ERAIS / capability acquisition
It didn't know the function. Then it built one.
ERAIS received four ordinary examples of an unknown integer function. It inferred a program, qualified it, published a native release, answered a new input, and then restored the capability in a fresh process.
The task
f(1) = 5 f(2) = 7 f(3) = 9 f(4) = 11 f(8) = ?
This is a bounded example-defined function, not a general programming claim.
The recorded sequence
From examples to a restored capability.
Every step below comes from the receipt. The page does not invent intermediate events to make the process look smoother.
- missing
The function is unknown
The request is f(8) = ?. No qualified capability named f exists in the starting bank.
- observed
Ordinary observations arrive
f(1)=5; f(2)=7; f(3)=9; f(4)=11
- qualified
A candidate is synthesized and qualified
The acquisition path produces an executable program, checks it against the development observation, and records the native backend as native_weight_circuit.
- released
The release is committed
Capability f is published under release 7a3a2effc4fd… and generation observation:053258…
- reused
The same process reuses it
f(9) = ? returns 21 without rebuilding the engine.
- restored
A fresh process restores it
Cold restore takes 1.74 s. f(10) = ? returns 23.
Native release
| Capability | f |
|---|---|
| Identity | general-synthesis:sha256:9e4bf82a0177ac8644950e7b536fe61df2d74278ef4eacf72202ee36923d74e4 |
| Manifest | 7a3a2effc4fd704b0b7d… |
| Backend | native_weight_circuit |
| Semantic authority | qualified_native_weight_program |
| Acquire run | 54b9e6678ae905… |
| Cold run | de74cab86622fd… |
| Cold restore | yes |
| Cold arithmetic | exact |
Cost / measured phases
The answer is fast. The setup is not.
The acquisition phase is dominated by building and exhaustively qualifying the native arithmetic engine. Reporting the warm answer as the whole cost would be misleading.
| Phase | Measured |
|---|---|
| Engine construction | 0.69 s |
| Engine qualification | 140.10 s |
| Store and root setup | 16.5 ms |
| Learning execution | 1.32 s |
| Warm reuse | 74.2 ms |
| Cold store open | 2.7 ms |
| Cold restore and lookup | 1.74 s |
| Cold execution | 28.7 ms |
Why report phases?
The six-family panel's recorded total is about 221 seconds, but that is not the cost of one warm answer. It includes engine setup, synthesis, native lowering and cold restoration for six tasks.
The current acquisition phase takes about 130 seconds, almost all of it exhaustive engine qualification. Reuse is 74.6 ms and the cold answer is 26.8 ms after restore.
Engine qualification dominates the 130 s acquisition phase. Warm reuse is 74.6 ms; cold execution is 26.8 ms.
Beyond one function
Five families, six recorded tasks.
A separate pinned panel exercises arithmetic, conditional, structured, sequence and stateful synthesis. All six tasks restored from cold and passed their checks.
| Task | Family | Cold | Time |
|---|---|---|---|
| affine | arithmetic | exact | 2.40 s |
| conditional_max | conditional | exact | 1.30 s |
| structured_pair | structured | exact | 1.00 s |
| sequence_sum | sequence | exact | 31.68 s |
| sequence_max | sequence | exact | 24.89 s |
| constrained_balance | stateful | exact | 36.98 s |
Native conversion matrix
| Curated cases | 74 |
|---|---|
| Strict native | 72 |
| Explicit mixed | 2 |
| Independent reference | 73 |
The two mixed cases are classified as mixed, not silently counted as strict native.
The family panel and the acquisition receipt are separate pinned recordings. The panel was captured at sourceee6469c; the acquisition at4621b361e5af….
Composition / real dependency releases
Retained experts, then a real composite.
Two separate current runs show the pieces. A multi-turn session retains glint and spark. A dependency run builds composite from increment and square. Both restore in a fresh process.
| Capability | Backend | Native regions | Release |
|---|---|---|---|
composite | native_regions | 1 | ea1dad146b1e… |
increment | native_regions | 1 | 14d3ff14da86… |
square | native_regions | 1 | 53ff9856b761… |
Executed graph
- →increment
1 argument · returns Q · identity d6518dbf0fd0…
- →square
1 argument · returns Q · identity 1ad0641d391c…
| Composite restored | exact |
|---|---|
| Dependency releases bound | yes |
| Glint cold | exact |
| Spark cold | exact |
Program conversion / supplied source
Code in. Classified execution out.
A separate mode converts conventional C into a typed native circuit. This is source conversion, not autonomous discovery of that source.
| Case | Status | Backend | Probes | Host compute |
|---|---|---|---|---|
integer.addadd | native | native_c | 6 | 0 |
float.sin_mixeddsin | mixed | mixed donor/host lane | — | — |
memory.structural_storeset_g | mixed | c_mixed | 6 | 0 |
The two mixed cases are labelled mixed. They are not counted as strict native execution. The 724-gate integer.add case has zero host compute and an independent reference.
Fresh Startmate native run
| Case | Status | Backend | Artifact |
|---|---|---|---|
integer.addadd | native | native_c | 586a64325cd8… |
float.addfadd | native | native_c | f3e031c6e0c3… |
float.sqrtdsqrt | native | native_c | e91387ffac52… |
Evaluation / development versus final
What the wider campaigns actually show.
The learner has a strong development panel and a useful 100-task diagnostic. Neither is presented as a final independent benchmark.
20 tasks, seven families
| Tasks fully correct | 20 / 20 |
|---|---|
| Scored rows | 60 / 60 |
| Candidate evaluations | 123,247 |
| Mean task time | 19.3 s |
| Source | cd812c42a88c… |
One development run across seven families. It is not an independent final-evaluation or cumulative-learning result.
The rejected 100-task run is still informative.
| Answered | 73 |
|---|---|
| Abstained | 20 |
| Partial | 6 |
| Wrong | 1 |
| Final report | rejected: source moved |
Every task result is retained. No validated final report was written because HEAD moved while the campaign was running; the frozen campaign must be rerun at the registered revision.
Cumulative bank: current development blocker.
| Condition | Rows | Tasks | Candidate evaluations | Wall time |
|---|---|---|---|---|
| C1 fresh | 36 / 36 | 10 / 10 | 28,350 | 386.6 s |
| C2 cumulative | 24 / 36 | 6 / 10 | 16,030 | 1440.4 s |
Claims and limitations
What this establishes.
- verified
ERAIS learned f from four observations and answered f(8)=19.
EL-LEARN - verified
The learned capability answered f(9)=21 without rebuilding the engine.
EL-REUSE - verified
A fresh process restored the published release and answered f(10)=23.
EL-COLD - verified
Six recorded tasks across arithmetic, conditional, structured, sequence and stateful families executed natively and cold-restored.
EL-FAMILIES - limitation
This is a bounded example-defined integer acquisition, not unrestricted programming or general intelligence.
EL-SCOPE
Known boundaries
- The demonstration covers a bounded integer function and five recorded program families.
- Engine qualification dominates the acquisition phase; the warm answer is much faster than the cold setup.
- The synthesis-family panel is a separate pinned recording at source ee6469c, not a rerun of the current learning phase.
- The page must not present this as unrestricted programming or general intelligence.