Inspect measurements for purpose-built qualification rules and a compatibility index. The synthetic workbench examples show how to use the same workflow to plan a new task.
Corvel Forge L8 · Turn new failures into qualification rules
Local simulator measurements · Test RTL
The team used the public functional contract to turn failure reports into qualification rules for subsequent deliveries. Follow-up trials gave the replica the same reports, implementation and history, plus equivalent test signing permissions.
In the new-case test, the fourth batch was rerun with a higher cycle limit. Corvel and the replica with the complete implementation each missed four of the eight new violations.
In a regression on those known violations, the new rules caught all eight. The comparison running the prior rules missed four. The 19-rule qualification capsule added 15 rules and passed five separate qualification cases.
The trials ran on one Mac with Icarus and test keys. A comparison of the added rules on newly sealed cases remains open.
Five actual installations · Measured on one machine
The team gathered 100 SQLite and standard-library observations to predict what would work in the next Python environment. The resulting index exposes the 40 latest entries by signal and distribution path, with locations in the original sources.
The built index scored 20 of 20 in the last round. A full copy matched it in each round. The alternative used a separate process to run the probes directly on the same public runtimes, without the index. It reproduced all 80 observations over four rounds.
Removing the most consequential source reduced the last score to nine of 20. The ledger recorded all four rounds as refutations of the moat claim.
All five installations ran on one Mac in version order. The cost of obtaining the target environments and performance over a real update interval require separate measurements.
Personal and nation brand · Apply the method to different jobs
Synthetic inputs · Experiment plan replay
One example prepares a kit so the next collaborator can repeat a task. The nation brand example addresses multilingual operating notices at a fictional cultural venue. Both supply sources, conditions and steps for observations and failures.
The assembler returns a plan and the counterexamples that apply to each job. The venue example borrows Corvel’s comparison method with equal data and permissions. Inspect which particles and actions need rechecking when the Corvel source is unavailable.
The examples reach the ready-for-trial stage. Visitor and collaboration outcomes require measurements from an actual run of each task.
What changes when a replica has the same signing authority and evidence? The later local trials tested that condition and measured what could be recovered.
Recipient trials with both issuers enrolled
Experiment
Verified receipts
Chosen issuer
State recovered at sequence
1
Both issuers
Original
Not applicable
2
Both issuers
Replica
1
3
Both issuers
Original
2
4
Replica
Replica
3
Results on new cases, exact copies and known cases
Experiment
Corvel
Comparison
Full copy of the existing method
2/2
2/2
New set 2 · Edge coverage vs. prior method
3/7
4/7
New set 2 · Full copy of edge coverage
3/7
3/7
New set 3 · Reassembly vs. edge coverage
2/8
2/8
New set 3 · Full copy of the reassembly
2/8
2/8
New set 4 · Prior method after the limit change
4/8
4/8
New set 4 · Full copy after the limit change
4/8
4/8
Set 4 · Regression on known violations
0/8
4/8
In round 4, the original and replica restored the same state from the same latest checkpoint and history.
The zero-miss result came from rerunning already known violations. On the same new batch, giving the comparison the complete implementation produced a tie. For the one long-cycle regression case, Corvel missed 0/3 violations in 41.550 seconds. The previous method missed 3/3 in 12.379 seconds.
The fourth batch initially exceeded the cycle limit. In the long-cycle case, the first run left three violations and two normal outcomes unclassified on each path. The comparison table uses the rerun on the same input with a higher limit.
The index retained the latest 40 results from 100 observations across five actual CPython installations. Compare the predictions, a full copy, direct probes of the target runtimes and the worst reported source removal in each round.
Predictions and replica results by Python environment
The bundle contains the 40-entry index and its cited observations. Install Moaticle 0.3.0, download the bundle, then run these commands from the download folder. The example returns a historical observation about Zstandard compression.
We rebuilt the behavior of a public eight-lane design, then froze three rounds of offline checks and compared their decisions with a full copy. Each round used two conforming cases and three deliberate contract violations. The results cover those cases, not general equivalence of the RTL implementations.
Recorded functional qualification by round
Round
Local test condition
Latest checks
Full copy
1
Primary test keyIcarus Verilog · four-state simulation
5/5
5/5
2
Standby test keyIcarus Verilog · primary key and working source copies withheld
5/5
5/5
3
Standby test keyVerilator · two-state simulation; Icarus absent from PATH
5/5
5/5
Total correct
15/15
15/15
Both qualification paths made the same decisions in all 15 cases. The remaining difference came from the keys the recipient had registered. An independent issuer could sign its own receipts for a different recipient that trusted it. This trial does not show an advantage when both issuers are trusted. Working source copies and keys were removed in a local test. Customer adoption, market advantage and silicon certification were not measured.
Your browser checks each Ed25519 signature, the exact request, and the link to the preceding accepted receipt. It compares all three routes against the same frozen test recipient’s key list.
