Which promises fit?
Which tests pass?

Protection guarantees collide in constructed cases. No loss-to-win score or rank reversal was found in the separately tested ranking fixtures. These are different findings with different limits.

A two-place ranking change can make eleven exemptions necessary.

In a nine-conference fixture, the ten pre-title top-ten teams all play title games. Swapping only final ranks 9 and 11 changes the selected champion set: five selected champions plus the retained top-ten claims require eleven opening exemptions. The fixed bracket offers two double byes and eight single byes: ten exemptions. After the double byes, nine claims compete for eight single byes. No tie-breaking procedure can preserve every guarantee. A separate, more extreme witness in the broad study requires fourteen exemptions.

A second boundary fixture swaps only final ranks 14 and 15. Four retained hosting claims still have four slots, but the claimants are not the first four opening teams in final-rank order. That is an ordering conflict, not a shortage of host slots. The two small-movement examples were designed as follow-ups after inspecting the broad stress study; their final ranks are declared inputs, not NRR score refits.

6,144constructed cases

1,024canonical allocations completed

2,048 / 3,072bye-capacity / host-order stops

The design enumerates 512 winner patterns, four participant rotations and three declared final orders. Every case has nine disjoint title games, one winner per conference and coherent participant identities. Final orders are declared stress inputs, not NRR refits or estimated postgame rankings. These counts do not estimate how often a real season would produce a conflict. No berth or double-bye overflow arose in this design.

Complete alternatives expose the sacrifice.

The canonical allocator still stops. The research alternatives below apply only after a demonstrated conflict; successful canonical cases are copied unchanged. All retain eighteen entrants, two double byes, eight single byes, four hosts and four visitors. The underlying final ranking stays unchanged for reseeding.

Unadopted policy choices; capacity changes require a different bracket study
Bye priority	Host priority	What can be lost
Selected champions first	Retained claims first	Some retained single-bye and hosting claims; final-order hosting can also give way.
Selected champions first	Final order	Some retained single-bye and hosting claims.
Retained byes first	Retained claims first	Some champions play an opener; final-order hosting can give way.
Retained byes first	Final order	Some champions play an opener and some retained hosting claims are lost.

All four alternatives sacrifice at least one promise in each of the 5,120 stopped cases. They are review options, not a change to the proposal. No extra berth or exemption is silently added. Within each priority class, retained claims use pre-title order; other places use final order.

Inspect the exact witnesses and alternative allocations

Full conflict identities, allocations and lost promises. Each stage retains an exact fixture with participants, champions and paired orders. Four existing NRR historical allocations across 2024 and 2025 reproduce unchanged; the two committee baselines remain preserved in the earlier report.

No reversal found in these full-refit ranking tests.

Each of two separately hashed generators—the original property-audit generator and the current source generator at this experiment’s freeze—was evaluated on 34,176 tournaments and 504,576 loss-to-win reversals. Every altered result recomputes opponent values and model outputs. Across both sources, 1,009,152 reversals produced no changed-winner Index decrease, strict-rank worsening or display-rank worsening.

Counts per generator; same fixtures for both sources
Fixture family	Tournaments	Result reversals
4 teams · complete · neutral	64	384
5 teams · complete · neutral	1,024	10,240
6 teams · complete · neutral	32,768	491,520
6 teams · ring-plus · neutral	256	2,048
4 teams · complete · venue-and-margin	64	384

Strict rank counts opponents whose reconstructed published Index is greater by more than 0.00001; display order additionally uses team strength and name. The complete four-, five- and six-team round robins include every outcome vector. The six-team ring-plus schedule tests one irregular topology. A separate four-team family changes venue context and coherent score margins.

These tests use zero team-efficiency inputs and no external polls, play-by-play or title context. A finite pass does not prove monotonicity generally, settle incentive compatibility or validate deservingness. It also does not test adding a weak win or every way an opponent can perform worse. The earlier fixed-component weak-win finding remains a separate result. No ranking weight or production formula was changed.

Turn the remaining questions into observations.

The separately labeled champion-first, retained-host candidate matches six historical allocations, but a controlled follow-up found winning can cost hosting despite an improved final rank. The candidate is now held for repair. The original stress results remain unchanged. See the candidate page for the exact rule and lost-claim ledger. More compatible pre-title/final snapshots require verified title inventories; the local ranking inventory currently covers 2024 and 2025. The ranking extension still needs title, poll, efficiency and broader irregular-schedule contexts before stronger claims.

The blank one-game travel inventory distinguishes a location-only scenario from a scenario where spectator origins and attendance change. It requires aggregate origins, transport modes, occupancy, permitted sources and unique journeys. The proposed Austin–San Antonio contrast is hypothetical; no travel total or collaboration commitment is claimed.

The reader-session kit records unaided explanations, score/weight/probability confusion and a win-cycle question. Five willing readers and their actual responses are still needed. Blank forms are not observations.

Suggested review wording: “When guarantees cannot all fit, the rules must say which takes priority before the season. The examples here expose the choice; they do not adopt one. Ranking checks report exactly which inputs passed, while broader properties and real-world benefits remain open.”

Reproduce and inspect

From the matching source, run node scripts/audit-protection-stress.mjs, node scripts/audit-nrr-monotonicity.mjs, then node --test tests/protection-stress.test.mjs. Each experiment was frozen before evaluation and checks source hashes. Original reports and ranking archives are preserved.

Historical protected fields · Original ranking properties · Evidence index