SEQUENTIAL PROMPT EXPERIMENT · ENGLISH ACCOUNT

What changed after the first request?

The initial request was: “Simulate SO₂ emissions and their spatial distribution in Hangzhou over the next 15 days.” A separate fresh task received only that sentence. After the first result was frozen, the controller supplied fifteen corrections one at a time, freezing each actual result before the next. The controller had seen the correction list; it did not claim to erase its own context. Initial-task isolation does not erase generic model knowledge.

Substantive changes

Source evidence. Early hypothetical sector sources were replaced with four published EIA design outlets, with outlet coordinates, emission rates, dimensions and temperatures. The conditional total is 4.519 kg/h, or 1,626.84 kg over 360 hours. Actual future operation, the gas-flow reference basis and geodetic datum have not been independently verified. Four documented outlets are not a complete Hangzhou emissions inventory.

Atmospheric transport. The early two-dimensional relative field evolved into a passive three-dimensional stochastic particle model with height-dependent forecast wind, parameterized turbulent diffusion, a bulk stability diagnostic and simplified buoyant rise. The real GFS forecast column is repeated horizontally. There is no resolved terrain-induced wind, ambient vertical flow or building downwash. This is a custom model, not WRF, FLEXPART, CFD or LES.

Continuity and accounting. Later versions explicitly verify state propagation through all 360 hours without daily reset. Airborne mass, deposition, chemical removal and export are tracked separately. The terrain model adds reversible and irreversible surface stores; re-emission recycles internal mass rather than adding an external source. A closed ledger establishes numerical accounting, not concentration accuracy.

Background and sampling. An inadequately validated surface-only background candidate was excluded. Results therefore show source increments with zero initial and incoming background, not total real ambient SO₂. “Ground” concentration is a defined 0–25 m AGL volume mean. It is not a 2 m measurement. Units, coordinates, source release heights and particle occupancy were checked explicitly.

Numerical refinement. Horizontal density bins were refined to 500 m, particle counts studied up to 353.89 million per realization, vertical bins increased to 52 layers and timestep reduced to 30 s. These are distinct studies with their own counts and controls. Several local and full-field convergence criteria remain unmet. Particle-count and bin refinement do not create finer meteorological information.

Terrain and surface exchange. Actual GLO90 DSM data provide bilinear terrain collision and terrain-normal reflection. Surface contact uptake, retention, fixation and conditional return are represented with uncalibrated coefficients. The forecast return gate stayed closed for all 360 hours. A warm/dry surface-only override demonstrated the mechanism; it is not a claim that real forecast weather produced re-emission.

Evidence that was not promoted to a success

Early flat-ground CPU/GPU agreement does not qualify the changed terrain model. Its strict long-run GPU comparison failed. A separate bounded localization found tiny initial floating-point differences amplified along state-dependent trajectories, without demonstrating a semantic implementation bug. No tolerance was relaxed and no “repaired GPU” was promoted. The original 345,600-particle terrain case also cannot inherit high-count convergence evidence from another physical scheme.

The requested 2 mm rainfall allowance is not an independent forecast-error result here: the preserved rainfall comparison uses the same supplied forecast on both backends and therefore gives zero implementation difference. It does not validate rainfall against observations or an independent weather solver.

Version-by-version record

Counts below are primary emitted numerical particles per realization, not extra emission mass. Horizontal sizes are concentration-estimator bins, not GFS weather resolution. Empty entries mean that this definition did not apply in the early two-dimensional model.

Version	Correction and actual result	Density grid / layers / timestep	Primary particles	Retained limitations
00	Initial prompt 2-D relative scenario	—	—	Relative index; hypothetical sector sources; repeated-weather spinup; not a concentration forecast.
01	Forecast provenance 2-D relative scenario	—	—	Frozen complete GFS run; single surface column; hypothetical sources retained.
02	Continuous 360-hour integration 2-D relative scenario	—	—	Continuous state and day-boundary hashes; zero initial load; transport is not strictly conservative.
03	Documented outlet scenario 2-D conditional column burden	—	—	Four EIA design outlets; commissioning and operating schedules unverified. Legacy citywide maps are not these outlet results.
04	3-D transport and plume rise 3-D conditional particle scenario	2000 m / 13 layers / 120 s	86400	Horizontally uniform forecast profiles, parameterized turbulent diffusion and buoyant rise; flow basis unresolved; no observed-concentration or convergence validation.
05	Deposition and mass tracking Inspect actual stage outputs	2000 m / 13 layers / 120 s	86400	Separate dry, wet, chemical, atmospheric and outflow quantities only where the completed stage actually provides them.
06	Background assessment Source-only3-D scenario	2000 m / 13 layers / 120 s	86400	Unvalidated surface-only CAMS background excluded; zero initial and external background is an attribution assumption.
07	Near-surface sampling Sampling qualification failed	2000 m / 13 layers / 120 s	86400	0-25m AGL volume means. Four outlets share a2km cell; seed and doubled-count tests expose severe local sampling noise.
08	GPU implementation Same-physics backend agreement	2000 m / 13 layers / 120 s	86400	Shared indexed RNG is new on both CPU/GPU. Compute speed excludes forcing, compilation and disk output; no concentration-skill claim.
09	Release implementation gates Implementation agreement only	2000 m / 13 layers / 120 s	86400	Exact-zero controls, nonnegativity and budget checks pass. Zero precipitation RMSE compares the same broadcast input; it is not an independent weather-solver validation.
10	Physical outlet release height Runtime birth-position verification	2000 m / 13 layers / 120 s	86400	Both solvers release at20/50/80/65m AGL before transport/rise. EIA-documented heights are not independently surveyed as-built heights.
11	Horizontal estimator refinement Local grid convergence not demonstrated	500 m / 13 layers / 120 s	5529600	2 km, 1 km and 500 m density bins use shared particle paths; the meteorological input resolution is unchanged. Fine-bin receptor differences remain above the declared criterion.
12	Particle-count refinement Sampling noise reduced; criteria not all met	500 m / 13 layers / 120 s	353894400	5.53 to 353.89 million particles per realization at fixed physical emission mass. Two-seed daily-field differences fall to 8.42%, but successive-count field and receptor criteria still fail.
13	Vertical estimator refinement Native profiles not converged	500 m / 52 layers / 120 s	353894400	13, 26 and 52 layers with shared trajectories. The lowest layer is 6.25 m; 26-to-52-layer global and outlet profile changes exceed the declared limits. Fixed 0–25 m agreement is a common-volume consistency check.
14	Transport timestep refinement Partial criteria pass; full-field criterion fails	500 m / 52 layers / 30 s	88473600	120, 60 and 30 s with paired stochastic increments and fixed 120 s release events. Process totals and receptors meet their finest-step subcriteria; near-surface fields still differ by 13.6%.
15	Terrain and surface exchange CPU terrain scenario; new GPU agreement failed	500 m / 52 layers / 30 s	345600	Actual bilinear DSM boundary, contact uptake and reversible/retained stores. Forecast re-release is zero under the assumed wetness gate. The CPU case has only 345,600 primary particles; previous higher-count convergence does not qualify it. Surface chemistry is uncalibrated; wind is not terrain-resolved.

Presentation and timing

Vancouver-style three-dimensional geographic rendering and separate Scientific analysis are reused on an independent presentation track. No Vancouver numerical field or city model becomes Hangzhou input. Early two-dimensional results remain two-dimensional. Preserved original versions use daily means. The completed high-count case additionally saves genuine hourly native fields: queries read all 360 hourly, 500 m, 52-layer float64 arrays. Its regional graphics use conservative 2 km averages every three hours, with scalar interpolation; display resolution is not native numerical resolution. Grey volume opacity encodes concentration illustratively: SO₂ gas itself is invisible.

Each version and stepwise comparison is planned as a 15-second presentation, after reference-frame review. Every frame is rendered from interpolated data or geometry, with no image crossfade, temporal RGB blending or overlaid neighboring frames. Near-outlet graphics integrate residence along actual accepted paths. This systematic diagnostic selects one member per source/release event through hour 336 and eight members during the final 24 hours, verified against the original subset and saved final states. Display expansion factors are separate from actual sampled mass; this is not the full-count near-outlet concentration or an unbiased random sample. Thirty-second path segmentation can remain visible. The low-to-high-count full-domain daily-field relative L1 difference averages 139.8%; this comparison does not establish local convergence.

Time is reported as observable Agent activity, data/model work, dedicated rendering, and other or inseparable work. Agent activity is a non-tool wall-clock proxy, not an internal inference timer. Overlapping intervals are unioned, and worker versus controller activity is distinguished. Read the current measured timing snapshot.