Exact prospective collection plan, 2026-10-03 UTC

512 episodes; at most4096 model calls; eight routes; seven templates representing
four mechanisms; two procedure arms; two fixed messages; one repetition per cell.
Proprietary frontier routes precede Chinese hosted open-weight candidates.

Conditional price USD691.233931264, using refreshed public endpoint metadata,
20,000 assumed input tokens and4096 billed output tokens on every possible turn,
advertised surcharge overrides and no cache discount. This is deliberately
conservative conditional arithmetic, not guaranteed cost. Actual early stops and
short responses may cost less. The20,000-byte input gate does not prove the assumed
token bound, and hosted checkpoint identity/license claims need verification.

Original authorized lifetime cap remains USD50. Accounted amount is
USD37.8088680376, including unresolved reservations. Remaining headroom is
USD12.1911319624. No additional funding is authorized by these files. A question
asks the owner whether to retain USD50 or authorize USD800 lifetime; absent an
explicit answer, retain USD50. Never treat elapsed time as approval.

This is a hash-linked matrix/analysis preregistration, not a ready-to-run complete
collection freeze. Authenticated route probes, collector/reanalysis source freeze,
final prior-case equivalence review and a compatible budget decision remain.
Any source amendment must preserve the original preregistration and identify
affected comparisons before new requests. Existing paid results are unchanged.

Holdout scope
256 planned episodes replicate development mechanisms.256 evaluate a pending-
physical-effect extension not used in paid model development. The latter was
scripted, inspected and published during simulator validation. No independently
authored, secret or contamination-free structural holdout is claimed. Four timing
variants represent one mechanism. One repetition supports descriptive paired
outcomes, not broad prevalence estimates, powered rankings or general alignment.

Review files
necessity-structural-collection-manifest.json contains every cell, public-packet
hash, source hash, instruction, message, route price and conditional assumptions.
necessity-final-analysis-protocol.txt freezes endpoint and missingness reporting.
necessity-expanded-prior-audit.txt documents close precedents and unresolved gaps.

Validate without network/model calls
python -c "import json; from benchmark.necessity_collection_plan import validate; validate(json.load(open('plans/necessity-structural-collection-manifest.json'))); print('Prospective preregistration validated')"