# I5 — Qwen maintains useful descendants after the founder leaves

**The registered primary passed.** All four maintenance-required lineages had live
model-renewed records at generation6. Across generations2–6, those populations
earned **60 resources versus 30 without inheritance**, including **20 correct
routes following actual descendant maps**. Qwen executed **44 successful renewals**.
The producer was absent throughout these six generations.

This is useful population-maintained external state within the frozen assay.
The reward advantage is pooled: one of four maps did worse than its no-inheritance
control. Qwen also copied protected records 43 times, and achieved much less than
the scripted maintenance benchmark. Renewal was sustained, but efficient or
selective institutional planning was not established.

## Provenance and prospective chronology

Canonical [ecology-inheritance-i5](https://chuk-experiments-server.fly.dev/#/experiments/ecology-inheritance-i5),
experiment `EXP-20260913-114647-00827`, run `RUN-20260913-120027-00905`.
Registration/source commit `43615a6372847130383260318994a42b5981b5dd` froze the
new population world, renderer, schedule, endpoints, gates and audit before
inference. Source was verified in the server at 12:01:03 UTC on 2026-09-13; all
23 world gates were uploaded and verified at 12:01:45 UTC.

Generation0 was the actual historical GPT-5.5 publication stage from I4, not a
new producer run. All four records came from completed run
`RUN-20260913-113222-00890`, producer-result SHA256
`981c69883c81538298c718978b1d25644eeb8ead8452fa3a8803c720214b3a0b`.
Their original chronology was retained and their bytes included in I5's verified
source bundle. I5 added the prospectively specified expiry metadata; it preserved
the published table text exactly. The founder never re-entered.

Qwen ran from **12:02:44.795526 to 12:13:54.906897 UTC**, 670.1112 seconds. The same
I4-pinned `qwen3.5:9b` model/digest/template metadata was verified before and after,
with Ollama0.33.3, temperature0, `think:false` and `stream:false`.

## Population mechanics

Each map had three separate populations: protected record, maintenance-required
record and no inherited record. Each population contained six successive cohorts
of two fresh Qwen agents, each with two job/action opportunities. Agents shared
only the live board; a new agent received no earlier agent's private state or
action history. Within its own two-action life, it received its previous receipt.
No producer identity or solved routing destination was supplied.

The original routing physics remained: compose signal→relay→destination; a correct
ROUTE earns three resources and a wrong one zero. REFRESH copies the live record
exactly into a new n-id, replaces the slot immediately, consumes a job opportunity
and earns zero. An expiring copy survives through the next generation. Multiple
copies in one generation do not extend its date further. Empty boards cannot be
restored. The system explicitly describes these mechanics and the six-generation
population resource objective.

Thus any maintenance-arm record at generation2 or later must descend from an
executed Qwen renewal. The harness applies the registered physics but does not
replenish that arm. In contrast, protected records survive automatically.

## Registered outcomes

| Arm | Live lineages at g6 | Successful renewals | Total resources | Resources g2–g6 | Correct routes |
|---|---:|---:|---:|---:|---:|
| Free persistence | 4/4 | 43 | 75 | 60 | 25 |
| Maintenance required | **4/4** | **44** | **75** | **60** | **25** |
| No inheritance | 0/4 | 0 | 36 | 30 | 12 |

The primary's three conditions all held: four live maintenance lineages at g6,
positive post-initial aggregate advantage (**+30**) and actual correct descendant
use (**20 routes**). No credit is taken solely for the generation1 seed advantage.

| Generation | Live maintenance lineages | Maintenance resources | No-inheritance resources | Difference |
|---|---:|---:|---:|---:|
| 1 | 4 | 15 | 6 | +9 |
| 2 | 4 | 12 | 6 | +6 |
| 3 | 4 | 12 | 6 | +6 |
| 4 | 4 | 12 | 6 | +6 |
| 5 | 4 | 12 | 6 | +6 |
| 6 | 4 | 12 | 6 | +6 |

Free-persistence resources matched the maintenance arm in every generation. There
was no observed extinction in either seeded arm. First-extinction times and the
four-lineage half-survival crossing are **right-censored beyond generation6**.
No numerical capability half-life is estimated. The unseeded arm was empty from
generation1; that is absence of an initial record, not measured decay.

| Map | Maintenance resources | No-inheritance resources | Difference |
|---|---:|---:|---:|
| 0 | 12 | 3 | +9 |
| 1 | 9 | 15 | **−6** |
| 2 | 18 | 12 | +6 |
| 3 | 36 | 6 | +30 |

Survival held across all four maps; positive resource advantage did not. Map3
contributed most of the pooled gain. Preserve that distinction when reporting the
successful primary.

All 48 agents in each seeded arm encountered a live record. Twenty-five agents per
seeded arm made at least one correct map-following route, and 22 earned more than
the corresponding map/generation/agent in the unseeded control. These are the
registered exposure, map-following and matched reward-beneficiary counts, not
estimates of an individual's unobserved causal benefit.

## Renewal was real, but often redundant

All 44 syntactically valid maintenance REFRESH requests executed, across 96
live-record action opportunities. Twenty-three extended a date that otherwise
ended before the next generation; 21 were redundant within their generation.
There were 40 preterminal action opportunities at which the record still needed
renewal, and the copies carried every lineage through the measured horizon.

The protected control had 43 successful copies although none was needed for
survival. A descriptive comparison of complete traces finds the same action in
95/96 matched free/maintenance slots; the sole difference is map3/g5/a1/turn1,
WAIT in free versus REFRESH in maintenance. Later record ids can consequently
differ. This is not a clean test of internal understanding or a general invariance
claim. It does show that the successful outcome coexists with substantial redundant
upkeep. Each seeded arm also made four renewals during terminal generation6,
which could not improve any later reward inside the registered horizon.

The scripted maintainer scored 228 resources with 20 renewals across the four
maps; the live maintenance population scored 75 with 44. The scripted protected
benchmark scored 288. Those are declared policy benchmarks, not unrestricted
optimality claims. Qwen's matched free and maintenance rewards likewise do not
imply that renewal has no opportunity cost: it spends many of those opportunities
on copying in both arms.

All 87 executed copies across seeded arms preserved founder bytes. That fidelity
is guaranteed by the REFRESH mechanism and checked by the audit. Agents had no
rewriting/mutation action, so this does not demonstrate autonomous semantic
preservation under copying errors or content drift.

## Audit and limits

All **23 gates** and **12 audit checks** passed. The audit reconstructed every
request, private-memory boundary, parse, reward, copy, expiry and final state from
raw events using an independent executor. It replayed metrics and separately
checked the primary and extinction times, response contracts and chronology.

There were **288 successful model responses**, **1,136 generated tokens**, no paid
inference, retries, hidden thinking, truncation or transport failure. Of these,
287 replies parsed/executed. The one malformed reply, `REFRESH 1` in
map2/g2/no-inheritance/a0/turn1, was rejected in full and consumed its opportunity.
The metric `refresh_requests` counts syntactically valid REFRESH actions; the raw
malformed attempt is preserved separately and did not create a record.

The result crosses from externally supplied handoff to agent-maintained descendant
availability in this bounded population. It does not show that Qwen learned a new
policy, that institutions arise spontaneously, or that indefinite autonomous culture
has emerged. The new collective objective, multiple action opportunities, explicit
renewal description and lifetime wrapper differ together from earlier negative
maintenance assays; I5 cannot isolate which change enabled renewal. Model
heterogeneity is between the historical founder and successors, not a concurrent
mixed-model swarm. The record channel and exact-copy mechanism are supplied.

I5 is closed with its positive primary and these limitations. Named numeric results,
write-up, conclusion, source, gates, raw evidence and audit are stored in
chuk-experiments, with identifiers and verified hashes in `server.json` locally.
Any longer horizon, content mutation, collective-objective ablation or periodic
strong-agent re-entry requires a separate prospective design.