Artifact index¶

Every measurement this study rests on is a committed file under datasets/results/, and every one of them was written by a script in this repository. This page is generated from those scripts: 682 artifacts under 67 writers, grouped by the script that wrote them, each described by that script's own summary line. Run make docs-index to rebuild it, and the test suite fails if it is stale.

A run that varies a knob names its output after the knob, so the variants of one measurement sit together: a cluster suffix keeps an off-TIGRIS run from overwriting a GH200 result, and RUN_TAG, SEEDS, TRAIN_SIZE and CLIENT_SIZE name a run that is a different study rather than a repeat of the same one.

scripts/adapter_geometry.py¶

How alike are two LoRA updates, and is it the data or the initialization that decides?

scripts/adapter_projection.py¶

Each client update as a short vector, so defences can be simulated honestly.

scripts/admissible_partitions.py¶

The ordered list of admissible partitions H1's runs are drawn from, fixed before any run.

scripts/aggregate_attack.py¶

RQ2 under secure aggregation: does a round's aggregate betray whose clients were in it?

scripts/ai_trailers.py¶

How many of an organization's examples come from a change whose commit names an AI tool.

scripts/backport_share.py¶

How many examples could carry code older than the checkpoint by way of a backport.

scripts/bot_sensitivity.py¶

The registered dev-window reading with examples that rest only on automated reviewers removed.

scripts/censoring.py¶

How much of each window's true cohort the collection boundary removes.

scripts/client_attribution.py¶

Read the client updates' geometry as a source-attribution attack, at three altitudes.

scripts/client_updates.py¶

Federated client updates for RQ2: many small adapters, one initialization, several sources.

scripts/cluster_informativeness.py¶

Is a change's size informative about its outcome?

scripts/comment_count_probe.py¶

Whether a host's total_comment_count of 0 means a row has no comments, asked of the host.

scripts/conjunctive_power.py¶

The power the RQ1 gate actually has, which is not the power that was computed.

scripts/contamination_battery.py¶

Outcome-neutral test 1 on real data: is the post-cutoff corpus less familiar to the model

scripts/contamination_blind.py¶

Outcome-neutral test 1, read against blind baselines: do the windows separate without the model?

scripts/crossed_coverage.py¶

How often does each interval exclude zero when there is nothing to find?

scripts/crossed_reread.py¶

Re-read single-seed runs as one multi-seed study, under both gates.

scripts/data_audit.py¶

A repeatable data audit of a built and refined corpus, one artifact per organization.

scripts/dataset_funnel.py¶

How each organization's corpus is filtered, step by step, from its build manifest.

scripts/decoder_read.py¶

Read the repetition-penalty check: what the checkpoint's penalty changed in stored results.

scripts/decoding_check.py¶

Is greedy decoding what amplifies a planted convention?

scripts/decomposition_pilot.py¶

The registered decomposition gate, read on development-window runs as a pilot.

scripts/decomposition_sensitivity.py¶

The smallest half-split effect the decomposition gate detects, and how often a null is bounded.

scripts/defence_curve.py¶

What masking an update with noise costs the attacker, and what it does not cost.

scripts/determinism_check.py¶

Where does greedy decoding's run-to-run variation come from?

scripts/dual_adapter_updates.py¶

The adapter-instance cut: two adapters per client, and only one of them leaves.

scripts/github_conversion.py¶

GitHub's thread-to-example conversion, read from a built pilot month of a GitHub organization.

scripts/github_sizing.py¶

Confirmation sizing of GitHub organizations: a two-stage sample of PRs, one row per draw.

scripts/github_sizing_report.py¶

GitHub candidate organizations sized from github_sizing.py draws: weighted totals and yield.

scripts/host_scoping.py¶

Which projects on a candidate Gerrit host could carry an organization, before collecting it.

scripts/interval_calibration.py¶

False-positive rate of the registered pairs cluster bootstrap, under a true null.

scripts/k_coverage.py¶

An organization's K under the coverage rule, as the K source a test read takes (--sizing).

scripts/label_audit_agreement.py¶

Agreement on the label audit: two blind model raters, and a human's blind check of one.

scripts/label_audit_by_org.py¶

The label audit read again on one organization's items, from the committed agreement report.

scripts/label_audit_open_rater.py¶

An open-weight, pinned model as a further blind rater on the label audit.

scripts/likelihood_score.py¶

Score a stored H1 partition run by the log-probability of each held-out reference.

scripts/likelihood_sesoi.py¶

The SESOI in log-probability per token, carried from exact match by the registered rate rule.

scripts/masking_mechanism.py¶

Why does masking an update with noise make the attacker better?

scripts/memory_probe.py¶

Does the registered 7B fit one GPU, training and inference, at the longest admissible item?

scripts/module_split.py¶

Does the organization survive in the modules a selective scheme actually transmits?

scripts/notedb_parity.py¶

The NoteDb git route against the REST-built AOSP corpus: enumeration, fields and examples.

scripts/partition_pilot.py¶

H1 over repeated partitions on real runs: the cell, the number of runs, and the runs' order.

scripts/partition_sensitivity.py¶

Operating characteristics of H1 over repeated partitions at the registered K, by simulation.

scripts/partition_variance.py¶

How much of H1's reading is the partition: alternative balanced splits against the seed effect.

scripts/pilot.py¶

Pilot: does an ADAPTED 7B clear the exact-match floor on this corpus?

scripts/placebo_corpus.py¶

Two pseudo-organizations built from one organization's own projects.

scripts/planted_exchange.py¶

The exchange rate between exact match and s, read on a planted code-level convention.

scripts/power_rq1.py¶

Minimum detectable matched-minus-mismatched exact-match difference, from an RQ1 pilot.

scripts/project_windows.py¶

How large will the sealed test window be, and does the capture model predict the dev one?

scripts/prompt_format_probe.py¶

How much of the pilot's base-model floor is the prompt format?

scripts/quasi_independence.py¶

Does the lag distribution hold still across creation cohorts?

scripts/retrieval_comparator.py¶

Few-shot retrieval as the adapters' comparator: the generations, one job per pool set.

scripts/retrieval_read.py¶

Read the retrieval comparator: its contrasts over partitions, beside the adapters' H1.

scripts/reviewer_overlap.py¶

How far an evaluated half's dev reviewers also reviewed its sibling half's training data.

scripts/rq1_pilot.py¶

Pilot-scale RQ1: does the matched adapter beat the mismatched one, per organization?

scripts/rules_distil.py¶

Distil the rules files the rules-file comparator prompts with, by the pinned distiller, greedy.

scripts/run_shift.py¶

How much run-wide shift a likelihood pilot carries beyond its change-by-run noise.

scripts/seed_effect.py¶

How large is the seed main effect: the shift a retrained seed gives every change at once?

scripts/sensitivity.py¶

The smallest organizational effect the confirmatory design detects, by seed count.

scripts/sensitivity_ids.py¶

The examples each registered sensitivity of H1 removes, by id, for one organization and window.

scripts/separability.py¶

Can a classifier tell the organizations apart from their review text?

scripts/separability_over_time.py¶

Is the organizational signal fading while the study measures it?

scripts/sibling_leakage.py¶

How much of an evaluated half's dev window a sibling half's training data already holds.

scripts/split_criteria.py¶

Whether an organization's registered split qualifies for a confirmatory H1 cell.

scripts/subspace_split.py¶

Would keeping the client-specific subspace local take the house style with it?

scripts/training_size.py¶

The fixed training size N every adapter trains at, derived from the organization's own splits.

scripts/window_report.py¶

What the collected windows contain, and how much leaks across their boundaries.

Written by no script this page can find¶

These are committed artifacts that no claim_result, no --out example and no script name accounts for. Each is a measurement whose producer has to be read out of the research log rather than out of the code.

• manifest-openstack-v2.json
• rq1-qtfull-fp32-seeds.json
• rq1-r256-seeds.json
• rules-reading-openstack-rp1.0.json
• rules-reading-openstack.json
• rules-reading-wikimedia-rp1.0.json
• rules-reading-wikimedia.json
• stratified-coverage-0.95.json
• stratified-coverage-0.975.json
• stratified-coverage-h1-0.95.json
• stratified-coverage-h1-0.975.json