFinish both replay lanes above. Then replace these illustrative assumptions with your own.
What this comparison includes
Include travel and handling in each count. Use zero history cost only for an existing observation. Shared baseline costs cancel. These inputs calculate time for the selected path. They do not change the count policy. They do not predict incident frequency.
03 / THE BENCHMARK
Compare all
test results.
1,944 generated cases. The same selected truth and production solver in each pair. These results do not establish customer savings.
662 cases improved with retained history.
1,280 cases tied.
2 cases got worse.
04 / THE APPROVAL
Review the result.
Check the approval.
A demo wallet recorded an approval digest on Base Sepolia. Read the live contract. Then change one quantity and repeat the check.
Verify the live Base proof ↗The approval is an existing testnet transaction. It identifies the approved artifact. It does not prove physical stock accuracy.
RECONCILIATION / APPROVAL RECORD↗
Ready for review.
A11Inferred
B9Direct count*
C10Inferred
*Synthetic measurement. Conditional on the declared scope and freeze.
BASE SEPOLIA · EAS904404f6…6628d
05 / OPEN THE EVIDENCE
Read the evidence
for each result.
Start with the product. Go as deep as you need into the process, controls and source.
01Watch the demonstrationContinuous capture and execution ledger↗02Restart, recover, removeFresh processes · recovered UUIDs · deletion and substitution↗03Agent handoffMCP contract · executed Hermes calls · precise verification boundaries↗04The customer-value hypothesisPrimary sources · existing alternatives · falsifiable trial procedure↗05Public sourceMIT license · real development history · reproducible checks↗06Run it yourselfLocal application and agent-host setup↗
THE OPERATING BOUNDARY
A limited case.
A result for review.
One SKU, 3–6 bins and up to eight ordered transfers. Trusted opening counts. Full-once or skipped movements. Complete records and a continuing final freeze.
Missing premises and contradictions stop selective inference. A consistent model cannot identify every omitted movement. Final quantities remain conditional; Countback never adjusts stock automatically.
Work in inventory? Challenge our assumptions ↗