COSS: Commission, Omission, and Source Selection

• COSS is a framework to detect news bias by jointly analyzing content inclusion (commission), omitted evidence (omission), and sourced materials (source selection).
• It employs a pipeline of candidate retrieval, text alignment, and directed graph construction to systematically reveal patterns of bias.
• The methodology extends to diverse domains, offering insights for news analysis, LLM behavior, event-stream modeling, and OSS evidence evaluation.

Searching arXiv for the cited COSS-related papers and adjacent literature to ground the article. Searching arXiv for “What's in the News? Towards Identification of Bias by Commission, Omission, and Source Selection (COSS)” and related papers. Commission, omission, and source selection (COSS) denotes a three-part analytic frame for identifying bias in news by examining what is explicitly included, emphasized, or reworded; what is left out; and which sources are chosen as inputs to reporting. In the formulation introduced for news analysis, COSS is treated as a joint three-fold objective rather than as three isolated tasks, with a pipeline that retrieves event-related articles, aligns reused or paraphrased text, builds a directed graph of text reuse over time, analyzes patterns of reuse and exclusion, and visualizes the resulting information flow (Zhukova et al., 4 Aug 2025). Related work extends the commission–omission distinction into other domains, including long-context LLM agents, event-stream modeling, and open-source software selection, but these extensions do not all preserve the full triad or define a single unified cross-domain taxonomy (Gamage, 22 Apr 2026).

1. Concept and scope

In the news-bias setting, COSS refers to three related forms of bias. Bias by commission is bias caused by what is explicitly included, emphasized, or reworded in a news article. The paper associates it with reuse of information from another source, paraphrase with different wording, loaded language, and changes in framing or polarity. Bias by omission is bias caused by what is left out, especially when a later article selectively excludes parts of an event, source article, or narrative. Source selection bias concerns which sources a news outlet chooses to rely on, including whether reuse is concentrated on outlets with a specific political orientation or on a narrow source set (Zhukova et al., 4 Aug 2025).

The same terminology appears in other research areas, but with different semantics. In long-context LLM agents, the closest equivalent to a COSS taxonomy is an explicit split between commission constraints, defined as rules that require the model to actively produce a specific token, string, or formatting feature, and omission constraints, defined as rules that require the model to suppress a default behavior. That work also states that it does not formalize a separate source-prioritization metric or taxonomy, although it argues that a policy document at turn 2 loses effective attention weight relative to more recent conversational tokens and injected schemas as context grows (Gamage, 22 Apr 2026). In temporal point process modeling, commission and omission denote contamination types in event streams: unexpected occurrences versus unexpected absences or delays of events (Zhang et al., 23 Jan 2025). In open-source software integration, the provided material relates COSS to component selection by framing source selection as the choice of OSS portals and evidence channels used to assess candidate projects, while warning that important signals may be omitted when selection depends too heavily on automatically retrievable metrics (Li et al., 2021).

This suggests that COSS is best understood as a family of distinctions centered on inclusion, exclusion, and evidentiary choice, with its most explicit three-part formulation currently appearing in news-bias identification.

2. COSS as a joint methodology for news-bias identification

The news-bias paper proposes automated identification of bias by commission, omission, and source selection as a joint three-fold objective, explicitly contrasting this with previous work that addressed these bias types separately. Its central assumption is that COSS bias can be inferred by finding related news articles about the same event, aligning text across them at paragraph or sentence level, detecting reuse and paraphrase, building a directed reuse graph over time, analyzing patterns of reuse, omission, and source choice, and visualizing the result for interpretation (Zhukova et al., 4 Aug 2025).

The described pipeline has five stages: candidate retrieval, source retrieval and text alignment, construction of a graph of text reuse, pattern analysis, and visualization. Candidate retrieval takes either an event-descriptive query plus a time frame or a seed document with its timestamp, and returns event-related candidate articles. The system searches large news corpora or databases such as LexisNexis, CommonCrawl, MediaCloud, and GDELT. In a closed evaluation environment, candidate retrieval can instead accept a pre-provided set of related articles with required metadata. Each article receives a polarity label induced from its outlet, such as Fox News mapped to “R” for right or conservative (Zhukova et al., 4 Aug 2025).

Source retrieval and text alignment constitute the core step. The method distinguishes itself from approaches limited to exact copy-paste or word-permutation detection, and instead targets paraphrased sentences or paragraphs and semantic similarity at higher granularity, especially paragraph level. The output is a set of aligned text units that separate original paragraphs, reused paragraphs, and candidate source links for reused segments. A subsequent polarity classification or relabeling stage checks whether outlet-based polarity labels match the textual content of each paragraph or article. This is applied to both original and reused paragraphs so that one can inspect how the same message evolves over time and across outlets. The paper notes that training a reliable classifier requires a large balanced dataset because biased language varies substantially (Zhukova et al., 4 Aug 2025).

The reuse graph is directed because time is known. It stores articles, paragraphs, extracted attributes, and assigned attributes, while edges encode semantic reuse between paragraphs with weights reflecting similarity. That structure supports both article-centric analysis, focused on one seed article and what is original versus reused, and collection-centric analysis, focused on event-level information flow across multiple articles. Within that graph, omission becomes inferable through absence under comparison, commission through reframing or polarity change under reuse, and source selection through the distribution and diversity of source outlets (Zhukova et al., 4 Aug 2025).

3. Signals, notation, and operational criteria

The formal apparatus in the news-bias proposal is relatively light, but several notions are explicit. Articles are assigned outlet-based polarity labels such as L for left, C for center, and R for right. The text-reuse graph is directed by timestamp ordering, so earlier articles can serve as sources for later ones. Relations between paragraphs encode the strength of semantic similarity, even though no explicit similarity formula is printed. The method also conceptually separates original text in an article, reused text from another article, and the original source article for reused text (Zhukova et al., 4 Aug 2025).

The example in Figure 1 is central to the paper’s operational interpretation. It describes an article $A_6$ with left polarity $L$, a paragraph $P_3$ that is original and labeled center-oriented $C$, and a paragraph $P_4$ reused from article $A_1$ and changing from center to left. This is used to illustrate polarity shift under reuse and thereby bias by commission. For commission bias more generally, the paper identifies paraphrased reuse, loaded wording compared to the source, polarity change between source and reused text, repeated reuse of politically aligned content, and longevity of reuse over time as relevant indicators. For omission bias, the indicators include source paragraphs present in an origin article but missing in descendant articles, selective coverage of event details, systematic exclusion of thematic aspects, and differences in coverage across outlets with different slants. For source selection bias, the analysis examines which outlets are used as sources, how many sources are reused, the source polarity distribution, whether the article mostly depends on neutral versus slanted sources, and the diversity or concentration of sources (Zhukova et al., 4 Aug 2025).

A recurring methodological point is that omission is difficult because it concerns missing content rather than directly observable text. The paper’s solution is comparative: absence becomes visible through reuse-graph structure. If a descendant article reuses only certain parts of a source while excluding others, that selective exclusion may indicate omission bias. This does not amount to a claim that every absence is bias; rather, the pipeline is designed to expose systematic asymmetries in what survives transmission across articles (Zhukova et al., 4 Aug 2025).

4. Commission and omission in long-context LLM agents

A distinct but closely related operationalization appears in the study of long-context LLM agents, which examines operator-defined behavioral policies embedded in system prompts. That work defines commission constraints as rules that require the model to actively produce a specific token, string, or formatting feature, and omission constraints as rules that require the model to suppress a default behavior. Its concrete examples include commission constraints such as ending every response with [AUDIT-OK], beginning every response with STATUS:, or including INC-08453 in every response, and omission constraints such as never using “however,” never using markdown headers, never using triple-backtick code blocks, never using bullet points, and never using first person (Gamage, 22 Apr 2026).

The paper’s main thesis is that long-context agents exhibit an asymmetry called Security-Recall Divergence (SRD): omission constraints decay with depth, while commission constraints persist. In a 4,416-trial three-arm causal study across 12 models and 8 providers at six conversation depths, omission compliance for Mistral Large 3 on the “no bullet points” constraint falls from 73% at turn 5 to 33% at turn 16, while commission compliance remains 100%, with a CMH test result of $\chi^2 = 147$, $p < 10^{-33}$. The six injection depths are 5, 10, 13, 16, 20, and 25. The three study arms are a no-dilution control, a schema-dilution condition that adds 20 semantically complex cloud-infrastructure tool schemas with about 10,000 tokens of overhead, and a token-matched padding control that replaces cloud schemas with semantically neutral “monitoring_probe” schemas of similar length and structure (Gamage, 22 Apr 2026).

The causal interpretation is that semantic content in the injected schemas, not just token volume, drives much of the dilution effect. In the two models with token-matched padding controls—Gemini 2.5 Flash and Llama 3.3 70B—the paper reports that schema semantic content accounts for 62–100% of the dilution effect. It also reports a regression result with token volume coefficient $\hat\beta = +0.19$, $p = 3.4 \times 10^{-8}$, and turn-depth coefficient $L$0, which supports a more nuanced statement: token volume contributes, but semantic load is the dominant causal factor in the padding-controlled models (Gamage, 22 Apr 2026).

This work is relevant to COSS because it shows that instruction-following is not uniform across instruction types and that visible compliance can be misleading. The paper defines a Zone of Exploitation as the interval where omission compliance has dropped below baseline but commission compliance remains high, and it argues that monitoring commission-style audit outputs is insufficient because omission-style security policies may already be failing. It defines Safe Turn Depth (STD) as the turn depth where omission compliance in the dilution arm crosses 50%, estimated by linear interpolation: $L$1 with reported examples including Mistral Large 3 at STD $L$2 turns $L$3, Qwen 3.5 at STD $L$4 turns $L$5, Nemotron 120B reported as $L$6, and Gemma 4 31B with no decay and undefined STD. The proposed mitigation is runtime rather than model-internal: re-inject constraints before the STD threshold and cap session length using an STD-derived safe token budget, thereby restoring compliance without retraining (Gamage, 22 Apr 2026).

5. Extensions beyond news: event outliers and OSS selection

In temporal point process research, commission and omission are defined as contamination modes in event streams. An omission outlier is an event that should have happened but is missing or delayed; a commission outlier is an unexpected event that happens, either too early or when it should not occur. The paper models $L$7 observed event sequences

$L$8

with event times on $L$9, counting process

$P_3$0

and conditional intensity

$P_3$1

Within contaminated sub-intervals, omission is represented by intensity depletion,

$P_3$2

and commission by intensity inflation,

$P_3$3

The proposed solution is a weighted temporal point process likelihood in which each inter-event interval is downweighted if it appears abnormal under time rescaling, so that likely contaminated intervals contribute less to estimation and clustering (Zhang et al., 23 Jan 2025).

The weighting mechanism is built around the fact that under the correct model,

$P_3$4

The class-specific interval weight is

$P_3$5

with overall interval weight aggregated across classes either by a maximum or by class-membership probabilities. The paper states four merits—adaptivity, robustness, local convergence, and outlier detection power—and compares the weighted method against a vanilla estimator using clustering purity as the evaluation metric. In simulations with non-homogeneous Poisson and Hawkes working models across $P_3$6, $P_3$7, and $P_3$8, the weighted method consistently outperforms the vanilla baseline, especially under Type-ii commission contamination. On IPTV viewing histories, it yields smaller median $P_3$9 error and higher comparison rate, and its learned weights highlight holiday-related anomalies around the Chinese Spring Festival and National Day (Zhang et al., 23 Jan 2025).

In OSS component selection, the cited empirical study is not a COSS paper by name, but the provided material interprets it as relevant to source selection because it investigates which factors and metrics practitioners use to select OSS for integration, what sources of information and portals can be used to assess those factors, and how much of that process can be automated (Li et al., 2021). The study reports 8 main factors, 74 sub-factors, and 170 related metrics, after dropping the Economic factor from an earlier set of 9 because it ranked below 3 in all interviews. The elicitation uses a 0-to-5 scale, and factors with Likert score at least 3 trigger deeper collection of sub-factors and measures. It also analyzes APIs and portals for the most starred 100K GitHub projects and concludes that only a small part of the factors can be evaluated automatically, with only 40 of 170 metrics available and only 22 returning information for all the 100K projects (Li et al., 2021).

A plausible implication is that this study illustrates an omission problem at the level of evidence acquisition: automation can support screening, but important factors such as community reputation and community sustainability may remain outside portal APIs. The provided material therefore characterizes the practical recommendation as a hybrid process combining automated evidence with manual judgment (Li et al., 2021).

6. Interpretation, limitations, and recurrent misconceptions

A common misconception is that commission and omission are symmetric. The long-context LLM study directly argues the opposite: commission-type audit signals can remain healthy while omission constraints have already failed, which makes the failure invisible to standard monitoring. Its asymmetry claim is formalized as

$C$0

where $C$1 is commission compliance and $C$2 is omission compliance (Gamage, 22 Apr 2026). In the news-bias setting, an analogous misconception would be to treat omission as merely the absence of evidence. The COSS proposal instead treats omission as something inferable only through structured comparison across related articles and reuse paths (Zhukova et al., 4 Aug 2025).

Another misconception is that source selection in COSS is equivalent to a formal source-ranking theory. The news-bias paper operationalizes source selection through source polarity, source diversity, and reuse concentration in a directed reuse graph. The LLM-agent paper, by contrast, explicitly does not define a formal source-prioritization metric or taxonomy, even though it argues that policy documents lose effective attention weight relative to more recent conversational material. The OSS-selection study uses “source” in yet another sense: portals, aggregators, repositories, and mining tools as evidence providers for adoption decisions. These are related but not identical notions of source selection (Zhukova et al., 4 Aug 2025).

The empirical status of the literature also varies. The news-bias COSS paper is primarily a conceptual and methodological proposal and does not report a full dataset or metric benchmark for the joint pipeline itself; the only numerical performance figures in the provided material, such as $C$3 for a reference-entity omission approach and $C$4 for a TF-IDF baseline, are explicitly attributed to earlier omission work rather than to the new joint pipeline (Zhukova et al., 4 Aug 2025). By contrast, the LLM-agent study reports a 4,416-trial causal design with model- and depth-specific results, and the temporal point process study provides both theoretical guarantees and synthetic and real-data experiments (Gamage, 22 Apr 2026).

Taken together, the cited work supports a restrained but technically specific conclusion. COSS is most clearly formalized as a joint framework for identifying news bias through commission, omission, and source selection; adjacent literatures show that commission–omission asymmetries recur in other systems and that source choice often determines what evidence becomes visible. This suggests that COSS is not merely a taxonomy of content errors, but a comparative framework for analyzing how inclusion, exclusion, and evidentiary dependence jointly shape observed outputs (Zhukova et al., 4 Aug 2025).