# Package caller construction observations

2026-10-09 KST. Compiler candidate `e6a22ba3329c4511ce53296b65b42dc79dbad263`,
Go 1.27.1, clean source build; PR 1402. Canonical CI and merge are still pending.
The installed baseline remains public 0.6.16 at `146a5085`.

The baseline uses the same package source and caller cases, chooses the locally
passing `late_unbounded` body and matches 0/1 caller expectations. Its original
execution is retained separately from five fresh candidate observations.

| Observation | Attempts | Local rejections | Native combinations | Native fault combinations | Evaluation |
| --- | ---: | ---: | ---: | ---: | ---: |
| budget 5 | 5 | 2 | 3 | 1 | 1/4 |
| budget 6 | 6 | 2 | 4 | 1 | 4/4 |
| three packages, source order | 48 | 16 | 32 | 8 | 4/4 |
| same packages, own model | 41 | 16 | 25 | 8 | 4/4 |
| saved model construction | all 41 rechecked | retained | fresh executions | retained | 4/4 |

`compare.go` decodes JSON with exact numbers, compares original package-keyed
caller cases, identical selected sources and every evaluation trace, including
input `9007199254740993` and output `-45035996273704965`. It also compares the
entire saved construction history, binds the supplied receipt digest and checks
zero new replay inference. `comparison.log` is its actual output.

Each `.json.gz` is the unchanged native JSON output, compressed. `source/`
contains original Gooo, manifests and cases. Decompress the three mixed JSON
files into one directory and run `go run compare.go <directory> source` to
repeat the comparison. Native replay uses the supplied source tree and the
candidate's `package construct --receipt` command.

The unchanged all-data demonstration model has metadata SHA
`3c68205a660695103712e2115ce90accd298b406dced8ad5ad5c9b84bab0a202` and weights SHA
`76f68845a03ed8c8bc261a57c919e96dcdd352d3c35bcdaeb3d9b51a88975c6f`.
One prediction took 29,416 ns; tensor memory was 2,096 bytes. The model command
recorded 8.42 seconds wall time, 4.09 user and 2.75 system CPU seconds, and
87,949,312 bytes maximum RSS. The command includes native compilation and 25
executed combinations plus later evaluation. Host-wide CPU utilization was not
measured. These related finite examples do not establish general speed or model
transfer; package enumeration differs from the older single-source example.

The local test records cover race checks, imported calls, explicit binds, integer
search, holdout/evaluation separation, source identities, exact integers and
replay. The retained 17 MiB red test exposed a reader using the earlier 16 MiB
file limit; the candidate reads the declared 32 MiB receipt boundary and passes
that regression. CI status is tracked separately from these local results.