evo-exp — Shadow Interface Evolution experiment archives

Complete, self-contained records of OSWorld-V2 interface-evolution runs: the evolved interface code, every agent trajectory, the builder logs, the evolution database, the run configuration, and derived result tables. Each archive is meant to be readable by someone with no access to the machine that produced it.

Three experiments are archived here. They differ in exactly one intended variable each, and the differences are listed before the numbers because most of the numbers are not comparable across them.

---

Which archive to take

file	size	contents	status
archives/qwen_syn30v15_128k-step15-slim.zip	95 MiB	no scratch VM, 16 variants, steps 0–15 — scores, code, config, derived tables	current
archives/qwen_syn30v15_128k-step15-redacted.zip	13.8 GiB	the same plus all 5,630 raw agent transcripts	current
archives/qwen_syn30v15_128k_scratch-step10-slim.zip	94.9 MiB	with scratch VM, 11 complete variants	current
archives/qwen_syn30v15_128k_scratch-step10-redacted.zip	5.6 GiB	the same plus transcripts, plus a partial 12th variant (see below)	current
archives/qwen_syn30v15_128k-bare-redacted.zip	1.0 GiB	no shadow interface at all — the baseline floor	current
archives/upstream_qwen_108-20260925T020227Z-keyframes.zip	138 MiB	upstream harness, no shadow interface — all 108 tasks, every attempt's full transcript, key screenshots only	current
archives/upstream_qwen_108_merged_vol60fill-20260925T164249Z-keyframes.zip	96 MiB	upstream harness, merged row — the 108 above with its 10 unscored tasks filled by a --volume_size 60 rerun; one counted attempt per task, key screenshots only	current (mixed conditions, see Run 5)
archives/if007_scratch_merged_108_20260925-redacted.zip	1.3 GiB	held-out + val eval of one evolved variant — qwen_syn30v15_128k_scratch if-007 on all 108 tasks; one counted episode per task, full transcripts, no recordings	current (5 tasks re-run, see Run 6)
archives/qwen_syn30v15_128k-step12-redacted.zip	10.6 GiB	no scratch VM, steps 0–12	superseded by step15
archives/qwen_syn30v15_128k-step12-slim.zip	96 MiB	no scratch VM, steps 0–12	superseded by step15
archives/qwen_syn30v15_128k-redacted.zip	3.3 GiB	no scratch VM, steps 0–5	superseded by step15
archives/qwen_syn30v15_128k-slim.zip	81 MiB	no scratch VM, steps 0–5	superseded by step15

The four superseded archives are kept for link stability. step15 is a strict superset of both step12 and the unsuffixed pair — same namespace, same config, more steps. Take step15 unless you need a byte-identical copy of what an earlier analysis read.

Take a slim archive to read results or redo analysis: all scores, code, configuration and derived tables are in it. Take a redacted (full) archive only to replay what the agent actually did, step by step.

Integrity

5ccfa693ba3e37d2122e32f6fffe2d3eb0573d4623cdffc763132afcbc832b9f  qwen_syn30v15_128k-step15-redacted.zip
3f97ecd726d0c38a61aa1b6d795c74c53a070594d865fdc484114617174c82a9  qwen_syn30v15_128k-step15-slim.zip
bbc2b8258d64cda3c0167e826bf3c512f52fa99ec81dc837e3395a71676997f9  qwen_syn30v15_128k_scratch-step10-redacted.zip
8a31ac3f19cd00d2668c1456f62fb92088135aea3ebc7b228fe67fe260209f38  qwen_syn30v15_128k_scratch-step10-slim.zip
095c08ab479f0efaf1235db50186a693275369fe12e63698d0e4696d6d897c48  qwen_syn30v15_128k-bare-redacted.zip

Credentials are redacted. An OSWorld agent explores the VM it is given, and one that runs env records the harness's API keys into its own transcript. This is not an occasional accident: grounding runs as a tool, so the agent's model credentials are deliberately readable by any bash action it takes. Every text file here has been passed through a redactor that replaces such values with <REDACTED> while keeping the variable name. For step15 that rewrote 4,099 generic assignments, 2,367 Hugging Face tokens, 109 Anthropic / 99 OpenRouter / 99 LiteLLM-proxy / 98 Parasail keys, and 55 private-key blocks or headers. Every archive was scanned before upload with a detection list wider than what the redactor rewrites, and none contains a credential-shaped string.

---

The settings

All three runs share this. Deviations are called out per run.

	value
harness	Prime Agent (--builder-harness prime_agent, evaluator_harness=prime_agent)
builder model	openai/simular-qwen-3p8-27b-evolving-agents (self-hosted Qwen 3.8-27B)
eval model	the same deployment — a variant is written and graded by one model
grounding model	also the same deployment. The evaluator exports SHADOW_GROUNDING_MODEL=<eval model>, so the interface's UI-grounding helper calls the island's own model. There is no separate grounding model. (Earlier project runs used Parasail UI-TARS; PARASAIL_API_KEY survives in the tool allowlist.)
context window	131,072 (128K), declared in shadow_interface/context_compaction.json with context_window_source = interface
train set	synthetic_train_30.json — 30 synthetic s_* tasks, UNGRADED
val set	osworld_v2_val_15.json — 15 graded tasks: 007 010 014 019 041 046 048 058 064 069 072 079 083 105 108
selection signal	val
parent selection	tournament
scheduler	deterministic:0,2,0 (alternating mutation / crossover)
elimination count	5 — no elimination event was ever recorded; every variant stays in the pool
per-task wall clock	7,200 s builder, 7,200 s eval
max agent steps	300
env concurrency	45
region	us-east-1, public IP
website deployment	52.91.237.94.nip.io — a self-hosted substitute, not the pinned site.hku.icu
egress proxy	allow_placeholder_proxy=1 — the proxy config holds placeholder credentials

Two settings that bound what these numbers can support

The proxy is not configured. With placeholder credentials, 6 train tasks (037 055 062 075 098 and 056) and 2 test tasks (036 050) fail for lack of an egress proxy. Those failures are indistinguishable from genuine interface failures, so they bias every selection step. The orchestrator says so itself on every launch: "This run is a diagnostic, not a reportable result."

The websites are a substitute stack. 40 of the tasks are website-backed. Scores on them are comparable only to other runs on this same substitute deployment — not to any run measured against the pinned site.hku.icu host.

---

Run 1 — qwen_syn30v15_128k: no scratch VM

--no-scratch-vm: the builder writes the interface without a VM to try it in. --max-steps 30, so this archive is a mid-run snapshot at step 15, not a finished run.

16 variants. val partial is the summed raw score over the 15 val tasks; val binary counts only full credit.

step	variant	op	parent(s)	val partial	val binary
0	seed	—	—	5.375 / 15 = 35.8%	1/15
1	if-001	mutation	seed	5.851 / 15 = 39.0%	1/15
2	if-002	crossover	if-001 × seed	5.540 / 15 = 36.9%	1/15
3	if-003	mutation	if-002	5.255 / 15 = 35.0%	1/15
4	if-004	crossover	if-002 × if-001	5.200 / 15 = 34.7%	3/15
5	if-005	mutation	seed	4.575 / 15 = 30.5%	3/15
6	if-006	crossover	if-001 × seed	5.600 / 15 = 37.3%	4/15
7	if-007	mutation	if-001	7.045 / 15 = 47.0%	2/15
8	if-008	crossover	if-001 × seed	7.180 / 15 = 47.9%	3/15
9	if-009	mutation	if-008	6.070 / 15 = 40.5%	3/15
10	if-010	crossover	if-001 × if-009	5.965 / 15 = 39.8%	3/15
11	if-011	mutation	if-002	6.731 / 15 = 44.9%	4/15
12	if-012	crossover	if-001 × if-010	4.738 / 15 = 31.6%	2/15
13	if-015	mutation	if-012	5.607 / 15 = 37.4%	3/15
14	if-016	crossover	—	6.645 / 15 = 44.3%	3/15
15	if-017	mutation	—	6.673 / 15 = 44.5%	3/15

if-013 and if-014 do not exist. On 2026-09-22 an expired AWS SSO token made VM allocation fail for all 45 tasks of both variants. The harness wrote those retry-exhausted tasks as real 0.000 scores with infrastructure_failed: 0, so they entered the DB as legitimate worst-in-pool individuals. They were purged and the two steps re-dispatched; the if- counter did not rewind, so step 13/14/15 carry if-015/016/017.

Other facts for this run: 5,630 trajectory files, 4,186 held-out files across 101 held-out tasks / 671 runs / 42 rerun pairs, and 65 of 240 val runs (27.1%) truncated by the 2 h wall clock.

---

Run 2 — qwen_syn30v15_128k_scratch: with scratch VM

Identical to Run 1 except that the scratch VM is enabled (the builder gets a VM to try the interface in while writing it) and --max-steps 15. Every other flag matches, which makes these two runs a clean A/B on the scratch VM.

dispatch_log records steps 1–10 complete. The variants table holds only 6 rows: seed, if-003, if-008, if-009, if-010 and if-011 were evaluated but their DB rows never landed. Their trajectories, git tags and builder logs are all present, so they are fully auditable; their lineage below was recovered from each tag's commit subject, and they are listed in MANIFEST.json → db_orphan_variants. Their step and val_score fields are empty in the DB-derived tables — the real scores are in results/val_matrix.csv, computed from result.txt.

step	variant	op	parent(s)	val partial	val binary	DB row
—	seed	—	—	5.375 / 15 = 35.8%	1/15	missing
1	if-001	mutation	seed	5.485 / 15 = 36.6%	1/15	✅
2	if-002	crossover	if-001 × seed	6.125 / 15 = 40.8%	3/15	✅
—	if-003	mutation	if-001	4.375 / 15 = 29.2%	1/15	missing
4	if-004	crossover	if-001 × seed	6.561 / 15 = 43.7%	3/15	✅
5	if-005	mutation	if-004	5.790 / 15 = 38.6%	4/15	✅
6	if-006	crossover	if-001 × if-004	6.220 / 15 = 41.5%	3/15	✅
7	if-007	mutation	if-001	7.197 / 15 = 48.0%	3/15	✅
—	if-008	crossover	if-005 × if-007	5.050 / 15 = 33.7%	0/15	missing
—	if-009	mutation	if-006	4.895 / 15 = 32.6%	1/15	missing
—	if-010	crossover	if-005 × if-002	5.150 / 15 = 34.3%	2/15	missing

There is no step 3 row in dispatch_log's variant output either; if-003 is the step-3 child.

The full archive holds one extra, incomplete variant

qwen_syn30v15_128k_scratch-step10-redacted.zip contains a 12th variant, if-011, with only 2 of 45 result.txt files. It appeared on disk between the slim and full packaging runs and was picked up mid-evaluation. Do not read it as a result. The slim archive holds the 11 complete variants only. The archive name reflects the last completed dispatch step (10).

Other facts: 3,932 trajectory files, 458 held-out files across 52 tasks / 133 runs / 0 rerun pairs, and 37 val runs truncated.

---

Run 3 — qwen_syn30v15_128k-bare: no shadow interface

The floor the whole experiment is measured against. Prime Agent with nothing deployed to the VM: no shadow interface, no setup.sh, no eval constraints, and the stock NO_TOOL_SYSTEM_PROMPT instead of TOOL_SYSTEM_PROMPT.

Three consequences:

1. --no_tool moves the action space, and with it the output directory, to shadow_eval_no_tool/. A reader that hardcodes the shadow_eval/ prefix reports this run as empty.
2. The leak guard is off. A baseline has no training set to leak from, and the point is to cover val_15 too, so the number is comparable to the 108-task denominator.
3. The proxy is off for this run, unlike Runs 1 and 2 — the orchestrator is already in the US, and a controlled probe showed the proxy changes no score while placeholder credentials only force failures.

108 held-out tasks | 190 runs | 6 rerun pairs | 935 files
115 runs carry a score: mean 0.3845, 17 at full credit

Batches: osworld_v2_full_108 (55), osworld_v2_bare_remaining (92), osworld_v2_bare_imputed_rerun (1), and _aborted_hku (42 — an abandoned attempt against the pinned host, kept and prefixed).

Read this before claiming the evolution works. The bare floor's partial mean of 0.3845 sits inside the 0.29–0.48 band that every evolved variant in Runs 1 and 2 occupies. The task sets and denominators differ, so this is not a head-to-head — but results/heldout_matrix.csv now carries a bare column, and aligning on the shared tasks is the comparison that decides whether an evolved interface beats having no interface at all.

---

Archive structure

Every archive unpacks to one directory named <namespace>-<UTC timestamp>/.

<namespace>-<timestamp>/
├── MANIFEST.json            provenance and self-description — read this first
├── config/                  what was run
├── db/                      the evolution SQLite database
├── results/                 derived tables (read these second)
├── interfaces/              the evolved code, per variant, plus a git bundle
├── builder_logs/            one JSONL per builder invocation
├── logs/console.log         the orchestrator's own output
├── trajectories/            <variant>/<task>/ — val + synthetic-train runs
└── trajectories_heldout/    <variant>/<task>/<eval set>__<stamp>/ — held-out runs

MANIFEST.json

key	meaning
namespace, created_utc, eval_model	which run, packaged when, scored by which model
variants	rows from the DB: id, origin, step, parent_id, parent_b_id, score, val_score, tag
db_orphan_variants	variants archived from disk that have no DB row (Run 2)
val_task_ids	the 15 graded task ids
task_cap_seconds	the per-task wall clock (7200) used to flag truncated runs
truncated_val_runs	how many val runs hit that wall clock
held_out_tasks / _runs / _rerun_pairs	held-out coverage; rerun pairs are the only direct measure of harness noise here
redacted, redactions	whether redaction ran, and how many substitutions per rule
slim, recordings_included	which variant of the archive this is
run_config	the full launch configuration

config/

run_config.json (full command line, models, timeouts, scheduler, parent selection, concurrency, AWS region/subnet/security group, seed SHA-256s), eval_sets/osworld_v2_val_15.json, eval_sets/synthetic_train_30.json, and seed_context_compaction.json (context_window_tokens, reserve_tokens, keep_recent_tokens).

db/

shadow_evolution_osworld_v2_<namespace>.db, SQLite. Tables: variants, eval_results (per variant per task), cost_events, dispatch_log, history, islands, eval_sets, val_set.

select id, origin, step, parent_id, parent_b_id, round(val_score,4)
from variants order by rowid;

eval_sets is empty in these runs — that table was never written; the task sets live in val_set and in config/eval_sets/.

results/

Generated at packaging time from the trajectories, so they agree with the raw files by construction.

file	contents
val_matrix.csv	15 rows, one per val task, one column per variant
variants.csv	lineage and final scores
truncated_runs.csv	read this before interpreting any zero
heldout_runs.csv	one row per (variant, run, task) — reruns included
heldout_matrix.csv	each task resolved to one score per variant, latest run wins
heldout_reruns.csv	tasks run more than once, with abs_delta

interfaces/

A git bundle carrying every variant tag and the commit graph, plus flat per-variant snapshots of shadow_interface/. The bundle is authoritative — it shows how each variant was derived from its parent, which the snapshots cannot.

git init evolved && cd evolved
git fetch ../<namespace>-<timestamp>.bundle 'refs/tags/*:refs/tags/*'
git log --oneline --graph --all
git diff evol/<ns>/seed evol/<ns>/if-001 -- shadow_interface

The bundles here record a complete history, including the upstream OSWorld commits the seed was cut from.

trajectories/ and trajectories_heldout/

Each task directory holds result.txt (the score, one float), result.json (grading detail including safety checks), traj.jsonl (one-line run summary with elapsed_seconds, hitl_turns), claude_trajectory.jsonl (the raw transcript — omitted from slim archives), claude_stderr.log, exit_code.txt (124 means the wall clock killed it), instruction.txt, env_spec.json, and sometimes phase_results.json or failure.json.

Token accounting lives on each turn_end event:

{"input": 5179, "output": 137, "cacheRead": 0, "cacheWrite": 0,
 "totalTokens": 5316, "cost": {"input": 0, "output": 0, "total": 0}}

cost is zero throughout: the model is a self-hosted deployment that declares zero per-token price. That is structural, not a measurement gap. Counting turn_start events gives the step count.

A failure.json next to a result.txt means a later retry succeeded and the failure record is stale; only a failure.json with no result.txt is a real gap.

---

Caveats

Beyond the two setting-level limits above, five things are easy to read wrong. All are recorded in MANIFEST.json.

1. Synthetic s_* tasks are UNGRADED — constant zero by construction, verification: dummy_constant_zero. Their result.txt is always 0.0, so the train score is 0.000 for every variant in Runs 1 and 2 and carries no information. Only the 15 val ids are graded, and the denominator is the full val set even when fewer finished.

2. A truncated run is not a capability zero. A task cut off by the 2 h per-task wall clock exits 124 (or reports elapsed_seconds above 7200), writes a real score — almost always 0.000 — and writes no failure.json, so infra-failure checks cannot see it. 27.1% of Run 1's val runs were truncated. results/truncated_runs.csv lists every one.

3. The context window is inside the evolvable interface. Because context_window_tokens lives in shadow_interface/context_compaction.json and that file is listed as evolvable, a child variant can change its own context window. Neither run is a fixed-window single-variable experiment.

4. The seed was not re-evaluated. Its 45 task directories were copied in from two earlier 128K runs rather than measured again, so the seed row is not comparable to its children on the truncation axis. Both runs' seeds report the same 5.375, which is that same cached measurement.

5. val at n=1 cannot resolve this ladder. Each variant's fitness is one run of 15 tasks. Across this project's earlier runs the spread between repeated evaluations of the same parent pair reached 1.64 points on the same 15-task scale — wider than most of the gaps in the tables above. Independent evidence: in the project's prime25v15r4 run, the two variants that also received a 68-task held-out evaluation ranked in the opposite order on the larger set (if-006 62.4% vs if-008 51.7%) from how they ranked on val_15. Treat single-variant val differences here as noise unless a held-out set confirms them.

---

Adding a run to this repo

The archives are produced by the evo-archive skill in the source repository: package_experiment.py builds the zip from an allowlist of filenames (never a directory sweep) with redaction on by default, and upload_experiment.py scans the result and refuses to upload if it finds a credential shape. Archives are keyed by namespace and name, so adding one overwrites nothing.

Run 4 — upstream_qwen_108: the upstream harness, for comparison

The same model under xlang-ai/OSWorld-V2's own Qwen scaffold (mm_agents/qwen + scripts/python/run_multienv_qwen.py) instead of the Prime Agent, on the same 108 tasks. It is the reference point the interface rows are measured against: not "no shadow interface" like Run 3, but a different harness entirely.

denominator	n	binary	partial
every failure counts 0 — recommended	108	6 = 5.56%	22.348 = 20.69%
drop only failures that were infrastructure in every pass	102	6 = 5.88%	22.348 = 21.91%
… also dropping the 3 unattributable ones	99	6 = 6.06%	22.348 = 22.57%

98 of the 108 tasks are graded. Four passes were run; the archive holds all 163 attempts, not just the scoring one.

Why 10 tasks have no score, and why it counts against this row

Eight died on the host-side screenshot path and two on the model's own malformed tool XML. The screenshot failure is deterministic, not infrastructure noise: across four passes the crashed set is strictly nested (108 → 31 → 14 → 10 → 10) and no task ever entered it, and 5 of the 8 deaths are monotonic — the guest serves 9–27 frames, then returns 500 twenty-four times running and never recovers.

The cause is ENOSPC. All 136 failed /screenshot bodies read OSError: [Errno 28] No space left on device. desktop_env.py:146 expands the guest partition only for tasks that declare volume_size, so these guests sat at 29.5 GB with ~400 MB free while snapd's boot-time auto-refresh grew /var/lib/snapd to 11–13 GB. Upstream's own --volume_size 60 fixes it — that is what the Prime-Agent runs already default to, which is why they are unaffected (59 GB with 26–28 GB free).

A defect in a component only one architecture depends on is that architecture's failure: upstream's agent runs on the host and pulls a frame through GET /screenshot every step, while an in-guest agent scores all 10 of these tasks normally. Hence the flat 108.

Caveats specific to this run

• 43 of the 98 graded tasks saw at least one screenshot 500 during their scoring attempt, so their episode ran on a guest that reached zero free disk even though a score came out (070 and 073 scored 1.0 with four 500s each). Exposure tracks episode length — median 144 steps against 112 — so the 55 unexposed tasks are the shorter ones, not a random subsample. "Never saw a 500" means "not observed to have hit zero", which is weaker than "ran with headroom".
• The snap auto-refresh runs on every guest of both harnesses, refreshing a varying number of snaps (7, 8, 10 and 11 observed), so which application versions a task faces depends on boot time and on the snap store at that moment. This is uncontrolled across every row in this repository. It is not an extra error term — the measured rerun noise floor (mean |Δ| 0.186 per task) already operated under it — but it is a plausible mechanism behind it.
• Agent-side flags are upstream defaults (max_steps 500, max_tokens 32768, image_max 20). Only infrastructure retry budgets were patched: HTTP statuses ×8 and PythonController.retry_times 3 → 24. All 98 scores come from patched runs, so the row is homogeneous on that axis.
• Step budget differs from the Prime rows: 500 here against 300 there. 12 of 108 tasks exceeded 300 steps and none scored 1.0, so the confound is worth at most ≈2.3 pp of partial and nothing on binary.
• Thinking is on by default and upstream's enable_thinking flag is inert on this deployment — only chat_template_kwargs is read. Reasoning tokens are generated and billed but never returned, so completion_tokens is not visible text. The deployment's context window is 262144, and no request in any pass was rejected for length.

What -keyframes means here

Unlike the -slim archives above, no transcript is omitted. All 163 attempts carry their full traj.jsonl, runtime.log and eval.log. What is trimmed is step screenshots: 19,494 PNGs are 7.8 GB against a few tens of MB of everything else, so the archive keeps every 0-byte PNG plus the last two non-empty frames of each task. The 0-byte frames are not corruption — they are the fingerprint of the crash above, since lib_run_single.py:948 opens the file "wb", truncating it, before the write raises on a None screenshot.

Redaction note

OSWorld task instructions state credentials in prose ("the password is '…'"), a shape the standard redactor — which matches NAME=value — cannot see. This archive was additionally passed through a prose-aware rule; 1,818 values are replaced with <REDACTED> while the surrounding sentence and account names are kept, so the trajectories still read as the task the agent was given.

Run 5 — upstream_qwen_108_merged_vol60fill: Run 4 with its 10 gaps filled

Run 4 left 10 of 108 tasks without a score. This archive keeps Run 4's 98 graded tasks exactly as counted there and fills the 10 gaps (020 037 041 050 062 074 097 100 103 104) with a rerun that passes upstream's own --volume_size 60. It holds one directory per task, the attempt the row counts, plus scores.csv (task, source, root size, score).

row	n	binary	partial
Run 4, as run (10 unscored = 0)	108	6 = 5.56%	22.348 = 20.69%
Run 5, merged	108	6 = 5.56%	23.405 = 21.67%

The 10 filled tasks: 104 = 0.45, 103 = 0.25, 097 = 0.25, 020 = 0.1, 062 = 0.008, 037 = 0.0, 041 = 0.0 (hit the 500-step cap), 050 = 0.0, 074 = 0.0, 100 = crash (malformed tool XML, counted 0).

Why the 10 were unscored

The guest root filled up. snapd auto-refreshes ~10 stale snaps at boot into /var/lib/snapd (11–13 GB), and desktop_env.py only grows the root partition when volume_size is set, so these guests stayed at 29.5 GB on a 40 GB disk with ~400 MB free. All 136 failed /screenshot bodies in the diagnostic pass were OSError: [Errno 28] No space left on device. With --volume_size 60 the guests run 59 GB with ~31 GB free, and the rerun saw zero screenshot 500s.

Caveats

• Mixed conditions. 98 tasks ran on the unexpanded 29.5 GB root, 10 on 60 GB. This is neither the as-shipped row (Run 4) nor a uniform --volume_size 60 row; a uniform 108 rerun is in progress.
• 43 of the 98 Run-4 scores came from attempts that hit at least one screenshot 500, i.e. the disk reached zero mid-episode.
• 050's website fetches returned HTTP 407 from the egress proxy, so its 0 is not a capability measurement.
• The rerun was relaunched once after the LiteLLM proxy's public IP changed; unfinished tasks restarted from step 1, and the two tasks graded before that were checked for Connection-error endings (none).
• Redaction: 30 NAME=value values and 607 prose-form credentials replaced with <REDACTED>; an independent wider scan found no credential shape.

---

Run 6 — if007_scratch_merged_108: if-007's full 108-task evaluation

The evolved interface evol/qwen_syn30v15_128k_scratch/if-007 (from Run 2), run with Prime Agent on simular-qwen-3p8-27b-evolving-agents over all 108 tasks: 93 held-out (heldout_93/) plus the 15 val tasks the evolution used as fitness (val_15/, taken from the evolution eval cache). One directory per task, the episode the row counts, plus scores.csv (score, source run, harness, replaced score and reason per task), MANIFEST.json, and the interface code (shadow_interface_if-007/).

row	n	binary	partial
as run, no replacements	108	16 = 14.81%	47.204 = 43.71%
strict: only 066 replaced (same harness as seed)	108	17 = 15.74%	48.204 = 44.63%
merged: all 5 replacements	108	17 = 15.74%	50.286 = 46.56%

For reference, Run 1's seed scores 47.121 partial on the same 108 tasks, and Run 3 (bare) 47.526.

Which episode each task counts

• Held-out: the latest completed run of full_minus_val_93 → remaining_41 → ce_rerun_4 → 499_rerun_1. The last two re-measure 042 067 094 106, whose episodes were cut by infrastructure (a Connection error while the LiteLLM proxy was resized, then a 499 Client Closed Request on 042).
• Five tasks count a re-run instead of the original episode:
  • 043 063 080 086 — the original ended on a reply that hit the 32k output cap with no tool call (stopReason=length), which Prime's loop treats as a finished episode, so the task was scored mid-thought (0.0625, 0, 0, 0). Re-run with SHADOW_EVOL_LENGTH_CONTINUATIONS=3: the harness resumes the same session after such a reply. Only 043 actually resumed (twice); 063, 080 and 086 did not truncate on the re-run. New scores 0.6875, 0.7, 0.657, 0.1.
  • 066 — the original ended on 422 No valid tool call after 6 attempts (0); plain re-run, 1.0.

Caveats

• Harness difference. The four length-continuation re-runs used a harness the seed and bare rows did not. Compare with seed on the strict row.
• n = 1 per task. Re-running one variant moves a task by ~0.19 on average in this namespace; single-task differences are not findings.
• The length truncation is task-driven, not specific to if-007: 043 truncates under seed and bare too.
• Recordings omitted. Redaction: 412 NAME=value credentials replaced with <REDACTED>; the wider pre-upload scan found no credential shape.