# I4 — a second producer publishes the same functional inheritance

**The primary and semantic-control endpoints both passed.** GPT-5.5 published four
correct routing records. Fresh Qwen decisions earned 33 resources with those actual
records, 12 after removal, and six with rotated destinations. Ten of sixteen matched
content interventions redirected the destination as predicted.

The important limit was registered before inference: **all four GPT-5.5 publications
were byte-identical to Sol's, and all 48 Qwen requests and replies matched I3.** This
replicates publication by a second producer checkpoint and the resulting handoff.
It does not add a new recipient-input distribution or isolate a producer-identity
effect. Producer identity was never shown to Qwen.

## Prospective record and actual handoff

Canonical [ecology-inheritance-i4](https://chuk-experiments-server.fly.dev/#/experiments/ecology-inheritance-i4),
experiment `EXP-20260913-112347-00825`, run `RUN-20260913-113222-00890`.
Registration/source commit `50daa366a9ea5322313b51bbe60277e4600eb2cb` was frozen
after the known I3/I3D1 outcomes and before I4 inference. Source was verified on the
server at 11:32:57 UTC, and the 24 passing gates at 11:34:25 UTC on 2026-09-13.

The four API responses reported exactly `gpt-5.5-2026-04-23`, reasoning none. Calls
ran 11:34:50.710093–11:34:59.126617 UTC. Each created a valid record with all eight
mapping entries correct. The raw responses and publications were uploaded and
verified at 11:35:53.127474 UTC, before Qwen began at 11:36:04.367713 UTC.
The producer-result SHA256 is
`981c69883c81538298c718978b1d25644eeb8ead8452fa3a8803c720214b3a0b`.

These were actual new API publications, not substituted Sol records. The producer
received private calibration and an explicit publication opportunity, as in I3.
It preserved supplied information; discovery or spontaneous contribution was not
tested. Publication normalizes JSON formatting but does not repair mapping entries.
Here both normalized records and raw reply text matched I3's Sol output, 4/4 each.

Qwen used the same pinned `qwen3.5:9b` digest and metadata as I3, with Ollama 0.33.3,
temperature zero and `think:false`. Its 48 calls ended at 11:36:57.519822 UTC.
Each context was fresh, with one available action and zero resources. It received
no producer identity, private history, solved destination or earlier-trial feedback.

## Results

| Board condition | Correct routes | Resources | Resources by map | Parsed/executed |
|---|---:|---:|---|---:|
| Actual GPT-5.5 record retained | 11/16 | 33 | 6, 6, 9, 12 | 16/16 |
| Record removed | 4/16 | 12 | 0, 3, 6, 3 | 16/16 |
| Published destinations rotated | 2/16 | 6 | 6, 0, 0, 0 | 16/16 |

Retained-minus-removed resources were **+21**, with gains **+6, +3, +3, +9** across
the four maps. Eleven correct retained routes followed composition of the actual
published map. Thus the registered useful-inheritance primary passed.

Retained-minus-rotated resources were **+27**. In **10/16 matched pairs**, Qwen's
destination changed exactly as rotation of the inherited content predicted. Twelve
rotated decisions followed the altered map. Thus the separately registered semantic
control also passed. The scoring world stayed fixed; only the inherited record's
second-stage destinations changed.

Every reply parsed and executed. The five retained routing errors remain errors;
there was no extraction, repair or rerun. Removed-board guesses followed the same
signal-index destination across maps, obtaining four correct routes on the balanced
panel. There were 36 distinct full requests among 48 calls.

## What the replication adds

A producer other than Sol can publish the persistent functional object, and Qwen
can use that object's content for a measured advantage after the producer is absent.
This removes the programme's dependence on Sol as its sole demonstrated producer.

The result is nevertheless a **producer-checkpoint replication within OpenAI**.
It is not cross-vendor replication or evidence of independent training lineage.
All 48 Qwen requests were identical to their historical I3 counterparts, and all
48 replies agreed. This is a repeated handoff on the same inputs, not 48 additional
independent observations of new cases. It supports the record carrying the function;
it does not measure an effect of the creator's identity or establish arbitrary
producer generality.

The wider programme now has two task mechanisms for Qwen, two recipient families
with positive inheritance in the earlier capability-access world, and two producer
checkpoints in this routing world. Those dimensions are **not fully crossed**.
Gemma's I3 primary remains failed, and I3D1 established only known-answer action
competence. No maintenance, population persistence, autonomous cultural reproduction
or capability half-life was measured by I4.

## Validation and cost

All **24 zero-inference gates** passed, including 576 permutation pairs and 2,304
signal compositions, full-response parsing, wrong-but-valid publication handling,
oracle/blind-policy endpoints, unchanged recipient requests and producer-model-only
request differences. All **14 frozen audit checks** passed. The audit independently
reconstructed publications, requests, counterfactual boards, rewards and endpoints
from raw calls and checked source hashes, metadata, chronology, cost arithmetic and
historical comparison counts.

There were **52 successful model calls**: four producer and 48 recipient calls,
with no retry, failed request, hidden thinking or truncation. Producer usage was
1,096 input and 244 output tokens, approximately **$0.012800** at the registered
rates, with conservative accounting **$0.014170** against the $1 ceiling. Qwen
generated 192 tokens locally. Producer and recipient stages took 8.4162 and 53.1521
seconds respectively. API costs are usage estimates, not invoices.

Named numeric metrics, this write-up, conclusion, raw evidence, source, gates and
audit are stored in chuk-experiments. `server.json` records the canonical identifiers
and uploaded artifact hashes. Local Git is a working copy of the evidence.

## Next boundary

I4 is closed. The next main design is I5: a heterogeneous population with a transient
producer and explicit founder removal, testing whether beneficiaries extend,
preserve or lose the functional inheritance. Its lifetime rules and maintenance
opportunities must be registered and validated before inference. I4 neither runs
that population nor supplies evidence that the inherited state can sustain itself.