LOCAL RESEARCH ARTIFACT - EYE172
AI-assisted finite-plan workflow study. Local execution date: 2026-10-10 Asia/Qyzylorda (UTC records may show 2026-10-09). Working paper; human review pending. No peer review, venue acceptance or external preregistration is claimed.

SCOPE
Original coding-agent experiment is not executed. Evaluated models can only return finite-ID JSON plans, a verification flag and completion claim. The AI-assisted, pre-authored trusted Python apparatus DOES execute; code produced by evaluated agents does NOT. Record field no_generated_code_executed refers only to evaluated-agent output, not to the research apparatus authoring process.

ROOTS
Study: D:\MiniProjects_1\02-agent-research\structured-study
Verified portable Ollama/weights: D:\MiniProjects_1\02-agent-research\runtime
Safety control: D:\MiniProjects_1\02-agent-research\v2-study\security\control
Emergency stop: D:\MiniProjects_1\02-agent-research\v2-study\security\STOP-NOW.cmd
All authored task/code/model/temp/result files on D. Installed tools are read from C. OS-managed pagefile, app and driver activity cannot be guaranteed to make zero C writes.

DESIGN
6 fixed templates x 4 structural settings x 5 evaluation instances = 120 parameterized draws/115 distinct semantic payloads. Six recovery draws exactly overlap preflight. Four conditions x three repeats x two related Qwen 0.5B variants = 2880 original assignments. The number of trials is not the number of independent problems. L1-L4 are structural settings, not validated universal difficulty. Development and qualification are separate 24-instance banks; their 96 model trials are excluded from main analyses.

FILES
protocol/freeze.json: actual pre-execution hashes and exact original randomized schedule.
protocol/ANALYSIS-PLAN.txt, SECURITY-AND-OPERATIONS.txt: frozen methods and security scope.
tasks/public.json, author-controls.json: all synthetic inputs and exact grading controls.
controller/: trusted finite interface, semantic oracles, role workflow, fixed model API and independent lifetime supervisor.
analysis/analyze.py: original frozen analysis, whose complete-series final gate is intentionally retained.
analysis/analyze_with_incidents.py: disclosed post-freeze continuation/missingness audit; no original outcome replaced.
output/AMENDMENT-01.txt and resource-review-*.json: original incident and review record.
runs/evaluation* and output/batch-*.json: all complete and partial observed traces and batch records.
output/statistics.json, audit.json, trials.csv, incidents.json: generated only after all assigned attempts are accounted.
output/research-report.html, pdf/finite-plan-agent-study.pdf, paper-source.txt: paper and offline report generated from audited measured data.

OFFLINE DATA AUDIT (NO NEW INFERENCE)
In PowerShell, set PYTHONUTF8=1 and PYTHONDONTWRITEBYTECODE=1; TEMP/TMP to D study runtime. Run existing Python with -B:
C:\Program Files\Python310\python.exe -B D:\MiniProjects_1\02-agent-research\structured-study\analysis\analyze_with_incidents.py
This checks preserved response-derived state, exact prompts, gates, token counts, external grades, hash chains, schedule and frozen inputs. It does not call Ollama. tokenizers 0.22.1 from the hash-verified prebuilt wheel is extracted under runtime/packages on D; numpy comes from the installed Python. Keep v2 supporting modules and control evidence with the original local artifact. Independent replication has not occurred.

FIGURES AND PAPER
make_figures.py uses installed numpy/matplotlib and reads statistics. build_paper.py uses the bundled read-only Python/reportlab runtime and reads the measured statistics/audit. No model calls. PDF pages must be rendered and inspected before delivery; visual inspection is recorded separately.

REPLICATION
Do not run evaluation on this original root: run directories must be newly created, never overwritten/adopted. Use a separately labeled replication root, verify pinned sources/weights and reserve policy, qualify the interface, and retain independent manifests. The current artifact is local, not a publicly released benchmark. Original study results and incident STOP files must stay intact.

SECURITY
No arbitrary generated-code execution, shell or file/network tools are granted to the evaluated model. The inference server is trusted native code and is not AppContainer-isolated. Windows Job resource/cleanup controls do not guarantee access isolation against a compromised server/driver. Persistent master stops fail closed; boundary violations suspend the whole study. Resource incidents are preserved, reviewed explicitly, and never silently retried or counted as measured success.

KNOWN REVIEW LIMITS
One host; concurrent applications; tiny related quantized models; fixed shared templates; five instances/cell; AI-assisted authoring; no independent human review. Missing outcomes are bounded, never fabricated. Public release, application claims and venue submission remain later owner decisions.

POST-FREEZE QUALITY REVIEW
Cross-split/deduplication sensitivities and independent-algorithm recovery audit are explicitly post hoc. Original recovery grader omitted max_steps and visible message said8 instead of12; stricter reference checks every actual outcome before publication. reviewed_apparatus/oracle.py fixes these defects for a NEW future protocol only; it was never substituted into main inference or replay. Frozen recovery partial scores omit length. Tests include a valid admitted16-step counterexample that proves the original omission.

PORTABLE STUDY STOP CONTROL
security/EMERGENCY-STOP.cmd opens the study-specific red button on this computer. security/safety_control.py --stop saves the persistent shared master stop; --status reads state without resetting it. The button stops owned inference processes, not unrelated applications. Python paths in the.cmd are machine-specific; use your own existing Python with -B onD for a separate replication. The older v2 control is retained and unchanged.

HOST I/O AMENDMENT
Original frozen scientific sources remain intact. output/heartbeat-review-*.json, diagnostic-instrumentation-pin.json and mutex-instrumentation-pin.json disclose retained heartbeat interruptions and additional host-only code. A diagnostic recorded read PermissionError; a rejected native reader failed its replacement-under-open-reader control. Qualified private mutex serialization has a100ms acquisition limit, retains the strict five-second age/identity/parser checks and fails closed. Its bounded monitor wait is disclosed; no model capabilities or scientific inputs changed. Rejected candidate tests remain in reviewed_apparatus/rejected and are not presented as passing qualification.

FINAL DELIVERY
All 2880 original assignments are accounted: 2876 completed and four infrastructure-interrupted with unknown full outcomes, never retried. 8549 completed-trial calls; 6155745 measured input+output tokens. All 186 owned inference jobs have cleanup attestation. Model inference has finished.
Start with output/FINAL-REPORT.txt (Russian), output/research-report.html (offline English paper) or output/pdf/finite-plan-agent-study.pdf. Raw records and code are in output/research-evidence.zip, once packaging succeeds. output/archive-check.json records its SHA-256 and CRC; output/portable-replay-check.json, created after extraction and successful replay, records the exact statistics comparison. Those checks do not constitute independent scientific replication.
The compact bundle excludes model weights, native Ollama binaries and private runtime homes/keys. It includes Windows/Python-specific tokenizer files; an installed compatible Python with numpy is required for analysis. package_evidence.py and verify_delivery.py provide reproducibility scripts. Use a NEW D-only destination, retain original records, and never rerun inference into this completed study.
Main result: no observed correctness benefit from the structural completion gate under these models/tasks; conditional uncertainty includes zero. Recovery shortcuts dominate successes, so this study does not establish broad reasoning ability or validation of the website catalogue. Human scientific review remains pending.
