# I6 — useful descendants survive when Qwen writes their contents

**The registered writing primary passed.** All four seeded writing lineages had
fully functional model-written descendants at generation6. Qwen executed 43
renewals supplying complete tables; every accepted table retained all eight founder
entries and all four composed routes. Generations2–6 earned **84 resources versus
27** under the matched writing interface with no inheritance, including **28
correct routes following written descendants**. No semantic mutation or functional
loss was observed during any accepted renewal.

This removes I5's mechanically exact copy operation as a necessary condition for
useful persistence in this bounded panel. It does not establish fidelity under
corruption, partial visibility or an indefinitely long succession.

## Registration and chronology

Canonical experiment [I6](https://chuk-experiments-server.fly.dev/#/experiments/ecology-inheritance-i6),
`EXP-20260913-122258-00830`, run `RUN-20260913-132239-00914`, programme
`cell-native-architectures`. Registration commit:
`43d559ca73dc5645b3cf903af75308380b40969c`.

The source and preregistration were uploaded and verified at
2026-09-13T13:23:14.837414Z, before world gates completed at 13:23:42.548860Z.
All 22 gates passed and their upload was verified at 13:26:38.554173Z, before
inference began at 13:33:33.294507Z. Inference ended at 13:56:08.304740Z;
the frozen independent audit passed all 12 checks at 13:56:34.004266Z.

Actual founder records are the four historical GPT-5.5 publications from I4,
`RUN-20260913-113222-00890`, producer-result SHA256
`981c69883c81538298c718978b1d25644eeb8ead8452fa3a8803c720214b3a0b`.
Their original provenance is preserved; no new producer was invoked or retrospectively
represented as prospectively sampled. I5 is the completed antecedent
`RUN-20260913-120027-00905`.

## Frozen comparison

Four maps × four arms × six generations × two fresh Qwen agents × two actions:
**384 local calls**, no retries, repairs or omitted cells. The producer was absent.
Only the live shared record and each agent's own first-action receipt were supplied;
no producer history, private calibration or previous agents' action logs were passed.

`copy_seeded` retains I5's `REFRESH n0` exact-copy operation. `write_seeded` instead
requires `REFRESH n0 <full-JSON>`; the supplied complete two-stage table becomes the
descendant. Each has its own empty-board control under the same system instruction.
The engine validates schema, not truth: a valid inaccurate table would replace the
record without repair. Gates demonstrated both accepted corruption and a live but
functionally failed lineage. Empty boards cannot be restored by this renewal action.

Renewal consumes a routing opportunity, pays zero immediately, and sets expiry to
the end of the next generation. Repeated renewal in one generation does not extend
that horizon further. All agents receive the same collective resource objective.

## Results

| Measure | Copy + founder | Copy, no founder | Write + founder | Write, no founder |
|---|---:|---:|---:|---:|
| Calls | 96 | 96 | 96 | 96 |
| Parsed / executed | 96 / 96 | 95 / 95 | 96 / 96 | 77 / 77 |
| Total resources | 75 | 36 | **105** | 33 |
| Resources, generations2–6 | 60 | 30 | **84** | **27** |
| Correct routes | 25 | 12 | 35 | 11 |
| Successful renewals | 44 | 0 | **43** | 0 |
| Needed renewals at action time | 23 | 0 | 23 | 0 |
| Live, fully functional lineages at generation6 | 4 | 0 | **4** | 0 |
| Correct post-initial routes following descendants | 20 | 0 | **28** | 0 |
| Unique complete requests | 96 | 36 | 96 | 35 |
| Generated tokens | 424 | 291 | 2,786 | 1,382 |

Copy's analogous secondary endpoint also passed. Post-initial inherited advantages
were +30 for copying and +57 for writing, a registered difference of +27; the
functional-lineage contrast was zero. Writing's instruction, protocol and output
burden change together, so the higher reward does not isolate a beneficial effect
of rewriting itself.

| Generation | Copy + founder | Copy, no founder | Write + founder | Write, no founder |
|---|---:|---:|---:|---:|
| 1 | 15 | 6 | 21 | 6 |
| 2 | 12 | 6 | 15 | 6 |
| 3 | 12 | 6 | 18 | 6 |
| 4 | 12 | 6 | 18 | 6 |
| 5 | 12 | 6 | 18 | 6 |
| 6 | 12 | 6 | 15 | 3 |

Writing's total per-map resource advantages were **+27,+6,+3,+36**;
its generations2–6 advantages were **+21,+3,+3,+30**. Thus its positive pooled
advantage was accompanied by positive advantage on each of these four maps, though
the effects were unequal. Copy's total per-map contrasts remained +9,−6,+6,+30.

The copy control repeated I5 exactly: all 96 seeded requests and replies matched
I5's maintenance arm, and all 96 empty-board requests and replies matched I5's
empty-board arm. This is a contemporaneous repeated anchor, not new input coverage.
Repeated complete requests within each I6 arm also gave identical replies. These
deterministic panel observations are not 384 independent estimates of population
reliability.

## Fidelity, refusal and efficiency boundaries

All **43/43** seeded writing attempts supplied schema-valid full tables and executed.
There were zero missing payloads, malformed payloads, changed entries relative to
parents, founder-content deviations or changed composed destinations. At every
generation start the four seeded writing records provided all 32 correct entries.
Writing renewals by map were 11,11,12,9. No lineage became empty through generation6;
survival is right-censored beyond this horizon, with no measured half-survival time.

The no-founder writing arm made 19 raw REFRESH attempts using numeric identifiers
such as `REFRESH 1 {...}` or `REFRESH 0 {...}` instead of concrete n-ids. The complete
responses were rejected as `MalformedRoute`: they did not satisfy the outer action
grammar and never reached payload validation. Hence its registered `valid_payloads`
and `malformed_payload` counters are both zero despite visible JSON in those invalid
responses. Raw responses are preserved; no first-action extraction or identifier
repair was applied. Copy's empty-board arm likewise repeated I5's one `REFRESH 1`
grammar failure. Neither control created a record.

Persistence was still inefficient. Twenty of the 43 writing renewals were redundant
within their generation; 21 of 44 copy renewals were redundant. Both seeded arms
made four renewals during the final generation despite no later payoff in the
registered horizon. Writing earned 105 versus the scripted faithful maintainer's
228 resources, while copying earned 75. Faithful content and useful availability
therefore do not establish efficient maintenance or accurate routing on every job.

The full current table remained visible on every renewal, and the writing instruction
explicitly asked agents to preserve useful routing information. This measures
successful written transmission under that support, not reconstructive memory,
error correction, selection against mutations, spontaneous discovery of renewal,
or robustness to lossy/hostile inputs. No mutation was imposed and none occurred;
there is no empirical mutation-repair rate to estimate. The explicit institution
and collective objective remain supplied by the harness, and agents act sequentially.

## Evidence and closure

All **22 world gates** and **12 frozen audit checks** passed. The audit independently
rebuilt all requests, private-memory boundaries, action parsing, supplied payloads,
record transitions, expiry and rewards, then checked metric replay, primary arithmetic,
metadata and chronology. The run used **4,883 generated tokens** in **1,355.010 seconds**,
with no hidden thinking or API calls. Registered limits were 384 calls, 60,000 total
tokens, 3,600 seconds, and per-response 512 tokens / 2,048 bytes / 120 seconds.

Full-result SHA256:
`a1ef383c7c6513be137a0fefb4c73d8765b947e2dcde837b2b20da47de1a145d`.
The canonical record carries named numeric results, this write-up, conclusion,
preregistration, frozen source, gates, full result and a durable evidence bundle
containing raw calls and audit. `server.json` is the local receipt for the verified
server identifiers, hashes and lifecycle states.

I6 is closed without extending the horizon or introducing post-result corruption.
A separate prospective experiment would be needed to test whether useful information
survives controlled damage, partial visibility or a longer transmission chain.