# I8 — complete diagnostic evidence increased copying without repair

**The registered repair primary failed.** At the start of generation5, neither
condition had a fully corrected descendant: **0/4 with diagnostics versus 0/4
without**, a contrast of zero. No partial functional improvement, entry change or
full repair occurred anywhere in the run. All **32 diagnostic-branch renewals**
preserved the damaged source table unchanged, as did all 24 control renewals.

The diagnostic panel changed behaviour without correcting the shared record.
During exposure, Qwen requested **REFRESH in all 16 decisions**, compared with
8 ROUTE and 8 REFRESH in control. Exposure resources were **0 versus 9**. After
diagnostic withdrawal, both branches earned **21 resources**, with the same
per-map totals. The secondary useful-repair-persistence conjunction also failed.

This removes absence of identifying information as a description of what the
harness supplied. It does not establish that the model successfully decoded that
information, preferred known errors, or is generally incapable of repair.

## Prospective record and chronology

Canonical [I8](https://chuk-experiments-server.fly.dev/#/experiments/ecology-inheritance-i8),
`EXP-20260913-202834-00836`, run `RUN-20260913-203908-00934`, programme
`cell-native-architectures`. Registration commit:
`9949cb85d4ef3d00c36521c9a5736c5285affc9a`.

Source and registration were uploaded and verified at 2026-09-13T20:39:55.571826Z,
before gates completed at 20:40:15.414085Z. All 27 passing gates and the generated
diagnostic grids were uploaded and verified at 20:42:31.432060Z. Inference ran from
20:42:49.882017Z to 20:49:17.565211Z. The frozen independent audit passed all 13
checks at 20:49:42.187335Z. No source or endpoint was changed after inference.

All four starting records are actual I7 corrupted descendants at generation4,
from `RUN-20260913-144529-00925`, full-result SHA256
`e850edd6b205d7de4b110c1bdba8bece9ff709ed8da68b1b369b93861549c14e`.
Their GPT-5.5/I4 → Qwen/I6 → corrupted Qwen/I7 history remains historical. No
producer, earlier generation or new corruption was sampled in I8.

## Information intervention

Each real starting state was cloned unchanged into `no_diagnostics` and
`diagnostics`. The true world, damaged record, id, expiry, next-id counter,
objective and action mechanics initially matched. Both used one common system
description and the same added DIAGNOSTICS observation field.

For each map, the harness executed all sixteen signal–destination combinations
against the unchanged device, in a frozen order. Each public row reported signal,
requested ROUTE, execution and earned resources. Exactly one destination earned
3 for each signal; the other three earned 0. All trials were shown, not only
successes. Their resources were excluded from the population ledger. No stage table,
target field, correction instruction or repaired record was supplied in the panel.

During generation4, only the diagnostic branch received the sixteen rows. The
control received an empty list. Both branches received an empty list throughout
generations5–6. Every fresh agent started with empty OWN and received only its own
first-action receipt on its second turn. No diagnostic history was retained in
OWN or passed separately to later agents. The common system still described the
DIAGNOSTICS field after withdrawal; it contained no map-specific calibration.

The full current shared table stayed visible. REFRESH still required the complete
JSON payload, consumed a job opportunity, earned zero and extended the descendant's
expiry to the end of the next generation. Schema-valid wrong tables remained
acceptable without truth checking. Empty boards could not be recreated.

## Main results

Four maps × two conditions × three generations × two agents × two decisions:
**96 local calls**, all parsed and executed, with no retries, repairs to output,
transport failures or execution refusals.

| Measure | No diagnostics | Diagnostics in g4 |
|---|---:|---:|
| Calls / parsed / executed | 48 / 48 / 48 | 48 / 48 / 48 |
| Fully correct new descendants at g5 | **0** | **0** |
| Fully correct new descendants at g6 | 0 | 0 |
| Partial improvement / full repair | 0 / 0 | 0 / 0 |
| Successful renewals | 24 | **32** |
| Unchanged copies of damaged source | 24 | **32** |
| Needed renewals at action time | 12 | 12 |
| Correct routes, exposure g4 | 3 | 0 |
| Correct routes, withdrawal g5–g6 | 7 | 7 |
| Resources, exposure g4 | 9 | **0** |
| Resources, withdrawal g5–g6 | **21** | **21** |
| Total resources | 30 | 21 |
| Live lineages at each generation start | 4,4,4 | 4,4,4 |
| Correct composed destinations across all four records | 8,8,8 | 8,8,8 |
| Generated tokens | 1,544 | 2,000 |

The primary contrast, diagnostic-minus-control corrected lineages at g5, was **0**.
No exposure repair existed to transmit, so there were also zero correct withdrawal
routes attributable to descendants of an exposure repair. The positive-preservation
phenomenon continued, but the records remained only 2/4 functionally correct.

| Map | Control total | Diagnostic total | Control withdrawal | Diagnostic withdrawal |
|---|---:|---:|---:|---:|
| 0 | 9 | 6 | 6 | 6 |
| 1 | 12 | 9 | 9 | 9 |
| 2 | 0 | 0 | 0 | 0 |
| 3 | 9 | 6 | 6 | 6 |

Generation totals were 9,12,9 in control and 0,12,9 in diagnostics. All of the
aggregate −9 diagnostic contrast arose during exposure. There was no measured
withdrawal reward difference, pooled or by map. Correctness did not improve in
either branch, and no record went extinct during the observed g4–g6 horizon.

## Changed behaviour, unchanged information

Each of the four diagnostic cohorts used all four generation4 opportunities for
REFRESH. Its control used ROUTE→REFRESH for both agents. At the first fresh decision
of each map, system and all observation fields except DIAGNOSTICS matched exactly:
adding the trial rows switched **ROUTE→REFRESH in 4/4 such pairs**. This descriptive
paired check is additional to the registered repair endpoint. Later paired inputs
also differed in resulting record ids and own-action receipts, so the whole
trajectory is not a collection of otherwise-identical single-field comparisons.

None of the extra writes modified an entry. Each diagnostic lineage made eight
renewals over g4–g6, versus six in control. Twenty of the diagnostic branch's 32
renewals were redundant within their generation, versus twelve of 24 in control.
Both branches also made eight renewals in the terminal generation, without a later
payoff inside this horizon. The additional exposure writes consumed opportunities
without correcting or extending the records beyond the same renewal lifetime.

After withdrawal, agents returned to ROUTE→REFRESH. The shared contents remained
identical between conditions, while diagnostic-branch record ids were further
advanced by the extra copying. Thus equal withdrawal reward is not a claim of
byte-identical requests or private behavioural retention.

## What the null does and does not establish

In I7, a zero-reward route did not identify the correct destination. Here the
exhaustive test grid supplied enough evidence to identify all four destinations.
A scripted policy using only those public rows and the visible stage1 mapping
repaired all four records. With subsequent maintenance it earned 108 resources
versus 57 without diagnostics, including 84 versus 42 after withdrawal. The repair
channel, its informational sufficiency and its later utility were all demonstrated
at zero inference before the model run.

The live model never routed while the diagnostics were visible. Consequently I8
does not demonstrate that Qwen decoded the diagnostic rows, resolved their conflict
with the board, or formed an accurate replacement mapping. Failure to use this
representation and failure to convert understood evidence into a corrective write
remain separate possibilities. More context, attention to repeated structured
material, the existing preservation instruction and other presentation effects
also remain unisolated: panel content, presence, length and salience change together.

The supported result is narrower: **this complete diagnostic panel elicited more
unchanged copying, not shared-record correction or later advantage**, under the
frozen interface. Do not describe it as knowingly choosing falsehood, established
irrationality, universal repair incompetence, or selection failing. There is still
one record per branch, no competition and no fitness-weighted reproduction.

## Validation and closure

All **27 world gates** and **13 frozen audit checks** passed. Besides the public-evidence
repair control, gates separated direct diagnostic routing without writing, repair
followed by expiry, and temporary repair overwritten before withdrawal. The audit
independently reconstructed public diagnostic rows, their complete withdrawal,
requests, private-memory boundaries, payloads, world transitions, expiry, rewards
and the primary, and verified metric replay, metadata, chronology and bounds.

**96 calls, 3,544 generated tokens, 387.683 seconds.** No paid inference, hidden
thinking, output salvage or exceeded bound. Full-result SHA256:
`45078d1ef8761941a27b575c0235487cd00573ea8c1db3251ea7c7675437a6d5`.
Canonical evidence includes numeric metrics, this write-up, conclusion, registration,
frozen source, probes/gates, full result and a durable raw-call/audit bundle.
`server.json` records verified server identifiers, hashes and lifecycle states.

I8 is closed. A separate diagnostic-use competence assay could establish whether
Qwen can recover a destination from the same test-grid representation before
attributing this null specifically to shared-record maintenance. No wording repair,
extra inference or stronger-model substitution is added to the completed I8 panel.