# Coherent implementation model

Coherent is an educational incremental economy, not a quantum-device emulator or a hardware forecast. All costs, laboratory seconds, staffing, qualification cutoffs, and numerical scenario coefficients below are game choices. Papers motivate the concepts; they do not validate the game's prices or performance.

**Reviewed campaign-depth revision, 4 October 2026.** This records the implemented engine and content coefficients accepted by the independent AI quantum and game-design reviewers. Legal strategy simulations and browser observations are recorded in `DEPTH-VERIFICATION.md`; they do not establish a 90-minute first human play or human enjoyment.

**Research-history extension, 5 October 2026.** The eleven annual discoveries add paid studies and purchases without changing the existing physical-error or full-workload coefficients. [RESEARCH-HISTORY-DESIGN.md](RESEARCH-HISTORY-DESIGN.md) records their intent; [RESEARCH-HISTORY-SCIENCE-REVIEW.md](RESEARCH-HISTORY-SCIENCE-REVIEW.md) records independent AI source/model acceptance, including the distinction between paper findings and authored study criteria. Older review dates below retain their historical scope.

## Architecture and campaign

The implementation uses static HTML, CSS, and classic JavaScript. A content table, a DOM-free deterministic engine, and a browser interface are separate files. There is no framework, backend or runtime network API. The 2D game requires no build; the separate 3D entry uses checked-in local Three.js assets generated by `npm run build:3d` during development. Both play offline from a local file or static HTTP server. Node's built-in test runner exercises the same engine used by the browser.

Six chapters implement the 30 foundation discoveries in `QUANTUM-DESIGN.md`: first signal, control, noisy circuits, correction, logical operations, and useful work. The campus adds three recent-research discoveries; the research-history extension adds eleven annual 2015–2025 discoveries, for 44 total and 51 academic sources. The full archive retains the foundational timeline from 1982 through 2026. Two of the recent-research discoveries become prerequisites for annual studies in the expanded programme. The content also defines 14 classical engineering advances, including two mutually exclusive choice pairs. These are laboratory-management game rules, not additional paper findings. Research effort is explicitly abstract staff work; repeated trivial measurements do not manufacture new scientific evidence. Qualifications and demonstration grants are awarded once per named experiment. Assignments and operating allocations are reversible; purchased equipment remains installed and the two engineering forks cannot be exchanged during a run.

## Research, storage, and engineering designs

The laboratory starts with one researcher, one notebook assignment, 60 funding, and no stored research or designs. Let `r` be reputation and `n` the number of discovered papers. Available trust is `2 + 2r + floor(n/4) + openBonus`, where open dissemination adds four. Each researcher and notebook assignment uses one trust slot. The opening two assignments therefore fit the opening two slots. Researchers can be assigned or released from 0 to 32; notebook assignments range from 1 to 64. Releasing a notebook is refused if its lower capacity would discard stored research. Assignments unlock with the first discovery and cost no purchase funding.

Research storage is `effortCap = 120 * notebooks * 4^storageTier`, with one tier for each of the three storage advances. Each advance multiplies capacity by four. Research accumulates only up to that capacity, and a discovery whose research cost exceeds it remains visibly blocked until storage expands. Paper and engineering purchases spend funding, research, and designs according to the price tables in `content.js`.

Let `a` be the number of classical automation stations currently running. Abstract research production is `(1.2 * staff + 3 * a) * researchMultiplier` per laboratory second. Workflow and scheduling each double `researchMultiplier`, for a maximum factor of four. After the universal-idea discovery, engineering-design production is `(.04 * staff^1.2 + .3 * a) * bankBonus * workflowDesignMultiplier * schedulerDesignMultiplier * synthesisMultiplier`. A full research store gives `bankBonus = 4`; otherwise it is one. Workflow improves design output by 25%, scheduling by 50%, and synthesis doubles it. Synthesis can be funded after memory qualification and classical scheduling, before logical factories; it remains a competing investment. Stored designs are capped at one billion. Designs represent classical engineering proposals, not verified scientific results or quantum states.

Spending research gives up the full-store bonus until the bank fills again. Expanding storage also moves the bank away from full. Allocating trust to notebooks makes large discoveries affordable but removes slots from researchers. These coupled choices adapt the original game's insight, notebook, credibility, and idea mechanism without its irreversible staffing trap.

Each first named qualification increases reputation by one, awards `80 + 120 * r` funding using the new reputation value, and adds eight research units up to current storage capacity. Repeating that qualification awards no further reputation, funding, or research. Each first completed workload gives its declared payout, two reputation, and 100 research units up to capacity. Recurring grants are only `1 + .12 * r` funding per laboratory second; they do not grow quadratically with chapter. Baseline upkeep is `.06 * staff + .1 * moduleLevel + .1 * rackLevel`, with additional operating costs below. The small stipend supports a minimal laboratory; large purchases require finite demonstration awards or delivered service revenue.

## Apparatus duty, maintenance, and service

Calibration duty `c` ranges from zero to 0.6. Service allocation `s` ranges from zero to 0.9 of the time left after calibration. In ordinary operation, effective service duty is `(1-c) * s`, while experiment progress uses `(1-c) * (1-s)`. Staff research continues independently. A memory, gate, or factory qualification experiment, or a full workload, is an **atomic protected schedule**: service, capacity-consuming automation, and construction pause; the job advances with duty `1-c`. The configured service and construction settings resume after that schedule finishes or is cancelled. This reservation changes laboratory pacing, not the separately modeled quantum microsecond schedule.

The maintenance target is the selected game rule `min(.55, (.08 + .018 * log2(active+1) + .04 * service) * rolloutMultiplier)`. During an atomic protected schedule the service term is zero. Verified rollout gives `rolloutMultiplier = .8`, rapid rollout gives `1.2`, and otherwise it is one. Maintenance grows with the active, commissioned footprint to represent the bookkeeping burden of this fictional controller scenario; it is not a measured law of device noise. Drift changes by `.004 * (maintenance - calibrationDuty)` per laboratory second and is clamped from .005 to one. Manual calibration reserves the entire apparatus for four laboratory seconds, pauses customers, automation, and commissioning, and resets drift to .005; the Ramsey scan resets it to .03. Researching randomized benchmarking sets the initial manual duty to .25. The later automatic-calibration advance permits `c = min(.6, maintenance+.02)`, explicitly retaining the apparatus cost of that reserve.

Service is a stipulated contract for batches of **1,024 known Bell-state preparations with computational-basis readout**. A batch is an economic unit distinct from the coupled-circuit tutorial's actual 2,048 browser samples. Computational-basis correlations alone do not certify entanglement, device fidelity, arbitrary-state reconstruction, or quantum advantage. The contract is available after the noisy-era discovery and the circuit demonstration. Its continuing cutoffs are the game conditions `pEff <= .004` and `drift <= .25`; these are not paper-derived accuracy claims or an experimentally certified success bound.

The selected controller/readout lane count is `min(1 + 2 * rackLevel, floor(active/2))`. Each lane contributes 1.5 batch-equivalents per laboratory second at full service duty, so shared capacity is `lanes * 1.5 * (1-c) * s`, or zero during an atomic schedule. These are superconducting controller presets for this narrow known-preparation service, not a universal throughput multiplier from qubit count. Classical automation takes its capacity share first, leaving `customerCapacity = max(0, sharedCapacity-a)`.

Fair price is `3 * (1+.2*r) * disseminationPriceMultiplier`, with multipliers .7 for open dissemination, 1.5 for proprietary dissemination, and one otherwise. Demand at fair price is `min(60, (1+.4*r) * disseminationDemandMultiplier)`, with multiplier two for open dissemination and one otherwise. At price `p`, demand is `min(60, demandAtFair * (fairPrice/p)^1.4)`. Delivered batches are the lesser of customer capacity and demand if current service qualification passes, otherwise zero. Revenue is delivered batches times price, so unmet demand earns nothing.

Manual price ranges from .5 to 200 funding per batch. The pricing advance permits capacity-matching automation: `p = clamp(fairPrice * (demandAtFair/customerCapacity)^(1/1.4), .5, 200)` when spare capacity is positive, otherwise 200. This is a declared demand curve and matching rule, not a real auction or proof of an optimal market price. Open versus proprietary dissemination is an exclusive game choice between more trust/demand and a higher price coefficient.

## Classical automation and construction

After the automation advance, up to 16 analysis stations can be purchased. A reversible `analysisShare` from zero to one controls requested station duty. Each consumes one controller-batch-equivalent slot while running, with `a = min(purchasedStations * analysisShare, sharedCapacity)`. Fractional running capacity denotes a duty share, not a fractional station. Station operation costs `.1 * a` funding per laboratory second in addition to staff and equipment upkeep. The stations perform classical analysis and organization; their abstract research/design output neither derives from Bell-state data nor claims a quantum acceleration of AI or general computation. Giving them priority over customers creates an income versus research tradeoff.

After the workshop advance, up to 12 construction teams can be purchased. Their hypothetical commissioning rate is `workshops * 1.2 * constructionMultiplier`, with multipliers .8 for verified rollout, 1.4 for rapid rollout, and one otherwise; it is zero during an atomic protected schedule. A slider `f` from zero to one divides this rate between chip fabrication and control/cooling integration. The funding-limited duty is `min(1, max(0,funds)/(workshopRate*.8))`; each completed construction unit costs .8 funding through the upkeep calculation. Fabrication rate is `workshopRate * f * duty` while installed capacity is below its cap; integration rate is `workshopRate * (1-f) * duty` while supported capacity is below its cap. Unused allocation at a completed cap is not automatically reassigned.

Fractional `fabricated` and `integrated` values store commissioning progress, each bounded at 8,192. Only their floored integer counts contribute physical capacity: `installed = min(8193, pool[moduleLevel] + floor(fabricated))` and `supported = min(8193, pool[rackLevel] + floor(integrated))`. Active capacity is their minimum. Partial construction does not supply fractional physical qubits or logical slots. Fabrication and integration here are hypothetical economic construction abstractions, not forecasts of wafer yield or refrigerator engineering.

All commissioned qubits use the same selected uniform noise/control scenario. Expansion does not predict real fabrication yield or experimentally validate a new device; current readiness still depends on the declared model and its qualification conditions.

Construction teams are distinct from the selected logical **magic-state factories**. Construction expands commissioned physical footprint; logical factories reserve footprint and have a separately qualified fresh-state resource scenario. Persistent factory rehearsal credits remain abstract schedule practice, accrue at `factories * 8/d` per laboratory second only while the current factory qualifies, and are capped at 2,000. Their pacing rate is separate from physical fresh-state throughput; credits do not store quantum states or create additional hardware.

## Measurements

Opening experiments sample explicitly prepared states and educational measurement models. A histogram records sample counts, not unknown amplitudes. Ramsey T2-star and echo T2 remain distinct. Displayed RB error is a separate illustrative characterization estimate, never a substitution for the threshold model's effective stochastic parameter.

The two-spin tutorial uses `H = Z0 Z1 + 0.6 X0 + 0.6 X1`, and the normalized known ansatz `cos(theta)|Phi+> - sin(theta)|Psi+>`. Its exact expectation is `cos(2theta) - 1.2 sin(2theta)`; the classical ground-state reference is `-sqrt(2.44)`. Each Pauli term is sampled separately. The browser really computes these four-dimensional expectations, sample counts, and the classical reference. This is a classically simulated tutorial, not a quantum hardware experiment or advantage demonstration.

For three independent measurement groups of `M` shots each, a simultaneous 95% Hoeffding energy bound is `2.2 sqrt(2 ln(6/0.05)/M)`. Ordinary standard error is also reported separately. Residual measurement bias is budgeted separately from statistical uncertainty and ansatz error. Mitigation models four times the raw-shot overhead and reduces the declared residual bias. The browser actually samples 3M Bernoulli trials; modeled acquisition is 12M with mitigation and is labeled separately. Funding costs `35 + ceil(modeledRawShots/2048)`, so both greater precision and mitigation consume a real game resource. This is a selected estimator/overhead scenario; no universal free-precision bonus is claimed.

## Hardware and noise

Purchased chip and control/cooling rack presets are `[1, 9, 25, 81, 257, 769, 2049]`. Integer commissioned construction adds to each preset as described above, up to 8,193 installed or supported qubits. Active physical capacity is the smaller of installed and supported capacity. These are illustrative game presets.

The selected independent stochastic threshold parameter is `pEff = 0.012 * 0.5^pulseLevel * readoutFactor * echoFactor * (1 + drift)`, where pulse levels run from 0 to 8, readoutFactor is 0.85 after its discovery, and echoFactor is 0.9 after its discovery. Factors are explicitly game scenario choices. A separate RB estimate uses a separate formula. Calibration duty competes for apparatus runtime while reducing drift. Calibration does not create additional physical qubits.

The no-fault proxy uses 12 independent stochastic gate events with separately declared probability `pCircuit = 0.01 * 0.55^pulseLevel * readoutFactor * (1+0.5 drift)`; here readoutFactor is 0.8 after readout research. This is an educational IID fault channel, independent of the RB display and threshold scenario. The proxy is not algorithm correctness. Coherence is not charged again to the same gate-error term. Ramsey values are selected scenarios; memory-bin counts are illustrative Bernoulli detection events, not a decoded circuit-level surface-code simulation.

## Memory, gates, and allocation

For odd distance 3, 5, 7, or 9, an ideal rotated memory patch uses `2 d^2 - 1` physical qubits. Below the selected threshold 0.01, use the dated toy fit `pL = max(1e-8, 0.1 (100 pEff)^((d+1)/2))`. The residual floor is an illustrative assumption. At or above threshold, reliable scaling is unqualified. The coefficients are inspired by [Litinski's selected resource model](https://arxiv.org/abs/1808.02892), not a universal hardware threshold or a fit to our RB display.

Protected memory, a historical discovery, and a currently qualified logical gate stack are distinct. Current drift, distance, footprint, and decoding capacity are checked whenever qualification or workload eligibility is requested. An unlocked paper never permanently certifies degraded hardware.

Logical operations reserve two standalone-patch-sized allocation units for routing and one for spares. Each factory reserves four additional units. Application memory slots are the remaining complete patches. This is a **fictional educational allocation**, not an exact real factory or Litinski tile layout. All allocated complete patches, including routing, factories, and spares, contribute syndrome load `(d^2 - 1)` measurements per selected 1-microsecond cycle. Decoder capacity uses measurements per microsecond; feedback latency uses microseconds and is checked separately.

Logical-operation error is a separately qualified scenario contribution `2 pL` per operation. It charges operation faults outside the idle-memory accounting and excludes magic-state preparation. Factory output error `pT` is a stipulated total accepted-output error, including internal factory faults: 1e-3 before distillation and 1e-7 after. Neither value is derived from a real distillation protocol.

## Factories and time

Persistent factory progress is **rehearsal credits**, a scheduling/economy abstraction. Credits are not stored magic states and carry no quantum-state quality. They cannot establish hardware readiness. Physical resource scenarios use freshly produced states with current qualified output error and footprint. A discovery never retroactively purifies quantum stock because no quantum stock is simulated.

Selected factory throughput is `factoryCount / (8 d)` accepted states per microsecond. This timing is a fictional scenario choice, not Litinski's actual 11-tile protocol. Credits accrue at a separate gameplay rate in laboratory seconds. Game project durations are pacing choices, distinct from the modeled quantum wall time.

Each resource recipe declares application width, gate count, gate depth, operation precision, magic-state count, repetitions, preparation/readout/classical times, and a total modeled time ceiling. An operation lasts `4 d` microseconds. Overlapping gate execution, fresh state production, and critical feedback take their maximum. Sequential preparation, readout, and classical work are added, and the result is multiplied by repetitions.

Idle memory exposure per repetition is `max(0, width * gateDepth - gateCount) * 4 d` qubit-cycles plus `width * max(0, parallelTime - gateTime)` qubit-cycles, using the selected 1-microsecond cycle. The recipe treats gate counts as single-application-qubit occupancy slots in its educational schedule; routing and multiqubit-operation overhead live in the separately qualified operation model and reserved layout. Additional factory or feedback waits therefore increase memory risk. This is a declared recipe convention, not a compiled circuit resource estimate.

Per-execution modeled risk is the sum of idle qubit-cycles times memory error, operation count times operation error, fresh state count times accepted-output error, preparation, and other terms. The entire repeated task uses `B_model = min(1, repetitions * perExecutionRisk)`. All inputs are point estimates/scenario assumptions; this is **not an experimentally certified upper bound**. Time, fresh states, and pacing credits are also multiplied by repetitions.

## Workloads and validation

The main ending's future recipe is a 32-site periodic nonintegrable Ising model: `H = sum Z_i Z_(i+1) + 0.7 sum X_i + 0.3 sum Z_i`, starting from all-zero computational-basis spins, evolution time 1 in units with coupling and hbar equal to 1, and target observable mean magnetization `sum <Z_i>/32`, precision 0.02. The game recipe selects 4,800 operations, depth 200, 192 freshly produced states, and one execution. These counts and its precision assignment are **educational recipe assumptions**, not a compiled algorithm or paper-derived estimate. The browser does not compute this large result. Finishing it displays **Modeled scenario completion** and no invented numerical magnetization.

Small factoring and search tutorials validate classical certificates in the browser. Their logical execution economics remain modeled; the browser does not pretend to execute Shor or Grover on a quantum device. The electronic-energy resource scenario uses a six-site periodic Hubbard ring, nearest-neighbor hopping t=1, on-site U=4, six electrons with balanced spin, and 12 spin orbitals; its target is ground-state energy per site within 0.01. Its resource recipe and precision assignment are educational assumptions, and no energy or catalyst is computed. No real-world speedup is claimed. The speedup-audit discovery explains why a matched task, accuracy, success probability, and dated classical baseline would be required.

## Persistence and time policy

Progress advances only while this game is visible and unpaused. Hidden and closed tabs receive no offline progress. Jobs retain their progress when saved. The engine's depth-campaign save envelope is `{game:'coherent', version:2, state:...}`, with `state.version` also two. Validation requires bounded values, known discoveries/qualifications/workloads, valid engineering prerequisites, exclusive choices, earned trust assignments, notebook capacity, and legal job/configuration data. Reputation must equal the number of named qualifications plus twice the number of completed workloads. Result samples and classical certificates retain their existing validation.

Version-one short-campaign saves are rejected with an explicit baseline explanation; they are not migrated into the new economy. The browser uses a separate `coherent.v2` storage key and reads any `coherent.v1` value separately for a baseline notice and raw JSON export. Saving, importing, and resetting the deeper laboratory leave that earlier key untouched. Browser checks verified exact preservation and export of a version-one storage sentinel, independent version-two storage, and current-state preservation after a rejected baseline import. The sentinel checks storage behavior, not restoration of a baseline campaign. An unreadable depth save remains protected until an explicit valid import or confirmed reset replaces it. Storage failures remain visible and export remains available. Reset requires an explicit confirmation. Sound is off until the player enables it; volume and mute are saved.

## Verification

Engine checks cover qualification, accounting, affordability, reversible trust assignments, storage pressure, full-bank opportunity costs, shared duty, pricing and delivered revenue, classical automation competition, atomic construction pauses, malformed saves, and whole-campaign completion with multiple legal strategies. `DEPTH-VERIFICATION.md` distinguishes current browser observations from retained baseline evidence and records the independent AI scientific, game-design, and code-review conclusions. Legal simulated completion is distinct from human pacing and enjoyment. Long action gaps and the small share of explicit experiment jobs remain disclosed limitations.

## Research campus extension — 5 October 2026

The rules below extend the reviewed baseline above. Their implementation review and native browser evidence are recorded in `CAMPUS-VERIFICATION.md`; design agreement is recorded in `QUANTUM-FRONTIER-REVIEW.md`. Baseline verification does not establish acceptance of this extension.

**Fresh measurement work.** Six finite frontier goals and six finite precision requests use the same exact classical two-spin model. A job captures its task ID, unique trial ID, angle, samples per group, mitigation, residual bias and recipe revision. Campus VQE duration is `8 + modeledAcquisitions/12288` apparatus laboratory seconds. This authored pacing rate is distinct from quantum microseconds. More shots and mitigation consume funding and time; three groups still produce `3M` actual browser samples, while mitigation charges `12M` modeled acquisitions.

A ground-reference task checks `abs(estimate-ground)+samplingBound+biasBound <= tolerance` and separately refuses an exact ansatz error above tolerance. The ansatz guard is not another additive charge. A named-angle task requires the actual recorded angle within one degree, then compares against its exact recorded-angle energy. Residual-bias ceilings are separate where declared. Bounds apply to each fixed trial; adaptive tuning and repeated attempts do not establish a campaign-wide 95% guarantee. A named-angle result can pass its request while failing the ordinary ground-reference tutorial. The result, message and receipt identify the actual task criterion.

Successful goals award their stated funding/designs once and add one trust assignment each. The baseline trust equation therefore gains `+ objectiveResults.length`. They do not increase reputation or quantum performance. Precision requests pay their posted funding once after a newly sampled passing trial. A historical result, ordinary trial, cancelled job or failed interval pays nothing. Both consume the existing apparatus duty; service/analysis allocation can slow them. Trial receipts preserve actual groups and are strictly validated on import.

**Procurement and commissioning.** Two final quotes buy either chip or control/cooling prefab units: 64 units for 220 funding and 12 designs with a 20-laboratory-second lead; or 512 units for 1,200 funding and 64 designs with a 90-second lead. At most two deliveries can be outstanding. Payment commits the resources immediately; there is no cancellation or appreciating financial asset. Countdown uses visible unpaused laboratory time. Arrival produces uncommissioned stock, not active physical capacity.

The existing workshop allocation commissions available prefab stock at twice its in-house stream rate, costing .4 funding per completed unit instead of .8 for in-house work. When stock runs out within a time step, the remaining stream time returns to the in-house rate. Interval accounting conserves stock, funding and assembly output. These are authored management coefficients. Workshops, funded duty and both chip/support streams remain necessary. Atomic apparatus work stops commissioning; external delivery countdown can continue during visible laboratory time. Delivered stock, pending shipments, preset upgrades and in-house work share the eventual 8,193 ceiling. Exact installed/support totals include only completed integer commissioning.

**Modern research decisions.** The optional controller study cites Q37/Q41 and costs 250 funding, 600 effort and 60 designs. It opens balanced `(1× throughput, 1× feedback latency)`, streaming `(2×, 1.5×)` and responsive `(0.75×, 0.5×)` classical presets. These tradeoffs are authored; they do not change `pEff`, `pL`, gate error or accepted factory-output error. Profile changes are refused during atomic jobs. The optional adaptive-control study cites Q38 and reduces only the selected maintenance requirement by 10%; calibration still consumes apparatus time and drift retains its floor. It does not copy the paper's injected-drift improvement factor into the game.

The optional state-readiness study cites Q39/Q40/Q43 and opens two alternative plans for the same 32-data-site task. Balanced retains width 32, depth 200, 4,800 operations and 192 fresh states. Compact uses width 32, depth 320, 4,800 operations and 128 states. Parallel uses 32 data plus eight workspace registers, depth 160, 6,400 operations and 256 states. Precision, repetitions and named task remain the same. Workspace, idle exposure, preparation, gates, state error, decoder streaming and feedback all remain in the complete budget. Compact trades greater depth for fewer fresh states, helping when factory throughput limits the schedule. Balanced or parallel can win as factories grow; increasing factories eventually meets a gate/feedback plateau. None of these counts is a compiled circuit or paper-derived resource estimate. Cultivation retries and alternative hardware connectivity are research readings rather than a silently substituted factory or patch model.

**Save transition and continued play.** The envelope remains version two. Newly created laboratories use `campusRevision:1`. A valid old version-two record is enriched with an empty campus block and revision zero only after strict baseline validation; partial new blocks are rejected. Explicit idle-only campus entry preserves all earned baseline fields, RNG state and historical evidence and grants no new receipts. A captured legacy VQE retains its ten-second schedule until it finishes or is cancelled. No legacy result becomes a new reward.

The final Audit cannot finish an ordinary campaign while a paid apparatus job is active: finish it naturally or explicitly cancel it, with entry costs remaining spent. Early Audit before scientific completion and unrelated research remain available during jobs. The first ending records its actual elapsed time, supported footprint, distance, discovery count, workload and executed recipe. Continuing the laboratory is a deliberate action that keeps this record and marks an epilogue. One-time grants remain one-time; remaining optional research and finite requests can be completed. Ending statistics and the gift postcard retain the first milestone even as continued laboratory time grows.

## Fresh dated studies and research persistence

Each 2015–2025 discovery needs a fresh successful paid study before its separate research purchase. Studies reuse the existing experimental apparatus. VQE jobs capture preparation, shot count, mitigation and sampled groups; logical studies validate the actual completion configuration. Successful studies award no separate repeatable grant. Cancellation or failure retains spent entry costs. The eleven purchases total 15,950 funding, 28,800 research effort and 4,090 designs, plus study entry fees.

The 2018 study accepts a precise measurement of an intentionally unsuitable preparation; it does not qualify the ground-energy result or simulate barren-plateau training. The 2022 study requires currently qualified memory, feedback ≤40 μs and zero allocated factories; its receipt records whether the actual stricter 20 μs comparison passes. Both comparison outcomes are legitimate. Planning credits never become stored quantum states. The 2024 condition checks authored decoder headroom, not decoding accuracy. The 2025 condition checks selected factory/gate lanes and footprint, not full workload success or cultivation yield.

New games have `researchRevision:1` and `researchResults:[]`. All eleven purchases gate Audit and the gift ending. An older complete version-two save with both fields absent retains revision zero and no inferred receipts; incomplete new blocks are rejected. Entering the programme is an explicit idle action after campus entry. For an already completed gift, Continue precedes that opt-in and the original ending record stays unchanged. Import recomputes captured receipts and their predicates instead of trusting success flags. Study purchases and ending checks preserve existing paid jobs.