anonymos-2321135/standing-not-truth

Standing, Not Truth — the data release

Every model response, hand label, ledger and per-experiment write-up behind the paper Standing, Not Truth: Who May Say What About Whom in Agentic Selection.

The code that produced all of it is a separate release: https://github.com/yedidel/standing-not-truth

What was measured

An agent chooses between competing offerings supplied by parties who benefit from the outcome. A lawful participant, whose identity fields are correctly bound and whose channel is not compromised, writes one false factual claim about a competitor. The claim carries no instruction. The experiments measure whether that moves the agent's selection, whether defenses placed on the origin of content or on its influence on the decision can separate it from benign traffic, and whether a check placed on the subject of the claim can.

Subsets

Each experiment is its own table, because each measured its own thing and they do not share a schema. The viewer's dropdown lists them.

subset	rows	what it holds
B35_model_panel	840	experiments/B35_model_panel/rows.json
B12_frontier_deciders	48	experiments/B12_frontier_deciders/rows.json
B13_harmonise_n	102	experiments/B13_harmonise_n/rows.json
B16_strip_prose	102	experiments/B16_strip_prose/rows.json
B17_endpoint_install	102	experiments/B17_endpoint_install/rows.json
B18_camel	102	experiments/B18_camel/rows.json
B19_gated_intervention	204	experiments/B19_gated_intervention/rows.json
B20_commerce_domain	120	experiments/B20_commerce_domain/rows.json
B22_frontier_endpoint	204	experiments/B22_frontier_endpoint/rows.json
B23_say_vs_do	102	experiments/B23_say_vs_do/rows.json
B24_matched_comparison	204	experiments/B24_matched_comparison/rows.json
B25_attriguard_matched	204	experiments/B25_attriguard_matched/rows.json
B26_wide_natural	353	experiments/B26_wide_natural/rows.json
B27_span_scoped	407	experiments/B27_span_scoped/rows.json
B28_multi_span	59	experiments/B28_multi_span/rows.json
B29_span_second_decider	425	experiments/B29_span_second_decider/rows.json
B30_attribution	425	experiments/B30_attribution/rows.json
B31_mood	216	experiments/B31_mood/rows.json
B33_inferential_gap	288	experiments/B33_inferential_gap/rows.json
B34_bracket	140	experiments/B34_bracket/rows.json
B39_avip_selection	90	experiments/B39_avip_selection/rows.json
B5_extend_n	102	experiments/B5_extend_n/rows.json
B9_signed_provenance	102	experiments/B9_signed_provenance/rows.json
labels	3942	the flattened view under data/
ledger	12340	the flattened view under data/

ledger and labels are flattened views, written by scripts/build_data_release.py in the code release. The ledgers carry two schemas across 27 files and a nested meta field, and the hand labels are dictionaries keyed by unit, so neither loads as it stands. Nothing is added: ledger drops an always-empty balance_after column and lifts the traceback out of meta, and labels turns each dictionary into one row per item. The per-experiment files under experiments/ remain the record.

One pass labelled four arms per unit and stored them as a list, so those rows carry a position and the rest leave it empty. Its own note gives the order: attack, control, check, sentence, with H the honest cheaper product, P the promoted one, A asserts, M mentions, and an empty label where the arm was not run for that unit.

Left out of the subset list, because they hold a mapping rather than a list of rows: B15_frame_anomaly, B36_extractor_transfer. Both are present as files.

Layout

LIMITATIONS.md     all 34 limitations, with their states
data/
  ledger.jsonl     every metered call, one row each
  labels.jsonl     every hand label, one row each
experiments/<id>/  one directory per experiment, 52 of them. `A*` are the
                   Phase-A feasibility experiments, `B*` the Phase-B programme
                   the paper reports.
  RESULTS.md       the write-up: numbers, denominators, reading notes
  rows.json        one row per unit, with the replies and the labels
  labels.json      hand-assigned outcomes, where a pass was made
  raw/             the stored response for every call, one file each
  ledger.jsonl     one append-only record per metered call

Forty-seven directories hold a RESULTS.md. The five that do not are B14_annotator_agreement (README.md), B2_binding_taxonomy (TAXONOMY.md), B35_model_panel and B36_extractor_transfer (run/RUN_REPORT.md, with the panel's reported numbers in B37_relabel/RESULTS.md), and B28_multi_span, which holds stored rows and no write-up because nothing in the paper rests on it.

Limitations

LIMITATIONS.md holds every limitation this work recorded, 34 of them, each with its state and, where one exists, the measurement that settled it. None is open. The paper's Limitations section carries the load-bearing entries; this is all of them, including the two pre-registered theses our own experiments refuted.

The ids are not contiguous with the order the limitations were recorded in. Five were added later and given ids an earlier block already held; they are L30 to L34. Two result files here, B19 and B27, cite L21 and L22 in the earlier sense, so the earlier block kept its numbers.

How to read a result

Four conventions run through every write-up.

1. Scoring is by reading, not by code. No script decided whether a reply was a hit. The annotator is an LLM reading replies one at a time, so its reliability is measured rather than assumed: against an independent human pass over a stratified sample of 201 items, Cohen's kappa is 0.933 on the choice task and 1.000 on the subject task, with every disagreement adjudicated, and a later 183-item pass over the thirteen-model panel gives 1.000. That frame covers B13, B5, B12 and the panel; the remaining experiments carry no second pass.
2. Denominators are stated and never pooled across populations.
3. Every rate carries a Wilson 95% interval, and where a unit is decided by several models the interval is also computed by cluster bootstrap, because treating thirteen decisions on one base pair as thirteen independent observations understates the width.
4. Negative and refuted results are kept. Two pre-registered theses were refuted by our own experiments and both are recorded rather than removed.

B37_relabel is the later pass in which all 1,689 stored replies of the three largest experiments, and then all 1,682 replies of the thirteen-model panel, were read again one at a time. It moved four published rates, left the attack outcomes unchanged, and supplied the panel result. The panel's blanket-removal arm is a baseline floor rather than a defense result, and reading it as one matters: command-r7b lands the attacker's package on 22 of 45 units with the description deleted entirely, against 7 of 45 with only the claim removed, because it reads the earlier of two publish dates as the more recent. Without that floor, one model's arithmetic would be scored as the defense failing.

Ledgers

Every metered call is recorded in an append-only JSONL ledger next to the experiment that made it: the request, the model, the cost, and a SHA-256 over the stored raw response. verify_ledger.py in the code release checks them and classifies what it finds:

ledger records examined : 12340
  digest verified       : 9863
  superseded by a retry : 1688
  empty, asked again    :   29
  no digest stored      :  760
  UNEXPLAINED           :    0

Use that script rather than hashing the files directly. Two undocumented facts stand in the way: the digest was taken before the writer applied the platform newline translation, and the backfilled records digest one row inside the named file rather than the file.

The two middle rows follow from one design decision. The ledger is append-only; the response store is keyed by unit id and is not. When a unit was called twice the second response overwrote the first file, so the earlier record's digest can never match again. A superseded record counts as accounted for only when a later record over the same file matches the file as it now stands; one with no such partner is reported as unexplained. The 29 are fifteen claude-opus-5 units that returned no content, which the harness recorded as ok with zero tokens without writing a file, and which retry_empties.py asked again under a |retry unit id. The 760 without a digest are calls that failed outright.

Some stored responses contain an opaque encrypted data field. That is the provider's own encrypted reasoning payload, kept as returned so the digests verify. Absolute paths inside stored tracebacks are replaced by <project>, <harness> and <python>; every frame and line number is untouched.

Licence

Data and documentation under CC BY 4.0. See LICENSE, which also covers the code release and the third-party material not redistributed here. The npm and marketplace text under study belongs to its original authors and is included as research data.