Coding-workload vocabulary — cited proposal v0.1.0

Status: proposal (research maturity hypothesis). This is a defensible vocabulary proposal grounded in the cited corpus as of 2026-08-10. It is not a claim of field-wide consensus, and no term here is a maintained fak contract. Issue #6209, parent #6208; generation stream gen/next.

Machine-readable companion: coding-workload-vocabulary.json (fak.workload-vocabulary.v1). Every identifier below appears there verbatim.

Next action: before reusing a name from this page, check its provenance label and its inclusion/exclusion criteria; if your case fails the exclusion test, it is a different element, not a variant of this one.

What this page is for

The naming risk this addresses is stated in #6209: inventing terms from local intuition either relabels an established concept or fuses distinctions that the literature already keeps apart. So every element carries one of three provenance labels:

Provenance Meaning
borrowed The name and the concept both come from a cited source; we adopt them unchanged.
adapted The concept is cited; the name, the boundary, or the scope is ours.
new-synthesis No cited source names this element; it is proposed here and the citation supports only its ingredients.

Citation verification method and its limit

Each source’s title, authors, venue, year, and identifier were checked on 2026-08-10 against indexed publisher/preprint metadata and abstract text. Direct page fetch was unavailable in this environment (the WebFetch capability is refused at the capability floor), so entries are verified at metadata-and-abstract granularity, not full text. Claims attributed to a source below are restricted to what its abstract or publicly indexed summary states. Where an antipattern maps to a source, the source names the nearest documented class, not necessarily the same term — that mapping is our reading and is labelled accordingly.

The four axes

The central proposal is that these are four independent axes, not four names for the same thing. An instance of work has a value on each. Collapsing any two is the source of the conflations rejected below.

Axis ID The question it answers Example values
Workload shape ax.workload-shape What is the job, and what artifact counts as done? the eleven wp.* patterns
Orchestration topology ax.orchestration-topology How is the work arranged across steps, agents, and turns? single interactive loop; fixed staged pipeline; fan-out/fan-in; tree search; human-in-the-loop
Verification strategy ax.verification-strategy What evidence closes the loop? executable oracle; derived/pinned oracle; relational (metamorphic) oracle; held-out oracle; independent judge; citation resolution; none
Failure mode ax.failure-mode How does it degrade into a plausible wrong answer? the nine ap.* antipatterns

Two sources make the axis split load-bearing rather than decorative. Agentless solves the same issue-resolution shape as an agent loop using a fixed three-phase pipeline of localization, repair, and patch validation — the shape is constant while the topology changes ([s.agentless]). And MAST’s fourteen failure modes are organised into system design, inter-agent misalignment, and task verification — failure classes that recur across workloads rather than belonging to any one of them ([s.mast]).

Patterns — the workload-shape axis

Eleven candidate patterns. A pattern names a goal and its done-artifact, nothing else.

ID Name Definition Aliases Include when Exclude when Provenance Sources
wp.issue-to-patch Issue-to-Patch Repair Turn a reported defect into a change that alters exactly the reported behavior. issue resolution; bug fix; test-suite-based program repair A defect claim exists and the done-artifact is a diff No prior defect claim; or the check itself is the deliverable borrowed [s.swe-bench], [s.apr-bibliography], [s.autocoderover]
wp.spec-to-feature Specification-to-Feature Construction Add an externally visible capability described in prose while preserving existing behavior. feature implementation; NL-to-code The described capability does not exist yet The capability exists but misbehaves (→ wp.issue-to-patch) adapted [s.swe-bench], [s.agentless], [s.grounded-copilot]
wp.behavior-preserving-restructure Behavior-Preserving Restructuring Change internal structure while asserting that observable behavior is unchanged. refactoring; cleanup; modularization The success oracle is “nothing observable changed” Any observable output is intended to change borrowed [s.swe-at-google-lsc], [s.parallel-change]
wp.mechanical-sweep Mechanical Fleet-Wide Sweep Apply one verified rewrite rule across many sites that cannot land as a single atomic change. large-scale change (LSC); codemod; migration sweep The rule is decidable per site and site count dominates the cost One site, or per-site human judgment is required borrowed [s.swe-at-google-lsc]
wp.interface-migration Interface Migration Under Live Consumers Change an interface incompatibly while its consumers keep working throughout. expand-and-contract; parallel change; strangler migration Consumers exist and cannot be updated atomically No consumer outside the change’s own footprint borrowed [s.parallel-change], [s.strangler-fig]
wp.env-remediation Environment and Dependency Remediation Restore a runnable build/test environment when the defect is in the environment, not the program. build fix; dependency bump; “make it run” The failure reproduces with no program logic being wrong Program logic is wrong under one toolchain (→ wp.issue-to-patch) new-synthesis [s.swe-bench-verified], [s.swe-bench]
wp.oracle-construction Oracle Construction Produce the check itself: tests, properties, metamorphic relations, or fixtures. test writing; coverage raise; property authoring The delivered artifact is the check A failing test written as a step toward a fix (→ sp.reproduce-first) adapted [s.oracle-problem], [s.metamorphic-survey], [s.codet]
wp.comprehension-report Comprehension and Explanation Report Answer a question about a system; the deliverable is a cited explanation, not a diff. code archaeology; investigation; “how does X work” No source mutation is required for success A diff is required for success new-synthesis [s.swe-agent], [s.oracle-problem]
wp.triage-and-route Triage and Routing Decide what happens to an item — route, assign, dedupe, defer — without changing the program. bug triage; assignment; prioritization The deliverable is a decision about an item The deliverable is the fix itself borrowed [s.who-should-fix]
wp.change-adjudication Change Adjudication Judge a proposed change and its evidence, returning a verdict and findings. code review; patch-correctness assessment Input is a candidate change plus its evidence The same actor also authored the change adapted [s.apr-overfitting], [s.mast]
wp.perf-under-invariant Performance Optimization Under Invariant Move a measured quantity while a behavior pin holds. optimization; tuning; performance-regression hunt Success requires a measurement against a named baseline Accepted on a plausibility argument with no measurement adapted [s.git-bisect], [s.oracle-problem]

wp.perf-under-invariant is grounded rather than invented: the git bisect manual states the search is not limited to bugs and explicitly gives “the commit that caused a benchmark’s performance to improve” as a target property ([s.git-bisect]) — a measured quantity is treated as a first-class searchable property.

Subpatterns — reusable moves

Twenty-two subpatterns. A subpattern is a move: it recurs inside more than one pattern, and naming it as a workload is the error rejected in cf.subpattern-as-pattern.

ID Name Definition Aliases Include when Exclude when Provenance Sources
sp.reproduce-first Reproduce First Build a deterministic check that fails now and would pass on success, before editing. failing-test-first; repro harness A behavioral claim can be executed The claim is about structure or prose, not behavior borrowed [s.swe-bench], [s.apr-bibliography]
sp.input-minimization Input Minimization Shrink a failing input while the failure is preserved, down to a 1-minimal remainder. delta debugging; test-case reduction; shrinking The failure is input-triggered and re-runnable The failure depends on unreproducible external state borrowed [s.delta-debugging]
sp.history-bisection History Bisection Binary-search revisions for the first commit at which a monotone property flips. git bisect; regression hunt History is navigable and the property is testable per revision The property is not decidable at an arbitrary revision borrowed [s.git-bisect]
sp.spectrum-localization Spectrum Localization Rank suspect code by coverage difference between passing and failing runs. spectrum-based fault localization; suspiciousness ranking A suite with both passing and failing runs exists No executable suite exists borrowed [s.fault-localization-survey], [s.autocoderover]
sp.hierarchical-localization Hierarchical Localization Narrow file → class/function → edit site in fixed stages, with no interactive loop. staged localization; localization funnel The repo is large and each stage is cheap Localization needs feedback from execution borrowed [s.agentless]
sp.structured-code-search Structured Code Search Retrieve context through program structure (AST/symbols) rather than text similarity. AST search; symbol retrieval A parseable program representation is available The artifact is prose or opaque data borrowed [s.autocoderover]
sp.graft-search Graft Search Look for existing in-repo code that already implements the change before writing new code. plastic surgery; donor search; reuse-first The change plausibly resembles code already present The change is genuinely novel to the repository borrowed [s.plastic-surgery]
sp.expand-migrate-contract Expand, Migrate, Contract Add the new interface, move consumers, then delete the old one — never broken in between. parallel change; expand-and-contract Consumers cannot be moved atomically A single atomic change is already safe borrowed [s.parallel-change]
sp.facade-and-drain Facade and Drain Put a facade in front of the old path, redirect call sites incrementally, then remove the old path. strangler fig; gradual legacy replacement The old component must keep serving during replacement The old component can be switched off at once borrowed [s.strangler-fig]
sp.rule-then-sweep Rule Then Sweep Validate the rewrite rule on a small sample, then apply and verify shard by shard. codemod-then-verify; sharded LSC The rule is mechanical and the sites are many Sites need individual judgment borrowed [s.swe-at-google-lsc]
sp.behavior-pinning Behavior Pinning Capture current observable behavior as a check before changing anything. characterization test; golden master; approval test Current behavior is trusted and re-observable Current behavior is the defect adapted [s.oracle-problem]
sp.metamorphic-relation Metamorphic Relation Assert a relation between outputs of related inputs when no expected output is known. metamorphic testing; relational oracle Expected output is unavailable but a relation is known A concrete expected output is available borrowed [s.metamorphic-survey], [s.oracle-problem]
sp.generate-and-filter Generate and Filter Sample many candidates and keep those surviving an executable filter or cross-agreement. sample-and-rank; dual execution agreement Candidates are cheap and a filter is executable Each candidate is expensive or unverifiable borrowed [s.codet]
sp.execute-observe-repair Execute, Observe, Repair Run, read the actual observation, patch, repeat under an explicit step budget. agent loop; reason-act loop; self-debug Execution feedback is available and cheap Feedback is absent, delayed, or systematically misleading borrowed [s.react], [s.swe-agent]
sp.verbal-retrospective Verbal Retrospective Carry a written failure summary into the next attempt instead of re-sampling blind. reflection memory; verbal reinforcement Attempts are repeatable and failures are describable The failure signal cannot be attributed to a step borrowed [s.reflexion]
sp.independent-adjudication Independent Adjudication A verifier that did not author the change judges it, with author-side rationale withheld. blind review; judge panel The change’s correctness is contestable The available check is a total decidable oracle adapted [s.mast], [s.apr-overfitting]
sp.held-out-oracle Held-Out Oracle Reserve a check that the producing step never saw and never optimized against. held-out tests; unseen regression suite A check can be withheld without blocking production Every available check must guide production adapted [s.apr-overfitting], [s.swe-bench-verified]
sp.interface-shaping Interface Shaping Change the tools and affordances the agent acts through, not the prompt or the model. agent-computer interface design; harness design Failures trace to action/observation format Failures trace to missing knowledge borrowed [s.swe-agent]
sp.blast-radius-scoping Blast-Radius Scoping Declare the write footprint before editing so concurrent work provably cannot collide. write lease; lane scoping More than one worker can write the tree A single serialized writer new-synthesis [s.mast], [s.swe-at-google-lsc]
sp.escalation-checkpoint Escalation Checkpoint Hand control to a human at a named decision boundary, not at a step-count limit. mixed-initiative handoff; human-in-the-loop gate The decision’s cost is asymmetric and its outcome uncertain The decision is cheap and reversible borrowed [s.mixed-initiative]
sp.citation-binding Citation Binding Bind every assertion to a resolvable pointer that is checked apart from the assertion. provenance binding; file:line citation The deliverable is prose or a report The deliverable is executable and self-checking new-synthesis [s.oracle-problem], [s.mast]
sp.budgeted-exploration Budgeted Exploration Spend a bounded budget in explicit explore mode before committing to one option. explore-then-accelerate; option sampling The approach is genuinely unknown The next step is already known (accelerate mode) borrowed [s.grounded-copilot]

sp.budgeted-exploration borrows a measured distinction rather than a slogan: Barke et al. report that programmer interaction with a code-generating model is bimodal — acceleration, where the programmer knows what to do next, and exploration, where they do not and use the model to survey options, with exploration being slower, more deliberate, and requiring more validation ([s.grounded-copilot]).

Mechanisms

A mechanism is a substrate capability, not a move and not a goal. It is what a subpattern needs in order to run at all. Listing mechanisms separately is what stops “we have a sandbox” from being mistaken for “we do reproduce-first”.

ID Mechanism Enables
mx.execution-sandbox Run untrusted code and observe its effects sp.reproduce-first, sp.execute-observe-repair, sp.generate-and-filter
mx.test-runner Select and run checks, reporting per-check status sp.reproduce-first, sp.behavior-pinning, sp.held-out-oracle
mx.coverage-spectra Per-test coverage or execution traces sp.spectrum-localization
mx.vcs-history Addressable revisions with checkout and diff sp.history-bisection, sp.rule-then-sweep
mx.program-index AST/symbol index over the repository sp.structured-code-search, sp.hierarchical-localization, sp.graft-search
mx.patch-applier Apply and revert a diff safely sp.rule-then-sweep, sp.execute-observe-repair
mx.write-lease Exclusive claim over a declared file region sp.blast-radius-scoping
mx.step-budget Bounded, observable step and token accounting sp.execute-observe-repair, sp.budgeted-exploration
mx.judge-channel A second model or human that never sees author-side rationale sp.independent-adjudication, sp.escalation-checkpoint
mx.reference-resolver Resolve a citation to its target and compare sp.citation-binding

Antipatterns — the failure-mode axis

Nine antipatterns. Each is stated with a detector, because a failure mode you cannot detect is a mood, not an axis value. Source column names the nearest documented class.

ID Name Definition Aliases Detector Provenance Sources
ap.oracle-overfit Oracle Overfit The change satisfies the checks it was optimized against and breaks untested but intended behavior. patch overfitting; test overfitting A held-out oracle disagrees with the optimized oracle borrowed [s.apr-overfitting]
ap.oracle-laundering Oracle Laundering The gate is weakened, narrowed, or deleted so that it passes. test gaming; check tampering The diff touches the check and the check’s strictness decreased adapted [s.apr-overfitting], [s.mast]
ap.silent-scope-narrowing Silent Scope Narrowing Delivered work covers less than was asked and the report does not say so. quiet descoping A requested acceptance item has neither an artifact nor an explicit deferral adapted [s.mast]
ap.unverified-completion Unverified Completion “Done” is declared with no verification step executed. claim without witness No gate invocation appears in the trace borrowed [s.mast]
ap.premature-termination Premature Termination The run ends before producing the information needed to close the task. early stop Handoff conditions are unmet at exit borrowed [s.mast]
ap.no-progress-oscillation No-Progress Oscillation Equivalent actions repeat without changing observable state. step repetition; thrash loop Repeated action/observation pairs above a threshold borrowed [s.mast]
ap.confabulated-reference Confabulated Reference A cited file, symbol, or source does not exist, or does not support the claim made of it. hallucinated citation The reference resolver fails, or the resolved target contradicts the claim adapted [s.mast], [s.oracle-problem]
ap.benchmark-validity-drift Benchmark Validity Drift The measuring instrument, not the system, explains the score. benchmark artifact; unsolvable items A human re-screen of the items changes the ranking adapted [s.swe-bench-verified]
ap.atomicity-overreach Atomicity Overreach A sweep is attempted as one atomic change at a scale where it cannot land. mega-commit The change cannot pass presubmit as a single unit borrowed [s.swe-at-google-lsc]

MAST supplies the empirical spine here: 1,600+ annotated traces across seven multi-agent frameworks, 14 failure modes in three categories, with inter-annotator agreement reported at kappa = 0.88 ([s.mast]). ap.unverified-completion, ap.premature-termination, and ap.no-progress-oscillation are its “No or Incomplete Verification”, “Premature Termination”, and “Step Repetition” modes under our names.

Rejected conflations

Six conflations that are tempting, common, and wrong. Each gets a counterexample that a reader can check.

cf.shape-is-topology — “multi-agent is a workload pattern”

Counterexample. wp.mechanical-sweep runs as a single scripted pass and as a fan-out of per-shard workers. The goal, the done-artifact, the oracle, and the dominant failure mode (ap.atomicity-overreach) are identical in both. Conversely “fan-out” says nothing about what done means, so it cannot be a workload. Agentless makes the same point from the other side: it solves the issue-resolution shape with a fixed localize/repair/validate pipeline rather than an agent loop ([s.agentless]).

cf.shape-is-verification — “if it has no test, it is not real engineering work”

Counterexample. wp.comprehension-report has no executable oracle by construction; its verification strategy is citation resolution. Meanwhile a fully test-green wp.issue-to-patch result can still be wrong: Smith et al. evaluated repair tools on bugs that each had a human-written patch and found that patches which pass the available tests can overfit them and break untested desired functionality ([s.apr-overfitting]). Test-green is neither necessary nor sufficient, so it cannot define the shape.

cf.subpattern-as-pattern — “bisect is a workload”

Counterexample. sp.history-bisection appears inside wp.issue-to-patch (find the breaking commit), inside wp.perf-under-invariant (find the regressing commit — the git bisect manual gives benchmark movement as a target property), and inside wp.env-remediation (find the breaking dependency bump) ([s.git-bisect]). A move reusable across shapes cannot itself be the shape.

cf.workflow-patterns-transfer — “the Workflow Patterns catalogue already names these”

Counterexample. Van der Aalst et al.’s workflow patterns are control-flow constructs that capture workflow functionality for comparing workflow management systems — they describe how tokens route through a process model ([s.workflow-patterns]). “Parallel Split” describes a mechanical code sweep and a document-translation pipeline identically, because it is silent about what the work is and what makes it correct. Borrowing that catalogue for ax.workload-shape imports a topology name where a goal name is required. It is the right vocabulary for ax.orchestration-topology and the wrong one for ax.workload-shape.

cf.repair-is-restructure — “refactor to fix the bug”

Counterexample. wp.behavior-preserving-restructure asserts that observable behavior is unchanged — its oracle is that the pin still holds. wp.issue-to-patch asserts that exactly one observable behavior changed. A change that does both has no single passing oracle: the pin fails at the fixed site, and the repair witness cannot distinguish “fixed” from “moved”. They must be sequenced, not merged. The expand / migrate / contract discipline exists precisely because the phases are kept separate and the code is never broken between them ([s.parallel-change]).

cf.environment-is-logic — “a red build is a bug in the program”

Counterexample. OpenAI’s SWE-bench Verified screening found development environments that were difficult to set up reliably, causing unit tests to fail regardless of the solution, alongside tests too specific to the original patch and under-specified issue descriptions ([s.swe-bench-verified]). Those are wp.env-remediation and ap.benchmark-validity-drift, not defects in the program under test. Routing them to wp.issue-to-patch produces edits to correct code.

Evidence table

Source What it establishes (per its abstract/summary) What it grounds here
[s.workflow-patterns] A catalogue of workflow control-flow patterns used to compare workflow management systems. ax.orchestration-topology; cf.workflow-patterns-transfer
[s.delta-debugging] Delta Debugging minimizes a failing test case to a 1-minimal failure-inducing input and isolates the passing/failing difference. sp.input-minimization
[s.oracle-problem] Frames the test-oracle problem; surveys specified, derived, implicit, and human oracles, including metamorphic testing. ax.verification-strategy; sp.behavior-pinning; wp.oracle-construction
[s.metamorphic-survey] Correctness can be judged by transforming an input and observing how the output changes, when a concrete expected output is unavailable. sp.metamorphic-relation
[s.fault-localization-survey] Surveys fault-localization techniques that guide developers to fault locations with minimal human intervention. sp.spectrum-localization
[s.apr-bibliography] Structures automatic repair into behavioral and state repair, and catalogues bug oracles and repair operators. wp.issue-to-patch; sp.reproduce-first
[s.apr-overfitting] Patches that pass the available tests can overfit them and break untested but desired functionality. ap.oracle-overfit; sp.held-out-oracle; cf.shape-is-verification
[s.plastic-surgery] Commits are graftable from code already present in the project to a degree largely independent of commit size and type. sp.graft-search
[s.who-should-fix] Bug reports must be triaged and assigned; a learned model reaches 57%/64% precision on Eclipse/Firefox. wp.triage-and-route
[s.swe-bench] 2,294 real GitHub issue/PR tasks; resolution frequently requires coordinated edits across functions, classes, and files. wp.issue-to-patch; wp.spec-to-feature; sp.reproduce-first
[s.swe-bench-verified] 500 human-screened samples; original items suffered under-specified issues, over-specific tests, and unreliable environments. wp.env-remediation; ap.benchmark-validity-drift; cf.environment-is-logic
[s.swe-agent] A purpose-built agent-computer interface materially improves an LM agent’s ability to edit, navigate, and run tests. sp.interface-shaping; sp.execute-observe-repair; wp.comprehension-report
[s.agentless] A three-phase localization → repair → validation pipeline, with hierarchical file → class/function → edit-site narrowing. sp.hierarchical-localization; cf.shape-is-topology
[s.autocoderover] Code search over an AST program representation, sharpened by spectrum-based fault localization when a suite exists. sp.structured-code-search; sp.spectrum-localization
[s.react] Interleaved reasoning traces and actions: thoughts guide action selection, observations inform subsequent reasoning. sp.execute-observe-repair
[s.reflexion] Verbal feedback is summarized into episodic memory and added as context for the next episode, without weight updates. sp.verbal-retrospective
[s.codet] Generated tests plus dual execution agreement select among sampled solutions, raising HumanEval pass@1 to 65.8%. sp.generate-and-filter; wp.oracle-construction
[s.mast] 14 failure modes over 1,600+ traces in three categories: system design, inter-agent misalignment, task verification (kappa = 0.88). ax.failure-mode; six ap.* entries
[s.mixed-initiative] Principles for coupling automated services with direct manipulation rather than choosing between them. sp.escalation-checkpoint
[s.grounded-copilot] Interaction is bimodal — acceleration (knows the next step) versus exploration (surveys options, validates more). sp.budgeted-exploration
[s.swe-at-google-lsc] An LSC is a logically related change set that cannot practically be submitted atomically; max atomic size falls as scale grows. wp.mechanical-sweep; ap.atomicity-overreach; sp.rule-then-sweep
[s.parallel-change] Expand, migrate, contract: keep the old design while adding the new, so the code is never broken. wp.interface-migration; sp.expand-migrate-contract; cf.repair-is-restructure
[s.strangler-fig] Gradual replacement: wrap the old system, redirect incrementally, retire the host. sp.facade-and-drain; wp.interface-migration
[s.git-bisect] Binary search over revisions for any property change, explicitly including benchmark-performance movement. sp.history-bisection; wp.perf-under-invariant

Sources

All entries accessed 2026-08-10. Verification granularity is metadata and abstract, per the limit stated above.

Twenty-four sources; the definition of done required at least twelve.

Machine-readable companion and its determinism witness

coding-workload-vocabulary.json carries every axis, pattern, subpattern, mechanism, antipattern, conflation, and source above under schema fak.workload-vocabulary.v1. There is no report generator to run twice; instead the artifact is canonical, so re-serialization is a fixed point and any editor that round-trips it must produce the same bytes:

python3 - <<'PY'
import json
p = "docs/research/coding-workload-vocabulary.json"
raw = open(p, encoding="utf-8").read()
canon = json.dumps(json.loads(raw), sort_keys=True, indent=2, ensure_ascii=False) + "\n"
print("canonical:", raw.replace("\r\n", "\n") == canon)
PY

Counts are asserted by the same artifact (counts block) and match this page: 4 axes, 11 patterns, 22 subpatterns, 10 mechanisms, 9 antipatterns, 6 conflations, 24 sources.

Promotion, demotion, and invalidating assumptions

Promotion evidence — what would move this toward gen/now. Two independent annotators classify a sample of this repository’s own trajectories using only this page’s inclusion/exclusion criteria and reach substantial agreement on ax.workload-shape without consulting each other, and at least one downstream consumer (a dispatch classifier, a report schema, or an issue template) reads the wp.*/sp.* identifiers rather than re-deriving its own names. Vocabulary that nothing consumes stays a hypothesis.

Demotion or retirement evidence. Retire an element when annotators cannot separate it from a neighbour without inventing a criterion absent from this page; demote the whole proposal if a survey published after 2026-08-10 supplies an equivalent axis split with broader evidence, in which case borrow it and keep only the new-synthesis entries that survive.

Invalidating assumptions.

  1. The axes are assumed independent. If, across real trajectories, workload shape predicts orchestration topology strongly enough that topology carries no additional information, then ax.orchestration-topology is a derived attribute and the four-axis claim is wrong.
  2. The four new-synthesis elements are assumed genuinely unnamed. wp.env-remediation, wp.comprehension-report, sp.blast-radius-scoping, and sp.citation-binding rest on a corpus verified at abstract granularity by one worker on one day; a full-text sweep, or a body of literature outside the surveyed streams, could show any of them is a rename.
  3. Sources are assumed to say what their abstracts say. Verification did not reach full text (WebFetch is refused at this host’s capability floor). Any attribution above that the full paper contradicts invalidates that row of the evidence table.