Anchor-fixed renaming: prospective study protocol

Status: prepared offline; no requests have been sent and no model outcomes exist. This is a new study informed by the completed September 2026 campaign. It is not an independent rediscovery or an external preregistration. freeze.json records the final prepared bytes; any review-driven change must receive a new manifest before execution.

Question and scope

Does consistent renaming of internal graph identifiers change Jev's selected labels beyond matched exact-request repeat disagreement when the literal question, answer candidates, abstract graph, record positions, and marginal identifier-role frequencies remain fixed? The response is a behavioral statistic. It does not identify an architecture, tokenizer, attention mechanism, or an optimal naming convention.

The previously observed unrestricted renaming effect motivated this design. No prior response or per-world error is read by the generator. No world is selected for model difficulty or expected effect. All published code is offline and uses the Python standard library.

Cohort and tasks

The main cohort contains 384 fresh independently seeded forest worlds, with 128 worlds in each of three fixed topology strata. Each world contains four roots, four cloned branches per root, and 32 non-root nodes per branch: 512 non-root nodes and 516 total nodes. The branch has an eight-edge spine and 24 additional nodes. Attachment laws are:

1. recursive: uniformly select an eligible existing parent whose depth is below eight.
2. spine_fan: attach additional nodes to a randomly selected spine position at depth one through seven.
3. clustered: choose three spine hubs, then mix hub attachments with recursive offshoots.

The query target is the same depth-eight spine leaf for all tasks in a world. Root queries use all four roots as alternatives. Fourth-ancestor and immediate-parent questions use the corresponding nodes from the four branches of the target's root. For each task, all four candidates are in one automorphism orbit of the directed rooted forest when the query marking and literal names are ignored. The independent auditor proves this using marked rooted-tree signatures. This excludes unique candidate depth, degree, subtree size, and branch-role explanations; it does not prove target-linked traversal by a model.

Target root and target branch, gold answer-key position, candidate lexical rank, and root declaration position use separate, independently seeded balanced four-cycles within each family. Name assignment is consequently conditioned on these constraints, not iid over structural roles. The audit reports query-blind first/last-key and lexical-minimum/maximum baselines. The three families are specified strata, not a sample of three real-world domains. Independently drawn worlds may share unlabeled topology; the audit reports collisions rather than treating topology uniqueness as a selection criterion.

Treatments

Each world has one preselected map per treatment:

• original: identity map.
• unrestricted: a uniformly sampled derangement of the eligible internal identifiers, conditioned on moving every eligible identifier.
• constrained: independent uniformly sampled derangements within each equal (child-record mentions, parent-record mentions) bin, conditioned on moving every identifier in each eligible bin.

The union of every identifier in all three tasks' instructions and alternatives, including all roots, is anchored. Unanchored singleton bins are fixed in both treatments. The eligible set is exactly the union of nonsingleton bins. Both treatments therefore change the identical node set and number of names, with no favorable-map search. Each abstract role receives the literal identifier of its image under the bijection. All names have the same character length. The constrained map preserves the complete name inventory, per-identifier role counts, full word/punctuation multiset and character bag. Both treatments preserve question/candidate bytes, roots, abstract graph, record positions, state byte length, and all exact answers. The constrained map can move which name occurs at a position: positional associations intentionally change. Equal bags and byte lengths do not guarantee equal contextual tokenization or model-internal serialization.

Two layouts are crossed with naming: an arbitrary shuffled edge order and a root-grouped order retaining the shuffled relative order within each group. Root grouping is computed from the graph and exposes membership. Grouped root success admits block-membership shortcuts. It is a secondary accessibility condition, not evidence that the model traverses edges.

Each task is served as a separate single-question Choice request, with model pinned to jev-1.13.0. This deliberately controls sibling-question context even though the vendor recommends batching in ordinary applications. The direct target-parent sentence is literal-identical across all naming treatments. Immediate-parent questions are separately budgeted descriptive controls, not extra primary endpoints.

There are 6,912 distinct request bodies, each scheduled twice. The sampling unit is the forest, not the request, map, copy, task, or candidate. Parent controls use the same worlds and do not add independent units.

Scheduling

Shuffle complete world blocks. Within each world, shuffle the six layout-by-task blocks. In each block, select uniformly among all 720 permutations of the six naming-by-copy conditions, including adjacent exact copies. Within each family/layout/task, pairs of consecutive world indices receive a randomly chosen permutation and its reverse. Thus each naming/copy has mean within-block position exactly 2.5 in every stratum. Worlds themselves are dispatched in shuffled order.

The audit exports within-body and cross-name pair-gap distributions, mean positions and body identities. Those empirical gap distributions need not be identical; elapsed time gaps are not assumed equal merely because start positions are balanced. Execution must retain actual request-start timestamps and dispatch/body-completion durations. Copies one and two are experimental labels and need not occur in numerical order. This design estimates local serving variation over short randomized blocks, not day-scale drift. Exact-copy repeats must bypass any client response cache. Hidden provider caching remains unobservable.

Prospective scheduling correction: the first offline draft excluded adjacent exact copies, which systematically restricted repeat lags relative to cross-name lags. Before any execution, that restriction was removed while retaining the mirrored role-position balance. The full request jobs, bodies, maps, oracles and counts are unchanged; only execution order changes. The previous schedule is retained in private working history and excluded from the public release bundle.

Primary estimands and inference

For one world/layout/task, let original-copy selections be a1,a2 and renamed-copy selections b1,b2. Define:

cross = [I(a1!=b1)+I(a1!=b2)+I(a2!=b1)+I(a2!=b2)] / 4

repeat = [I(a1!=a2)+I(b1!=b2)] / 2

excess = cross - repeat.

Negative estimates are retained; the observed world statistic lies in [-0.5,1]. Under independent, exchangeable copies from stable categorical selection distributions its expectation equals half their squared L2 distance. That interpretation is conditional and is not required to report the empirical difference. Independence and exchangeability of API calls are unverified.

Exactly two endpoints form the primary family: constrained-renaming excess for shuffled root and constrained-renaming excess for shuffled fourth ancestor. Average within each topology family, then weight the three family means equally. Use 10,000 stratified whole-forest bootstrap draws for percentile 95% intervals and 10,000 two-sided sign flips of world contrasts with the plus-one correction. Holm correction retains a family of two even if an endpoint is unavailable. Sign-flip inference uses a symmetry/exchangeability approximation; it is not an exact randomization test of a randomly assigned representation. Bootstrap intervals are not simultaneous, and saturated intervals can be degenerate.

The smallest effect of engineering interest is 0.10, or ten percentage points of excess disagreement. Report effect sizes, individual intervals and corrected p-values together. An interval wholly above 0.10 would support a materially large effect for this workload. A nonsignificant result is not evidence of invariance. Do not turn a comparison between statistical significance levels into a direct treatment comparison.

Secondary, without additional p-values: unrestricted treatment, grouped layouts, parent controls, signed accuracy changes, repair/harm, raw complete-vector half-L1 distance, and constrained-minus-unrestricted excess on the same complete worlds. All are reported; no best map/layout is selected. Raw probability coordinates are not normalized or treated as confidence in correctness. Available-call accuracy/probability summaries are distinct from complete-world paired estimates.

Precision and power planning

The cohort is fixed at 384 for this prepared study. precision.json reports normal-approximation planning for assumed world standard deviations 0.20, 0.35, 0.50 and 0.75, using a conservative two-test Bonferroni critical value. It also exactly enumerates the moments of several explicitly hypothetical independent-copy categorical distributions. None is fitted to new observations.

The statistic's range gives a worst-case SD bound of 0.75. At SD 0.50, 384 worlds are compatible with approximately 95% normal-approximation power for a 0.10 effect; at SD 0.75, power is only approximately 64%. This is not a power guarantee for the registered sign-flip procedure. Family heterogeneity, dependence, missingness and a smaller true effect can make the study less informative. Any later sample-size change must be documented before outcomes; there is no optional stopping or sample expansion rule in this package.

Integrity, missingness and endpoint validity

The analyzer first verifies the frozen source and artifact hashes, then requires exactly one attempted row per planned name across the provided logs. Requests, expected labels and metadata must match the frozen job in ordered JSON, and the request-body hash must agree. Duplicates, unplanned jobs, missing attempts, reordered candidates and mixed synthetic/real evidence cause a hard failure. An interrupted cohort cannot silently become a smaller completed study; record the interruption and finish only untouched jobs under the original plan, or declare the study incomplete.

A selected label is valid only for HTTP200, no error field, explicit validated_success:true, the pinned response model, a Choice answer and a string among the four declared keys. Invalid probability maps do not erase an otherwise valid selected label. Complete vectors require exactly all four finite coordinates in [0,1]; totals within 0.02 plus numerical tolerance of one are accepted as independent cent rounding. Missing coordinates remain missing. Confidence and nonnegative integer usage have separate validity checks. A selected label that differs from the printed argmax is flagged; selected-label estimands retain the actual returned selection.

Each naming contrast requires all four selected labels (two original, two named) in a world. The constrained-versus-unrestricted secondary difference requires all six. Probability distance requires all four complete vectors separately. Failure in an unrelated treatment does not exclude a valid contrast. Report all exclusion IDs, family counts, failure categories and available-call denominators. An equal-family mean is undefined if any family has no complete worlds. Complete-case inference does not repair nonrandom service failure; report finite-cohort missing-outcome bounds using the statistic's [-0.5,1] range. No failed request is silently retried or replaced.

Remaining limits and execution boundary

These are three synthetic attachment laws with cloned branch symmetry and a fixed depth-eight target. They do not represent arbitrary graphs, natural language or application performance. One random internal map per treatment per world does not measure every possible renaming. Frequency-constrained and unrestricted permutation spaces differ by design. Root grouping exposes graph structure, and the same-root ancestor task still permits layout cues. Token identity at the service is unknown. One model version and one serving period do not establish general architecture or cross-service behavior.

All tasks are mechanically solvable; the exact code oracle is the appropriate production implementation. The service calls would be diagnostics. This package has no API client, key, .env file, paid execution command, global installation step or public deployment. Before future live execution, the coordinator checks model availability, current limits and budget, runs a disjoint operational smoke test if required, disables client retries/cache, and uses a monitored central runner. Independent review is recorded separately from the byte freeze. Preparation requires no additional approval flow; paid execution remains outside this task.