This record separates the author's experience, published cost references, synthetic task assumptions and the successive experimental settings. It also retains the earlier results that motivated changes to the task. The current study is cliff_realistic_dense_v1; earlier series are not pooled with its 504 responses.
Personal motivation and experimental scope
In the second half of 2025, the author used Gemini 2.5 Pro to help prepare a Cambridge application. The author obtained approval for one year of joint training as a master's student, with CSC funding of £1,000 per month. Money pooled for visa financial evidence was subsequently used for the author's grandfather's cancer treatment, and the placement could not be taken up.
This experience supplies the question. The experiment tests payment timing in a separate synthetic scenario and does not test whether a model caused that personal outcome. The historical £1,000 is not presented as a current CSC standard. At the experiment's fixed conversion it would equal RMB 9,000 per month, below the chosen living-cost range. To isolate timing with enough annual support, the experiment instead supplies an independent award sufficient for the year's living costs.
Cost references and synthetic assumptions
Sources were checked on 1 October 2026. The scenario uses a fixed £1 = RMB 9 conversion. The PBOC central parity on 30 September 2026 was 8.8918; 9 is a rounded experimental convention, not a transaction rate.
| Item | Scenario | Role of the source or assumption |
|---|---|---|
| Monthly living L, including rent | RMB 14,500 / 15,500 / 16,500; twelve months | Cambridge's published 2027/28 forecast was £1,725 per month / £20,700 per year, estimated in May 2026. It supplies a cost scale. |
| Monthly rent and deposit D | RMB 7,200 / 8,100 / 9,000; deposit equals one month's rent | Cambridge accommodation information describes deposit and initial-rent payments and the five-week deposit limit. Rent is already included in L. |
| Arrival H | L + D | A complete first living month plus deposit; no additional setup charge. |
| Later living payments | L at H+30, 60, …, 330 | Eleven further payments. Full deposit return at H+360 is a synthetic guarantee. |
| Living award | S = 12(L + δ), with δ = 500 / 750 / 1,000 | A sufficient annual award is an independent task assumption. The original releases are 80% after AR and 20% after the matching DE. |
| Administration V | RMB 16,000 / 17,000 / 18,000 | A £558 Student-route fee plus £1,164 IHS for the specified hypothetical duration total about RMB 15,498; the remainder covers scenario expenses such as translation and medical paperwork. |
| Paid English test | IELTS RMB 1,990 | NEEA's 2026 IELTS/UKVI fee. Approval of an alternative campus certificate costing 0 / 100 / 200 is synthetic. |
| Transport and support | T = G = RMB 8,000 / 10,000 / 12,000 | Fixed return-travel and transfer budgets, not live fares. Advance and settlement refer to one payment; the 50% refund rule is synthetic. |
| Research and application materials | R = 300 / 500 / 700; A = 200 / 400 / 600 | Synthetic printing and preparation costs, not a purchase of academic eligibility or a uniform university fee. |
| Tuition, bench fees and proof | Host covers or waives them; P is free | Host coverage is explicitly supplied. P checks this coverage and the award; it imposes no additional balance hold. |
Primary source links: PBOC exchange-rate table, Cambridge maintenance costs, Cambridge private-rental payments, deposit information, NEEA 2026 test fees, Student visa, IHS amounts and duration.
Cambridge visiting-student information and registration categories and fees motivate specifying status and coverage, rather than reconstructing the author's historical process. The latter page's CSC example refers to the Commonwealth Scholarship Council, not the China Scholarship Council. The China CSC 2026 programme guide mentions return international travel and scholarship support; this study does not use it to infer a current uniform UK monthly rate.
Where the payment mechanisms came from
The UCL Turing Scheme describes, for 2025/26, an 80% release after agreement and arrival evidence and a final 20% after departure evidence. Its target of fourteen working days is not a guaranteed receipt date.
The Manchester Turing Scheme describes travel advances, receipts and repayment of excess support for 2026/27 summer or short-term activities; that provision excludes semester and full-year placements.
These are distinct examples. Their combination here is not a real Cambridge/CSC annual policy. The task independently specifies sufficient annual support, guaranteed receipts fourteen calendar days after AR and seven after DE, an early or late AR appointment, and full deposit return. The six Chinese notices explicitly state all of those conditions.
Fixed main design
The main configuration draws twelve complete financial profiles once using random.Random(20261005) and the recorded field order, including singleton domains. No draw is rejected, filtered or repeated. Initial cash takes the shared values 20,000; 25,000; 30,000; 35,000; 37,500; 40,000; 42,500; 45,000; 47,500; 50,000; 52,500; 55,000; 57,500; 60,000; 62,500; 65,000; 67,500; 70,000; 80,000; 100,000; and 120,000 RMB. Model seeds are 1729 and 2718.
All outputs and all budgets are reported. Rates summarise this nonuniform grid and are not directly comparable to pooled rates over the earlier seven cash roles. Descriptive percentile intervals resample twelve whole profiles 5,000 times, retaining every budget and seed within a profile; bootstrap seed 20261006. The curves connect saved points without a fitted sigmoid, a fitted B50, enforced monotonicity or removal of contrary results.
The checkpoint is Qwen3.5-4B revision 851bf6e…, with native thinking, BF16, context 131,072, output cap 32,768 and batch 12. The complete sampler and runtime are in methods. The current experiment does not run 2B.
For the cheapest permitted campus-certificate, travel-advance and early-AR route, write q = R + campus-certificate cost + A + V, where R here is the research-material cost. Then R_H = q + D + L and R* = q + D + 2.4L − 9.6δ. Travel support arrives before V and offsets T. Future living support cannot fund arrival before its release. The exhaustive search, rather than a model judgment, supplies each profile's reference values.
Development record
| Series | Observed result and subsequent development |
|---|---|
main_v1 |
4B completed 320/328 reference-feasible requests; there were zero cash-only failures among 320 valid executable schedules. The original 2B outputs, fixed diagnostics and schema clarification remain. This task did not support the initial proposed cash-failure pattern. |
cliff_dev_r1 |
Twelve responses: eight positive, four negative, five completions and zero cash false assurances. Higher-cash responses also had procedural or date errors. |
cliff_dev_r1_lang |
Eight fixed requests clarified synonymous administrative terms and receipt semantics. Seven executed. A higher-cash plan still omitted the invitation, so the failure was not purely financial. |
cliff_dev_r2 |
Introduced explicit conditional 80/20 support and advance versus reimbursement. All twelve answers were positive and structurally valid: eight reference-feasible plans completed; four at R*−500 failed on cash. Ambiguities about payment direction and protected living costs were retained in the record. |
cliff_dev_r3 |
Clarified that the student pays providers and that guaranteed ordinary living applies only before departure. Of twelve responses, ten were positive. Four reference-feasible plans completed; six of eight physically infeasible requests produced unconditional cash false assurances. |
cliff_dev_r4 |
Two new profiles, twelve responses. All eight physically infeasible requests produced unconditional cash false assurances. Of four reference-feasible requests, three completed; one negative answer selected late AR and rejected its own inferior schedule. |
cliff_confirm_v1 |
Twelve new profiles × seven cash roles × two seeds = 168 responses. There were 166 readable Booleans, 151 unconditional positives and 155 structurally valid plans. Execution was 90/96 among reference-feasible requests. Fifty-seven cash false assurances comprised 56 physical and one avoidable failure; four additional positives failed structurally and two had interface problems. |
cliff_realistic_dev_r1 |
Two new profiles drawn with seed20261004, six cash levels of 20/40/50/60/70/100 thousand RMB and model seed1729: twelve responses. Eight unconditional positives, eight structurally valid plans, six completions, two physical cash false assurances, one positive structural error and one own-plan false negative. All three arrival-unaffordable requests were correctly rejected. |
cliff_realistic_dense_v1 |
The current 504-response experiment. Its complete results are in methods; the original requests, outputs and scores remain available. |
In the realistic pilot, both cash false assurances were profile 02 at RMB 40,000 and 50,000. Their valid optimal plan required RMB 53,900. On day 688, a RMB 15,500 bill met RMB 1,600 or 11,600: shortfalls of 13,900 and 3,900. Profile 01 / B50,000 used G_adv rather than the specified action ID; its original structural error was not silently repaired and counted as a cash failure. Profile 02 / B60,000 denied its own executable plan. Incomplete negative plans were not relabelled as late cash failures.
The realistic-cost version is not a uniform rescaling of the preceding experiment. It changes the cost domains, the full first-month payment at H, twelve award standards, travel support equal to transport, removal of an additional 28-day hold, a shared absolute-cash grid and batch size 12. The earlier conditions and outputs remain. This is disclosed adaptive development, not an unchanged replication, a claim of preregistration without seeing development results, or an isolated causal estimate of monetary scale.
The initially proposed main grid was RMB 20,000–120,000 in increments of 5,000. After the complete twelve-response pilot was read, it was revised once to put 2,500 increments between 35,000 and 70,000. The total remained twenty-one cash values and 504 requests. Profile-draw seed, model seeds, costs, prompts and sampler were unchanged, and no new main profiles or outputs existed at the time. The original grid snapshot, pilot answers and scores are retained.
Interpretation of the final answers
Six GPT-6.1-sol MAX agents each read the complete final answers and JSON reasons for two profiles: 84 responses each, with no overlap. Their merged annotations are in the saved annotation file. The own_plan_commitment categories are unconditional, withdrawn, conditional_external_funds, negative, contradictory_negative and ambiguous; each note cites the answer's text.
These are auxiliary model interpretations, not human gold labels or an inter-annotator agreement study. The original Boolean, structural scores, cash scores and model text are unchanged. Missing interpretations are not assumed unconditional. Permitted future receipts within the task are not treated as outside funding. Visible thinking and explanations do not establish the hidden causal mechanism of a response.
There are 311 readable positive Booleans, all classified as unconditional. The semantic category contains 313 unconditional texts because two additional texts lack a readable Boolean; those two are interface cases and are not added to the 311. Other text categories include 189 negative and two ambiguous answers. Cash false assurance requires a readable positive, a full structurally valid schedule and failed cash execution.
Illustration selection and timing replay
Before the main profile draw, the remote calculation established exact bounds across the fixed cost domains: R_H = RMB 38,200–45,000; original R* = RMB 48,900–63,300. The full-year requirement exceeds arrival by at least RMB 10,700. These calculations generated no model responses and did not select profiles. The relative ordering of the advance, V and early-AR receipts is preserved by calendar shifts.
Figure 3 fixes the first main profile and its original optimal reference schedule. Its low illustration amount was changed from RMB 40,000 to 45,000 before the draw, because 45,000 permits arrival throughout the cost domain but cannot fund the original full year. RMB 70,000 is sufficient throughout. The calculation is saved in the domain-bound record.
The timing diagnostic moves the final 20% to the original AR receipt day, retaining total support, costs and the original action dates. DE remains required and earns no duplicate instalment. All original structurally valid schedules are replayed, with no new model calls. Of 97 cash false assurances, 95 disappear and two remain. Among 348 valid schedules, completions rise from 194 to 305. These are changes in environmental payment timing, not evidence of learned model improvement.
The seed-1729 answer shown for profile 01 / B45,000 was inspected after generation. Its day-702 shortfall is RMB 10,800. The profile and cash amount were fixed beforehand; the actual answer and the reference trajectory are documented separately. The two residual false assurances are profiles 08 and 12 / B42,500 / seed2718, whose bills occur before their first receipts. All residual failures, correct warnings and high-cash exceptions are included in the complete results.
Research context, files and reuse
DeepPlanning studies global time, space, cumulative cost and total-budget constraints. SCHEDBench studies natural-language resource-constrained scheduling and constraint fidelity under equivalent wording. This evaluation focuses on when a conditional entitlement enters the student's spendable balance and whether that balance supports the model's full-plan commitment.
Inference, reference search, scoring, statistics, replay and original plots ran on Molab RTX PRO 6000. Publication figures were later redrawn locally from saved results; the report was typeset and checked locally. Full final answers and thinking, runtime records, original scores and earlier development records are retained. The data guide provides an English map of the original Chinese experiment and the reading copies prepared for this release.
Original synthetic data are CC BY 4.0; original code is MIT. Qwen weights and third-party sources retain their own terms. Model weights are linked to the official checkpoint rather than duplicated. The personal account is separate from the fictional benchmark profiles.