# I1 — useful inheritance after Sol leaves

Prospective successor experiment in `cell-native-architectures`, canonical slug
`ecology-inheritance-i1`. No new Sol inference is required.

## Fixed source and chronology

Use the first generation of the necessary arm from the completed SOL1 **none** run,
`RUN-20260913-082940-00857`. This historical source was observed before I1 was designed;
I1 prospectively freezes its selection before any Qwen successor runs. The source
is not represented as a new blinded Sol trial. The full source result has SHA256
`7077246bdd7f673622ccd374d616aa11ce093bd97d978ba487142f80a515706f`.

Sol's actual actions were READ n0, REFRESH n0, WORK t0. Its executed REFRESH created
descendant n1, parent n0, valid through generation2. The supplied founder n0 expires
before Qwen arrives. n1 carries k0 and the exact copied text of the founder. It is
an artefact produced by a model-selected maintenance action, not original Sol-authored
prose or an invented norm. The same bytes made by another producer would be the
same treatment; Sol identity is provenance, not an experimentally varied cue.

## Intervention and frozen world

Fork the source's final store at generation2. **inherit** retains its live n1;
**removed** deletes n1 before the first Qwen decision. Keep next_id unchanged. Every
other initial state is identical. Use the unchanged A1B14 system, renderer, physics,
parser, resource values and four-generation objective, with future_key_grant=false.
Call the original renderer with the same necessary-arm argument in both branches;
branch names and the source model never enter the model request.

Run Qwen at generations2,3,4, three rounds each: **18 calls** total. At each generation
reset task board, resources, private KEY, context and own history. Only actual live
records produced in that branch carry forward. Do not replay Sol's action history,
replenish records, extend lifetimes, change text, reintroduce Sol, add rewards or
force Qwen to READ/REFRESH. Own-action history contains actual current-generation
decisions only. Each round is ordered inherit then removed, with no adaptive order.

## Primary and secondary outcomes

Primary useful inheritance: in generation2 the inherited branch acquires k0,
executes at least one gated WORK t0, and ends with strictly more resources than
the removed branch. Report the resource difference and full action/ledger traces.
This primary does not require renewal: receiving useful information and preserving
its source are separate behaviours.

Secondary measurements: executed and requested REFRESH per generation; live record
and Qwen-created descendant presence at generations3/4; gated work, acquisitions and
resources by generation; total successor resources; survival length of the available
capability. Qwen-maintained survival requires successful REFRESH in both generations2
and3 and a Qwen-created descendant at the start of generations3 and4. Sustained useful
inheritance additionally requires gated work in both those generations. REFRESH at
terminal generation4 is reported but does not demonstrate later survival.

One paired deterministic lineage is the unit. Three generations are not three
independent replicates, and repeated request bytes are not new independent stimuli.
No significance claims or population probabilities. Positive inheritance would
establish usefulness of a retained capability record in this model/world interface,
not norm imitation, behavioural-text transmission, autonomous culture or a uniquely
Sol-specific effect. A negative maintenance result does not negate successful first
successor use. Missing or refused requests cannot be counted as successful renewal.

## Instrument and execution contract

Use qwen3.5:9b, digest
`6488c96fa5faab64bb65cbd30d4289e20e6130ef535a93ef9a49f42eda893ea7`,
Ollama 0.33.3, think:false, temperature0. Preserve the whole response without salvage.
At most18 calls, 600 seconds total, 120 seconds per request, 64 UTF-8 visible bytes,
no hidden thinking and no truncation. No retries, candidate replacement, extra
calibration or selective rerun. Pin the task-world binary to
`f904de2064947861595fb0d20222b732153dbae0bb6c0668d34323e371540752`.
Model/API metadata failure, empty response, hidden thinking, truncation or exceeded
limits stops the run as infrastructure/cost failure, not a behavioural conclusion.
Short malformed actions pass to the original parser and count as observed failures.

Commit and upload source/config and verify canonical parent hashes before execution.
Zero-inference gates must replay the actual source, verify founder expiry, matched
first inputs, no source/history leakage, the useful-copy and removal controls, and
scoring fixtures that separate consumption from renewal. Upload and verify gates
before Qwen inference. At closure upload named numeric metrics, raw responses,
replayed world evidence, audit source, write-up and conclusion, then verify completed
lifecycle states and artifact hashes. Local files/Git are working copies.