Qualify recurring work
Retain the procedure; measure failures, intervention, latency and total cost per accepted outcome while its preconditions hold.
AI workflow software for continuing work across engineers, models and infrastructure.
Cantos connects a job's inputs, decisions and results to reusable procedures across engineers, models and infrastructure. Explore the historical corpus, inspect exact conditions and producing sources, then use Second Run, its execution-and-reuse capability, to run supported procedures, reuse verified results and identify requalification when declared dependencies change.
Start with public history. Export records, compare observations and continue supported work from your own collection.
START WITH THE HISTORICAL CORPUS
Inspect observations, sources and conditions.
The corpus brings together InferenceX history, MLPerf, captured issues and PRs, prices, workload traces, model and demand history, and research. Use our runs and worked scenarios as labeled examples within that record.
01 / INSPECT SOURCE HISTORYWhat was observed, when and by whom?Browse historical records, inspect exact conditions and producing sources, and compare observations across the retained corpus.Open the historical corpus ↗
02 / TRACE AND NORMALIZEFollow a record into the local interface.Use the local source-to-history-to-normalization interface to inspect retained collection structure and source boundaries.Open the local corpus interface ↗
03 / RUN SUPPORTED PROCEDUREContinue work with Second Run.Run supported procedures, reuse verified results and identify requalification when declared dependencies change.Read the integration work ↗
HOW WORK CONTINUES
Second Run carries resolved judgment into execution and reuse
Restore and reconstruct
Carry the source and processing record through restoration, rendering and reconstruction. Those production examples are held in estate records, not published here.
Test consequential claims
Reconstruct a claim, test its decisive assumptions and identify the evidence that would change the conclusion.
Genesis binds claims to source identity; journals retain how knowledge changed; Arc separates artifacts, execution and consequences.
Start with the ecosystem guideIN THIS REPOSITORY
Implemented work and entry points
CAMPAIGN / MEASURED WORKHot Aisle · Workload reportConnected runner, retained Run 1–3 records, price history and rented-seat qualification kit.Open report ↗
EXPERIMENTAL DESK / PUBLIC EVIDENCEDistance to the floorExperimental desk rating from dated public reviews and provider surfaces; shortfalls and unknowns remain explicit.Open floor ↗
TECHNICAL INVESTIGATIONClusterMAX challengeFrozen retrospectives, the 3.0 binding and a standing incident collection.Open investigation ↗
MANUAL DATA CYCLECirculation receiptsManually invoked public-data cycles, receipts, retained-study checks and probes.Open receipts ↗
BOUNDED CLAIMSPublic shelfA public shelf of bounded claims with their evidence limits.Open shelf ↗
ECONOMICS / INSTRUMENTCompute deskProvider-neutral price and material inputs for bounded decisions.Open Compute desk
RESULTS / INDEPENDENT COLLECTIONSCompare, challenge and continueRetained results, claim checks, portable decisions and a downloadable kit for running your own collection.Open the community board ↗
HISTORICAL CORPUS / PUBLIC ENTRYSource historyExplore InferenceX history, MLPerf, captured issues and PRs, prices, workload traces, model and demand history, and research. Records retain their producing source and conditions.Inspect source history ↗
LOCAL INTERFACE / CORPUSSource to history to normalizationInspect and normalize the retained collection locally. Imported rows preserve their histories and may require a supported benchmark binding before execution.Open local interface ↗
SECOND RUN / SUPPORTED PROCEDURERun and reuse supported workDocumentation for running supported procedures, reusing verified results and identifying requalification when declared dependencies change.Read integration work
WORKSPACES / EXAMPLES
Open a concrete decision
These are two examples inside the program. Run 3 is a retained measurement; the invoice scenario is synthetic.
This repository includes the software campaign and technical investigations shown above. Media restoration and reconstruction examples remain in estate records. Published examples do not establish production qualification.
DECISION / 01
Synthetic worked example
Invoice extraction · 24 invoices. Two implementations. The same required fields.
Synthetic24 authored invoicesEngineResearchCore 1.0.2 candidatelocal, verified
LOWEST COST / ACCEPTED INVOICE
Opening the decision…
AcceptedFormatIncorrectLate / incompleteSelect any invoice to inspect
Decision history 01
Earlier deliveries stay intactOPERATING EVIDENCE / 24 SEP 2026
Same work. Different bill.
Run 3 · 8,622 coding requests · MI300X versus H100
38%
Lower cost per accepted request
$0.74 against $1.20 per 1,000 accepted, whole run, at list prices. Accepted counts sit within 1.3%, so the gap is price.
Run 3list pricesone run per arm$ per 1,000 accepted
Run 3list pricesA/T0 own-seat equivalent, 64.7 min; the whole shared seat was $0.90. N/T0 fully closed ledger, 69.8 min. Invoices unreconciled.
Accepted, of 8,622
Run 3Baseline 1Correct under base and plus tests, first token ≤1 s, done ≤60 s. The 26 Sep repeat is one fresh allocation; the comparison is exploratory.
First token p50 · p95 · p99
Run 3vLLM 0.30.0H100 is quicker in the typical case; MI300X holds the tail through the bursts. Log scale, milliseconds.
Where the ranking flips
Run 3rescaled list priceArithmetic on the measured H100 run (4.41 × 0.74 / 1.20): below ~$2.72/h the H100 is cheaper. Not separate runs.
THE CONTRACT
Correct + first token ≤1s + completion ≤60s
Qwen3-Coder-30B-A3B-Instruct-FP8 on vLLM 0.30.0; the Azure code trace replayed for one hour at factor 0.95. One run per arm.
| Measure | A/T0 | A/T1 | N/T0 |
|---|---|---|---|
| List rate | $2.99/h | $2.99/h | $4.41/h |
| Completed | 8,622 | 8,615 | 8,622 |
| Correct | 4,371 | 4,334 | 4,329 |
| Accepted | 4,336 | 4,292 | 4,280 |
Sources & accounting
RUN3-RESULTS.md · 24 September 2026
MI300X own-seat equivalent: 64.7 minutes; H100 closed interval: 69.8 minutes. The whole AMD seat also contained the smoke and second AMD arm ($0.90/1k across 9,086 accepted). The comparison preserves those scopes. AMD credit-funded; H100 operator-funded.
FIRST-BASELINE-RESULT.md · 26 September 2026
The later Hot Aisle baseline recorded 4,274 accepted requests, 62 fewer than the single Run 3 observation. Observed credit change $3.79, not an invoice.
CORRECTION RETAINEDCorrectionpost-hoc
4,371 → 6,192 correct
The historical sanitizer regrade improved correctness. Its deadline intersection was not recomputed, so the change in accepted requests is unmeasured. The invoice laboratory evaluates both fields and deadlines on its separate, fictional records.
SYSTEM MAP / 01
A decision has a route.
Three owners. One packet carried forward.
01PAGE / LOCAL
ResearchCore
IN THIS PAGEEvaluation → dependencies → review → edition
Owner & version
The local, tested ResearchCore 1.0.2 owner candidate. The published 1.0.0 application remains unchanged.
02PROCEDURE / ESTATE
Second Run
ESTATERetained judgment → checked procedure → reuse
Execution boundary
integration/work.py binds exact input, code and runtime identities. The browser does not execute estate jobs.
03CUSTODY / SEPARATE
AXM foundations
SEPARATE OWNERIdentity → knowledge custody → continuation
Integration boundary
Genesis and Core retain their own identity and knowledge responsibilities. This unsigned page does not perform Genesis signing or activate CAIRN.
PACKET / DECISION PATH06 STEPS
- 01Evidence
- 02Assumption
- 03Calculation
- 04Review
- 05Edition
- 06Successor
The page shows the decision. Each owner keeps its own execution and custody boundary.
THE WORKING OBJECT
A decision you can continue.
Evidence, assumptions, calculation, review and editions travel in the packet. Cantos supplies this interface; the native owner supplies the calculation and journal.
Independent demonstration. Synthetic invoice data. No BEP commission or endorsement.