thwalsh/moderators-playbook-benchmark

The Moderator's Playbook Benchmark

A five-tier offline contextual-bandit dataset with exact logging propensities on a fully crossed core and scorable off-policy evaluation (OPE).

Version 4.0.3, 2026-10-16, changes the documents and the graders only: it names the reader of the readability judge's validation labels, a Claude model session; it specifies Task 2 on the core with its evaluator and four baselines; and it adds a graded positivity variant of Task 1. Version 4.0.2, 2026-10-15, changed the documents and the Task 1 grader only: every off-policy estimator is now fit on the logged interactions alone. Version 4.0.1, 2026-10-14, changed the documents only: what a policy learned from the log reaches, and the page's history. Every data file is version 4.0.0's, byte for byte. Version 4.0.0 regenerates version 3.0.0, withdrawn on 2026-09-30 because its participant posts decayed round by round, with the same cells, levers, contexts, seeds and logging propensities. Every participant post now goes through a readable-post gate and is graded by an independent readability judge, and every link, email address, user mention and participant label in the text is masked before release. CORRECTIONS.md carries the record.

What it is

19,520 logged transitions from a simulated moderation forum, across five tiers. Each row is one transition: a context (state), a generation-lever action, the logging propensity mu under which that action was taken, and seven objectives. Further columns record each cell's text-health verdict, its readability as graded by the independent judge, and the accelerator that produced it.

The corpus ships in five tiers:

tier	cells	arms	contexts	seeds	mu	positivity
core	12,000	20	12	50	1/20 = 0.05, exact and uniform	240/240 (arm,ctx) filled, 0 empty
periphery	6,000	200	12	15	76 distinct (69 at six decimals), 0.001826 to 0.030659, recorded	336/2,400 empty (14.0%), by design
controls	720	6 named arms over 4 lever sets	12	10	mechanism tier, not an OPE surface	6 arms x 120 cells
stress	600	10	6 OOD	10	1/10 = 0.10, exact	60/60 filled
decode	200	20 arm indices over 14 distinct lever settings	1	10	characterization only, never merged into the reference

The core grid is fully crossed: every one of the 20 arms is logged in every one of the 12 context strata. Because the core behavior policy is the designed grid rather than a scraped log, mu is exact by construction (a design fact, not a fitted estimate), and every arm's outcome is known for every interaction, so a target policy's value on any logged interactions is computed rather than estimated. That pairing is what lets an OPE estimator be scored against a known target, which a scraped log cannot do.

Positivity holds everywhere in core: mu(a|s) = 0.05 (1/20), uniform across all 12 strata, and 0 of the 240 arm-context pairs are empty. Periphery is logged by a recorded softmax policy with the propensity written down at sampling time (76 distinct values, 69 when rounded to six decimals, importance-weight ratio 16.79x, effective sample size against a uniform target 55.0%, 3,298 of 6,000). Its 336 coverage holes are deliberate: they are the scraped-log analog with propensities that are still known, and an estimator that requires support has nowhere to hide there.

Schema

Read the column lists and types from croissant.json, the generated consumer contract, and the objective definitions from the table below. Do not hardcode the s_* / a_* / r_* names.

group	columns	notes
state (state_columns)	s_pool_imbalance, s_bad_actor_frac	the 2 pre-action exogenous context fields; the 12 context strata are their crossing
action (lever_columns)	a_clamp, a_delta_scale, a_temperature, a_rank, a_mc_dropout, a_cfg_scale	the 6 portable levers (a_rank is the adapter rank mask, a_mc_dropout the MC dropout rate); the arm is keyed by config_idx (unique per lever tuple), 20 arms in core
steering coefficient	a_arditi_alpha	swept, at 8 values from 0.0 to 2.0. It is the axis along which the between-cohort effect is attributable, so removing it would remove the benchmark's most informative dimension. PORTABILITY CAVEAT: the patch needs our Pythia-1.4B direction tensor, so a consumer reproducing generation must supply their own; a consumer using the logged trace is unaffected.
behavior policy	mu_logging	the logging propensity mu(a
reward	the raw objectives in the table below, plus two blends shipped beside them for lineage, r_total = r_util + r_safe + r_fid with the gap SIGNED and r_tension = r_util + r_fid + 0.5 r_fair + 0.25 r_sys. The OPE task grades r_util + r_safe + abs(r_fid) (deviation D17) and every objective alone, computed from these columns	none are pre-scalarized; components live on different scales; r_fid is on the decision channel and its sign is set by the steering label
id / bookkeeping	cell_id, tier, config_idx, seed	tier in {core, periphery, stress, decode, controls}; there is no qc_reject column, every shipped cell is a valid transition

Reward

All objectives ship raw, none pre-scalarized:

All objectives ship raw, none pre-scalarized, one value per cell. Definitions are read from the trace builder at generation time and this table is regenerated, never hand-edited.

objective	definition
r_util	moderation utility: true-positive rate on bad-actor and canary removals minus the false-positive rate on benign posts
r_safe	1 minus the mean toxicity of the posts the moderators KEPT, scored by the primary RoBERTa classifier
r_fid	the decision-channel cohort gap, empath minus rage act rate on participant posts; its sign follows the labels supplied to the steering (deviation D7)
r_indiv	within-cohort individuation, binomial-corrected excess variance of per-moderator act rates; defined on readable forums only
r_fair	minus the participant-cohort disparity in benign removal; reported, never optimized
r_sys	minus the deterministic count of forward passes per generated token; never a clock
r_read	coherence (deviation D18): the share of a forum's participant posts that the independent readability judge reads as comments
r_coh	lineage only, the sign gate D18 replaced: 1 when the gap is positive and the canary was caught, else 0, blind to broken text

Two further columns ship beside them: text_health, the joint verdict of the text check and the readability judge on the forum (HEALTHY, MARGINAL or BROKEN), which is the column to filter on for readability, and fleet, the accelerator that generated the cell.

Three readability columns ship beside r_read: r_read_gate, the share the generation gate itself passed, the filter's own view and never the grade; read_first_draw, the share that passed on the first draft, the setting's own ability to write before any redraft; and read_source, the instrument behind r_read, the judge on all 19,520 cells.

Two objectives can be undefined on a cell, and the count is stated rather than filled: r_fid on 400 of 600 stress cells (every cell at pool imbalance -0.5 and +0.5); r_safe on 25 of 600 stress cells. The cohort gap is undefined where every moderator in the forum comes from one cohort, so there is no second cohort to difference, and kept-post toxicity is undefined where every post was removed. The shipped r_total treats an undefined component as zero on those cells; recompute it from the raw columns if that is not the convention you want.

The decision channel carries the persona and its direction is the label's. Over the 12,000 core cells the mean cohort gap (empath minus rage act rate) is -0.160 and 87.3 percent of cells share its sign; the prespecified direction was the opposite one and the reversal is filed as deviation D7, so nothing here rests on the sign. The gap is a dose response in the steering coefficient: mean absolute gap 0.013, 0.114, 0.190, 0.238, 0.266 at steering 0, 0.25, 0.5, 0.75, 1. The injected canary is caught in 85.0 percent of core cells overall, falling from 1.000 at zero steering to 0.592 at full steering, which is the trade the levers impose: steering buys separation and costs moderation competence (gap magnitude against canary catch, -0.39 across cells). Fairness does not move with fidelity in one direction: the signed gap and r_fair correlate at -0.14 across cells, the gap magnitude and r_fair at +0.14 across cells and at -0.45 across the 243 arm means, which is why r_fair ships as a reported outcome rather than a term in any return.

How to load

import json
import pandas as pd

tiers = ["core", "periphery", "stress", "decode", "controls"]
df = pd.concat([pd.read_parquet(f"{t}/trace.parquet") for t in tiers], ignore_index=True)

state_cols  = [c for c in df.columns if c.startswith("s_")]
action_cols = [c for c in df.columns if c.startswith("a_")]
reward_cols = [c for c in df.columns if c.startswith("r_")]

# the graded composite (deviation D17): utility, safety and the SIZE of the cohort gap, whose sign the
# steering label sets; r_total, the same sum with the gap signed, ships for lineage
target = df["r_util"] + df["r_safe"] + df["r_fid"].abs()

# the posts and the individual votes behind any cell, keyed by cell_id
posts = pd.read_parquet("controls/forum.parquet")
votes = pd.read_parquet("controls/moderation.parquet")

The OPE task

Given a log drawn from the core grid under a known logger, estimate a target policy's value over the 20 core arms, and compare the estimate with the value computed on the same interactions. The score is each estimator's mean absolute error, with an interval, and every comparison between two estimators is a paired difference with its own interval.

Task 1 is a logged bandit drawn from the full grid (deviation D16). In each of 200 repetitions, 10 of the 50 seeds fix the benchmark's choices, the greedy arm that defines the targets and the tilted logger, and no estimator sees them; on the other seeds every context logs ONE arm drawn from the logger, with its propensity recorded exactly, 480 logged interactions a repetition. Every estimator is fit on that log alone: the outcome model is a ridge regression of the logged reward on the context and the arm's three levers, cross-fitted over two halves of the logged seeds for DR and SwitchDR, and the constant floor is the mean logged reward. The truth is the target's full-information value on those same interactions, computed because every arm was run on every interaction, so an estimator's error is its off-policy error and nothing else. Two loggers ship: uniform, the core's design propensity, and tilted toward the fit seeds' better arms (median ratio of largest to smallest propensity 14.2). The targets mix the uniform policy toward the pooled greedy arm at weights 0, 0.25, 0.5, 0.75, 1. The graded composite is r_util + r_safe + |r_fid|, the cohort gap's size in place of the sign the steering sets (deviation D17). Mean absolute error over the repetitions, uniform logger:

shift toward the greedy arm	DM	IPS	SNIPS	DR	SwitchDR	constant
0	0.0033	0.0068	0.0068	0.0034	0.0034	0.0068
0.25	0.0072	0.0729	0.0077	0.0051	0.0082	0.0282
0.5	0.0139	0.1451	0.0102	0.0084	0.0146	0.0560
0.75	0.0207	0.2175	0.0131	0.0119	0.0212	0.0839
1	0.0275	0.2899	0.0167	0.0156	0.0277	0.1117

Tilted logger:

shift toward the greedy arm	DM	IPS	SNIPS	DR	SwitchDR	constant
0	0.0244	0.0369	0.0086	0.0055	0.0054	0.0481
0.25	0.0171	0.0631	0.0082	0.0052	0.0054	0.0205
0.5	0.0099	0.1176	0.0091	0.0065	0.0049	0.0104
0.75	0.0058	0.1766	0.0112	0.0089	0.0063	0.0354
1	0.0079	0.2370	0.0143	0.0116	0.0080	0.0632

Plain importance weighting loses to the constant floor at every shift above zero under both loggers; DR and SwitchDR beat the floor at every one, each paired interval clear of zero; DM and SNIPS do too except at shift 0.5 under the tilted logger, where their error and the floor's cannot be told apart. At full shift under the uniform logger DR minus IPS reads -0.2744 [-0.3070, -0.2433]; under the uniform logger DR's error is below the direct method's at every shift above zero; under the tilted logger the direct method's error falls below DR's from shift 0.75. The same ordering, importance weighting above the floor and DR below importance weighting at every shift above zero under both loggers, holds on 10 of the 10 graded rewards (core/ope_bandit.json).

The positivity variant (core/ope_positivity.json). The same repetitions, drawn by a holed logger: uniform over each context's settings except 5 of the 20, which it never plays, 60 of the 240 (setting, context) pairs, fixed by seed 142 and listed in the file. Every setting was still run on every interaction, so the truth covers the holes and an estimator's error on the target's mass there is measured; that mass is 0.25 of the target at no shift and 0.2517 at full shift. Importance weighting loses to the constant floor at every shift, each paired interval clear of zero: its error is 0.4386 at no shift and 0.4970 at full shift, against 0.0068 and 0.2899 under the uniform logger, because no weight reaches a setting the logger never played. Under the holed logger DR has the lowest error at every shift above zero; at full shift DR reads 0.0172 against 0.0156 uniform, SNIPS 0.0183 and the direct method 0.0337 against 0.0275; DR minus SNIPS at full shift reads -0.0011 [-0.0023, +0.0001], the two cannot be told apart there.

Holed logger:

shift toward the greedy arm	DM	IPS	SNIPS	DR	SwitchDR	constant
0	0.0035	0.4386	0.0071	0.0036	0.0036	0.0071
0.25	0.0103	0.4508	0.0084	0.0063	0.0063	0.0287
0.5	0.0181	0.4635	0.0115	0.0098	0.0186	0.0565
0.75	0.0259	0.4783	0.0148	0.0134	0.0263	0.0843
1	0.0337	0.4970	0.0183	0.0172	0.0339	0.1122

Task 2 on the core (core/task2_core.json, python evaluate.py --root . --task 2). A submission names one of the 20 core settings for each of the 12 contexts, learns from 25 selection seeds and is scored on the other 25 (both lists in the file), where every setting's outcome is known in every context: the violation rate, the share of contexts whose chosen setting breaches the envelope on the evaluation seeds; the value, the mean of Task 1's graded composite; and the hypervolume over utility, safety and the gap's size, each scaled to [0, 1] by the settings' range. Intervals are 95 percent from bootstrap draws of the evaluation seeds. The envelope binds on the core context by context: 30 of 240 (setting, context) pairs breach it over all 50 seeds, every one on the canary floor, in 5 of the 12 contexts.

baseline, chosen on the selection seeds	violation rate	value	hypervolume
uniform	0.117 [0.100, 0.138]	1.7486 [1.7444, 1.7532]	0.232 [0.219, 0.246]
random acceptable	0.014 [0.004, 0.033]	1.7270 [1.7225, 1.7315]	0.242 [0.231, 0.253]
best fixed	0.000 [0.000, 0.000]	1.8552 [1.8444, 1.8661]	0.421 [0.367, 0.473]
per-context greedy	0.083 [0.000, 0.167]	1.9826 [1.9730, 1.9922]	0.472 [0.393, 0.551]
per-context oracle	0.000 [0.000, 0.000]	1.9954 [1.9920, 2.0088]	0.553 [0.480, 0.674]

The per-context greedy choice gains +0.1274 [+0.1180, +0.1367] over the best fixed setting and breaches the envelope in 0.083 of contexts on the evaluation seeds; over 200 random splits it breaches in 0.041 of contexts on average and at most 0.25. The oracle chooses on the evaluation seeds themselves and is a reference, never a baseline.

The policy-learning challenge, cross-fitted. On the graded composite, a per-context policy chosen on half the seeds and scored on the other half beats the pooled best arm chosen and scored the same way: a gain of +0.0474 [+0.0358, +0.0538] from a seed-cluster bootstrap of 1,000 draws, positive in 100.0 percent of them. On the full data the best arm takes 3 distinct values across the 12 contexts. On the all-objectives sum, utility, safety, the gap's size, fairness and readability unweighted, the gain reads +0.0435 [+0.0303, +0.0524].

Learned from the log. With one setting logged per forum, a per-context direct-method learner gains +0.2119 [+0.2103, +0.2135] over the uniform policy on held-out seeds under the uniform logger and +0.2109 [+0.2090, +0.2127] under the tilted one, against +0.2339 for a choice made from the full grid, 91 percent of it. Doubly robust gains +0.2125 and +0.2090; importance weighting per context +0.0733 and +0.0133. 200 repetitions, each scored on 20 held-out seeds (core/policy_baselines.json). The best fixed setting learned from the same log gains +0.1821; the per-context learner beats it by +0.0298 [+0.0277, +0.0320] under the uniform logger and +0.0283 [+0.0262, +0.0304] under the tilted one.

Entry point

python evaluate.py --root .

Run it in the release directory, whose tier folders are the corpus root it needs. evaluate.py runs the grading script's own validation suite first (code/ds4_ope_bandit.py), then Task 1 under both loggers along the shift curve, and prints the tables above. --holed 5 grades Task 1's positivity variant, and --task 2 runs Task 2 on the core (code/ds4_task2_core.py).

Splits

Task 1 draws its own fit seeds and logged seeds in every repetition, from seed 142, so its tables reproduce exactly. splits.json freezes one further division for a consumer who wants a fixed one: a seed split stratified within context, half of each stratum's seeds to train and half to test, with uniform context weighting.

Memorization and privacy

For every generated post, the audit measures the longest run of consecutive words it shares with that post's own author's real posts, against a random-author control. 1,701,130 generated posts from 390 authors across 12,000 core cells were audited. 0 reproduce fifteen or more contiguous tokens from their own author (0 control leaks); by the rule of three a zero count over 1,701,130 trials places the 95 percent upper bound on the per-post leakage rate at 1.8e-06. The mean self-overlap is 2.290 tokens against 2.291 for the control, an excess of -0.001 tokens: a post is no more similar to its own author than to a stranger. Posts shorter than the window and posts whose text has degraded cannot trigger the test, so the audit bounds leakage in longer intelligible generations and says nothing about the rest.

Every user and moderator identifier in the released files is a salted SHA-256 digest cut to 16 hexadecimal characters, and the salt is kept outside the release, so a digest cannot be mapped back to the internal user index or to any account. The export hashed 39,040 identifier columns across 19,520 cells and dropped each moderator's raw decode string, keeping the parsed vote. A digest is stable across cells, so per-agent analyses remain possible: recomputing the per-moderator statistics of the controls and core tiers from the released files reproduces the source values exactly on 12,720 cells.

Scope and disclosures

State these as bounds on the benchmark.

• Single backbone. The trace comes from one frozen Pythia-1.4B backbone with a hypernetwork-emitted per-user adapter. Generalization to other backbones is not established here.
• Single trace topic. Every row shares one moderation topic ("Should communities ban members who post hostile comments?"), so there is no topic column; the topic is held constant on this corpus.
• Synthesized shift. Task 1 logs the core grid under a chosen logger; it is not a deployment log. The generated content is what makes the rewards realistic.
• An answer key conditional on the grid. A target's value on the logged interactions is computed, not estimated, because every arm was run on every interaction; its value beyond these 12 contexts and 50 seeds is a Monte Carlo quantity. Only mu is exact by design, and only in the core tier.
• Levers held fixed on the core, disclosed. The core holds the three decoding levers fixed (a_temperature 0.7, a_mc_dropout 0.0, a_cfg_scale 1.0). The decode tier isolates them against one context, the periphery moves them jointly with the other levers inside its sampled arms, and the stress tier sets a_cfg_scale to 1.0 or 5.0. A permutation importance of zero for them on the core is an artifact of having no core variance, not evidence that they do nothing.
• Caricature personas. The personas are extreme-group cases on the neutral axes, so the effect sizes are upper bounds.

Instrument checks

Text quality, per tier.

Every forum is scored by the text instrument (rq4g1_text_health.py, selftested against fifteen cells with recorded verdicts), per cohort with the worse cohort deciding, and by the independent readability judge, whose readable share on the worse cohort is HEALTHY at 0.90 or above and BROKEN at 0.50 or below; the worse of the two verdicts decides. Counts and percent by tier:

tier	cells	HEALTHY	MARGINAL	BROKEN
core	12,000	11,903 (99.2)	77 (0.6)	20 (0.2)
periphery	6,000	3,732 (62.2)	454 (7.6)	1,814 (30.2)
stress	600	148 (24.7)	24 (4.0)	428 (71.3)
decode	200	187 (93.5)	5 (2.5)	8 (4.0)
controls	720	689 (95.7)	19 (2.6)	12 (1.7)
all	19,520	16,659 (85.3)	579 (3.0)	2,282 (11.7)

The periphery and stress tiers sweep levers into degradation on purpose; their broken share is the design, and the text_health column is what lets a consumer restrict to readable forums.

The generating accelerator.

The corpus was generated on two accelerators under two container builds, a40 and gh200, and the fleet column records which one made every cell. The tiers were divided between them by design: controls and periphery on a40 only; decode and stress on gh200 only; core on both. Pooled over tiers, a comparison would set tiers against each other, so the question is answered inside the tier both accelerators generated.

On the core tier, where 431 of 12,000 cells came from gh200, every quantity was compared between cells of the same arm and context, so the settings each accelerator happened to draw cannot enter the difference (198 strata holding both, a bootstrap over strata at the family-wise level 0.0045, each difference a40 minus gh200). Read against a ruler of a tenth of the steering effect on the cohort gap (0.0272) for every rate and share and a tenth of the pooled level for individuation, no quantity differs materially; r_fid +0.0020 [+0.0001, +0.0040] differs by a nonzero amount inside the ruler; r_coh and canary_caught cannot be placed at this sample. The cohort gap reads +0.0020 [+0.0001, +0.0040], moderation utility reads -0.0015 [-0.0112, +0.0088] and individuation reads -0.0072 [-0.0285, +0.0150]. Compared without matching, r_fid, r_coh and fpr separate; matched, r_fid reads nonzero inside the ruler, r_coh reads inconclusive at this sample and fpr reads flat.

Determinism.

Generation is deterministic for a given accelerator build, measured on repeats the design already contains (ds4_determinism_check.py). The controls tier regenerates 240 core cells under identical settings: 232 of 232 pairs made on the A40 fleet are byte-identical in both generated files, and 0 of 8 pairs split across the two accelerators are. With the per-user offset at zero the clamp has nothing to act on, and 30 of 30 unconditioned groups of the clamp ladder are byte-identical across all 4 clamps, each regenerated by 3 to 4 separate workers. The fleet column records the accelerator behind every cell.

The text channel.

The sentiment scorer first has to see a prose difference known to exist: bad-actor posts read 0.079 to 0.086 lower in polarity than participant posts in every arm and at every steering level. Against that, the empath-minus-rage polarity gap among participants is at most 0.0039 in any arm of the controls tier and 0.0028 at any core steering level. The distribution over the scorer's sentiment bins does not separate the cohorts in any of the six arms or at any core steering level: every excess total-variation interval covers zero. A style reconstruction, correlating fifteen writing statistics re-extracted from each user's generated posts with the same statistics on that user's real history, reads -0.048 to +0.018 across the six arms, every interval covering zero. The prose carries no measurable trace of the cohort label beyond that polarity gap, and none of the individual user (rq4g1_text_channel.py).

A second safety scorer.

The safety objective was rescored by an independent second classifier, unitary/toxic-bert, against the primary s-nlp/roberta_toxicity_classifier, on 19,495 cells. Cell-level r_safe from the two models correlates at +0.873 with a mean offset of -0.0031; the ranking of the 195 lever settings by safety agrees at Spearman +0.825; per-post scores correlate at +0.966 within cells. Verdict at the 0.7 threshold on both: model independent.

Transfer across discussion topics.

Every forum in the corpus is on one question. To measure transfer, the treated setting and its steering-off floor were regenerated on 3 unseen questions at matched contexts and seeds, cell for cell against the core tier. On readable forums the cohort gap on the unseen questions is -0.2565 against -0.2658 in distribution, a retention of 96 percent [95, 98] on 83 against 29 cells (97 percent on every cell, 90 against 30), with the steering-off floor on the same unseen questions at +0.0057 over 90 cells and the in-distribution canary rate 0.633.

License

See LICENSE.md in this repository, which points to the repository-root LICENSE. The upstream Reddit-derived corpus lineage governs redistribution; consult LICENSE.md before reuse. Upstream is Pushshift Reddit 2010 to 2016 (Baumgartner et al., ICWSM 2020), determined NOT IRB-REGULATED RESEARCH by the William and Mary IRB on 2026-09-25 under IRB-2026-354, because the dataset used is public; the signed letter is irb/IRB-2026-354_NHSR_signed_20261005.pdf, transcribed in irb/IRB-2026-354_not_regulated_determination_20260925.md. Cite it as a determination and never as an approval. Not redistributed.