Source-linked AI summary
From Human Corrections to Learned Segmentations: A Critical Mathematical Review of Information, Learning, and Human Effort
Jon Sporring
TL;DR
The review asks when corrective feedback is informative and what establishes better segmentation at matched human effort. It compares selection, updates, and evaluation mathematically, finding that response protocols, assimilation, and effort units shape what empirical comparisons establish. It identifies open tests for selection, assimilation, and interaction policies.
Problem
The review asks when corrective feedback becomes useful learning and what establishes improved segmentation per unit of human effort.
Source: §1 · p. 2Method
It compares selection, learner updates, and evaluation through a mathematical account of response information, predictive error, and human work.
Source: §6 · p. 18, §8.2 · p. 21, §10 · p. 25Results
Response protocols affect measured performance, assimilation matters, and fewer annotations do not establish less human effort or reusable-model performance.
Source: §10 · p. 25Takeaways & Limitations
The open question is whether modelling selection and sequencing improves assimilation and segmentation at matched effort beyond addressed labels and fitting resources.
Source: §10 · p. 25Observed marking frequency alone cannot identify both error prevalence and the probability of detecting and marking an error.
Source: §9.4 · p. 24
Abstract
from arXiv · showhide
When does a human correction become useful learning, and what establishes better segmentation at matched human effort? This critical narrative review follows feedback through its selection, learner update, and evaluation. It compares active learning, machine teaching, interactive segmentation, and sequential predictive evaluation using a common mathematical account of target information, predictive error, response predictability, and human work. An exact Bayesian update converts evaluation-label information into expected log-loss improvement; an excess-risk decomposition explains what an approximate learner can lose. Conditional information decompositions separate the roles of selected locations, labels, and displayed context. Empirical comparisons show why response protocols, update mechanisms, and effort denominators matter, distinguishing current-mask refinement from reusable learning and simulated interaction from human evidence. We identify what logs can reveal, what requires independent evaluation, and what remains model-dependent. The resulting research priorities are to test what selection reveals, whether using it improves assimilation, and whether the resulting interaction policy benefits segmentation at matched effort. These are distinct claims: predicting a teacher's responses need not identify human intentions or establish better learning. No new acquisition algorithm or empirical superiority is claimed.
1 When do corrections become useful learning?
The review asks when corrective feedback is informative, becomes useful learning, and establishes improvement per human effort. It compares selection, learner updates, and evaluation without treating distinct methods as approximations to one objective.
1 When do corrections become useful learning?
- A correction may refine the current mask, train a reusable model, or do both, as RootPainter illustrates [Smith et al., 2022a, Smith et al., 2022b].Success at current-image refinement does not establish success for reusable predictors.Source: §1 · p. 2
- Query selection, pedagogical example selection, mask refinement, and sequential compression control different decisions and define success differently.The review compares their assumptions using a common mathematical description rather than treating them as approximations to one objective.Source: §1 · p. 2
- Its three questions concern who selects supervision and under what observation model, which objective is pursued, and what assumptions connect corrections to lasting learning.The review uses progressive teaching to describe this coupled process, not to propose an acquisition algorithm or claim that humans teach optimally.Source: §1 · p. 2
- The review separates evidence in a selected response, the effect of a learner update, and the comparison needed to call that effect efficient.A teacher-response model is useful only insofar as it explains or improves a specified outcome.Source: §1 · p. 2
Notation at a glance
This section previews notation used in the review, distinguishing random objects from realized values and specifying base-two logarithms.
Notation at a glance
- The notation distinguishes random objects from their realized values, such as R_i and r_i.All information quantities use base-two logarithms unless stated otherwise.Source: Notation at a glance · p. 2
- Table 1 is a quick reference for core symbols, meanings, and relations.The boxed figures, rather than the table, develop a four-pixel example.Source: Table 1 · pp. 2–3, Notation at a glance · p. 2
- Measured human cost is denoted C_n, with W_n as its mark-count specialization.The review also defines ideal response length ℓ_i from the response probability and cumulative expected length ℓ_PT(n).Source: Notation at a glance · p. 3
2 Related perspectives: who, how, and what makes it efficient?
This critical narrative review selects studies by their observation, update, and evaluation mechanisms to connect assumptions about selected feedback, learner updates, predictive outcomes, and human effort. It distinguishes target information, individual responses, learner improvement, and efficient teaching to motivate a shared formal language rather than a replacement algorithm.
2.2 Relation to existing reviews
- The review extends prior human–learner-loop and active-learning reviews by asking who selects spatial support, what makes it informative, and whether gains concern current masks or reusable learners.It makes explicit the assumptions connecting observation, update, and effort; similar interfaces may optimize different quantities.Source: §2.2 · p. 4
2.3 Who selects supervision, and how do the participants interact?
- Corrective segmentation distributes choices among system and teacher, while teacher-aware learning can fail when assumed teaching behavior mismatches reality or sparse corrections lack labelled alternatives.Inferring information from a teacher’s choices is not itself novel; the application-specific question jointly concerns displayed predictions, selected supports, labels, updates, and human work.Source: §2.3 · p. 4
2.4 What changes when a segmentation is corrected?
- Interactive correction spans mask refinement, image-specific adaptation, prompt reuse, and reusable-model training, which must not be conflated; RClicks models spatial responses, whereas simulated click trajectories do not validate human behavior.MedSAM measures live refinement rather than reusable learning, while SPA reuses preference state; these are distinct from natural spatial-support selection.Source: §2.4 · p. 5
2.5 What does the empirical evidence establish?
- The response model is supported by 475,544 real clicks; its multi-click simulations required approximately 5–29% more clicks than conventional simulated clicking to reach 90% IoU, capped at 20 clicks.Source: §2.5 · p. 7
- Microscopy SAM annotators averaged 0.7 seconds/object for both models, while test mSA was 0.51 ± 0.02 versus 0.42 ± 0.05; CEREALS reached 95% of full-data mIoU at 17.07% versus 33.76% of recorded cost.The microscopy annotation and test rankings differ, and CEREALS uses a robot oracle and recorded polygon-click cost rather than measured human time.Source: §2.5 · p. 7
- Selection, assimilation, and efficiency are distinct: changing correction use alters the update, while response prediction or lower mark counts alone do not establish better reusable learning at matched human effort.RClicks and CEREALS test different parts of the chain; RootPainter and RootPainter3D savings depend on training histories, and late or optimized-model savings are not whole-session savings.Source: §2.5 · p. 7
2.6 Which labels should the learner request? / 2.7 Who should annotate, and at what cost?
- Acquisition criteria differ in target and assumptions: BALD discounts within-model label variability, batch scoring needs a joint-label model, and cost-aware selection or work proxies do not generally establish optimal multi-step effort savings.XpertLoop’s size–entropy selection can outperform random selection in fewer loops, but plain entropy can underperform, and loop counts omit total cost; click counts, inspection, and changing correction difficulty matter.Source: §2.6 · p. 8, §2.7 · p. 9
2.8 What else must a correction model distinguish?
- Correction logs filter exposure, attention, error detection, and marking, so non-correction is not automatically acceptance and adaptive logs alone do not provide an unbiased evaluation sample.Ignoring the observation mechanism requires assumptions; displaying an image does not reveal human attention or marking probabilities.Source: §2.8 · p. 9
2.10 Synthesis: which part of the corrective loop is modelled?
- Table 3 compares modeled mechanisms rather than benchmark rankings, separating controlled actions, observations, and choices left outside the model.Source: §2.10 · p. 10, Table 3 · p. 10
- Three distinctions organize the review: a target is not an individual response, response information is not learner improvement, and predictable supervision is not necessarily efficient teaching.Source: §2.10 · p. 10
- The formalism provides shared language for these distinctions without replacing surveyed algorithms, and Section 9 revisits the literature through objective comparison and open problems.Source: §2.10 · p. 10
3 Targets, predictions, and the corrective interaction
The formalism separates a distribution-valued target from individual human responses and learner predictions, then records the interaction as an ordered sequence of displays, responses, and learner states. Its core scope is pointwise prediction on a fixed collection, with actual correction addresses distinct from evaluation addresses.
3.1 A fixed collection and a distribution-valued target
- The target assigns a label distribution at each address, with deterministic segmentation as the special case of point-mass distributions.Across sessions, the target distribution may itself vary; evaluation addresses are drawn independently using ν, whereas correction addresses are interaction-selected.Source: §3.1 · p. 11
- The formal core treats pointwise prediction on a fixed collection; random views within that collection are represented, but other applications may require extensions.The learner’s predictive distribution, target, source distribution, and beliefs over candidate models are distinct objects.Source: §3.1 · p. 11, Table 4 · p. 11
3.2 Pre-response and post-response memory
- Each round records the learner’s prior memory, selected view, displayed context, and response in order, preserving what was shown separately from what was corrected.An empty response may be recorded; timestamps, erasures, and navigation require their own action records when relevant.Source: §3.2 · p. 12
- The response model describes where and how a teacher responds to a display, while the learner update produces a declared next checkpoint or terminal readout.If teacher navigation or overrides convey target information, they must be logged as an earlier teacher message or retained as context information.Source: §3.2 · p. 12
4 Evaluation-equivalence and predictive distortion
Evaluation-equivalence groups targets that agree at addresses weighted by ν, while predictive distortion measures a learner’s error under those weights. Changing evaluation weights can change both the equivalence classes and the measured improvement from the same correction.
4 Evaluation-equivalence and predictive distortion
- Targets are evaluation-equivalent when their conditional label distributions agree wherever ν assigns weight; changing ν can change the target class.Changing the learner family does not change that class.Source: §4 · p. 12
- Predictive distortion is the evaluation-weighted discrepancy from the target, distinct from the label-variability floor and from optimization or finite-data error.For log loss, the fixed-target cross-entropy risk separates into the target entropy and predictive distortion.Source: §4 · pp. 12–13
- Changing only evaluation weights changes both the equivalence class and the measured improvement from the same correction.Figure 2 applies these definitions to the first correction in Figure 1.Source: §4 · p. 13, Figure 2 · p. 14
5 Information gain, expected improvement, and their differences
The review distinguishes uncertainty about evaluation-relevant target classes from predictive improvement and from fixed-target response gains. Under a shared source model, evaluation-label information equals the expected log-loss benefit of a Bayes update, but an approximate learner’s excess risk can reduce or reverse that benefit.
5.1 What is uncertain, and under which distribution?
- The information account assumes a discrete evaluation-equivalence class with finite entropy; its probabilities come from the source and interaction model, not automatically from learner beliefs.A realized response can broaden the posterior, and full-round gain includes information from the view and display as well as the response.Source: §5.1 · p. 14
5.2 Fixed-target expectations versus source averages
- For a fixed target, expected response gain is not generally mutual information or nonnegative; mutual information arises when targets and responses are averaged under the same joint source model.The fixed-target log-loss change has no general equality with target information gain.Source: §5.2 · pp. 14–15
5.3 When information improves prediction
- The Bayes update’s expected log-loss improvement equals I(Y; Ri | X, Zi = z), which is nonnegative under the stated joint source model.The equality is specific to log loss; other losses require their corresponding decision model.Source: §5.3 · p. 15
- An actual learner improves on average precisely when predictive information exceeds its increase in excess risk; fitting corrections poorly can forgo the benefit.Reducing existing excess risk can add to the information benefit, but the identity does not guarantee improvement for each target or response.Source: §5.3 · p. 15
- Equal information need not yield equal distortion reduction when evaluation weights differ, and learner beliefs or updates that differ from the source posterior can also separate information from predictive benefit.These comparisons qualify the proportionality in Figure 3, which averages over the next view and correction under a history-insensitive response policy.Source: §5.3 · pp. 15–17, Figure 3 · p. 16, Figure 4 · p. 16
5.4 When does this perspective recover BALD?
- BALD’s candidate-model information matches response information about evaluation classes only under calibrated beliefs, a matching label-response model, and class-dependent label probabilities.Information about candidate parameters outside the support of ν need not be evaluation-relevant; Figure 6 also changes available actions.Source: §5.4 · p. 17, Figure 6 · p. 18
6 Human work and teaching objectives
Human work cannot be represented by marked locations alone when interaction policies change inspection, navigation, or time. Comparisons at matched effort require declared policies, resource units, and comparable experimental conditions.
6 Human work and teaching objectives
- Marked-location work counts each round’s response support, including repeated marks, but does not measure retracing, inspection, navigation, or time.Human effort should specify a declared unit; mark counts and active or elapsed seconds measure different resources.Source: §6 · p. 18
- A policy comparison at a common budget needs a shared action set, initialization, evaluation distribution, and declared training resource, or must vary these explicitly.Equal mark counts or an “information per click” ratio do not remove those alternatives.Source: §6 · p. 18
- Target information and log-loss distortion are measured in bits, while work is measured in marks or separately in seconds; local gain-per-work ratios are not generally optimal multi-round policies.A ratio of expectations differs from an expectation of a ratio, particularly when responses can be empty.Source: §6 · p. 18
7 Prequential evaluation: what response predictability establishes
Prequential evaluation forecasts each response before observing it, using only earlier decoded observations. Its fixed-stream response-prediction scope is distinct from measuring useful supervision or human effort.
7 Prequential evaluation: what response predictability establishes
- Prequential coding predicts and encodes each observation using only already decoded observations, instantiating sequential forecast evaluation with logarithmic scoring [Rissanen, 1978, Grünwald, 2007, Grünwald and Roos, 2020].Two-part MDL instead describes a model and then data conditional on that model.Source: §7 · p. 19
- For well-specified i.i.d. data in a fixed regular k-parameter family, suitable codes have leading redundancy (k/2) log2 N relative to fitted likelihood.The stated regularity conditions include an interior parameter, nonsingular Fisher information, and concentration.Source: §7 · p. 19
- This benchmark is not an arbitrary neural-training guarantee or a bound on human effort, and adaptive responses do not justify replacing N with rounds or marks.For continuous parameters, a two-part code must specify finite precision and a decodable parameter description.Source: §7 · p. 19
8 Information in adaptive response streams
The section separates information in an adaptive correction stream from response variability and coding-model error, then asks when response codes reflect full-memory gain. It further distinguishes information in selected addresses from information in labels and from the work required to mark locations.
8 Information in adaptive response streams
- Expected response-code length decomposes into target information, residual response variability, and coder mismatch; the code covers the correction process, including support selection.Long codes can reflect any of these components, while short codes need not indicate useful teaching.Source: §8 · pp. 19–20
8.1 A measurable comparison on the same interaction stream
- A normalized causal response model compares forecasts on the same recorded stream, with each forecast fixed before its response; description costs are excluded.Figure 7 makes same-stream scoring operational without estimating the teacher’s target or information gain.Source: Figure 7 · p. 20, §8.1 · p. 20
8.2 Which information belongs to the response?
- Full-memory information gain includes both context information and response information; response length bounds the gain only when pre-response context supplies no fresh target information.Teacher navigation may violate that condition and must be coded as a message or accounted for as context information.Source: §8.2 · p. 21
8.3 Support, labels, and work
- Work counts marked locations, whereas response coding weights events by predictability; marked pixels are not independent sample counts.Labels may be dependent, and coding them sequentially does not require retraining after each label.Source: §8.3 · p. 21
- The support-first decomposition assigns information to selected addresses and then labels, but is order-dependent and not a causal attribution.The four-pixel example shows that a selected address can distinguish targets even when both targets agree on the displayed labels.Source: §8.3 · pp. 21–22, Figure 8 · p. 22
9 What does an improvement per human effort establish?
The section argues that response predictability, learning benefit, and efficiency are distinct claims requiring different comparisons and evidence. It sets out tests for selection, assimilation, and closed-loop interaction while noting limits on what logs and formal guarantees establish.
9.1 A map of objectives rather than a single efficiency score
- A short response code on one stream does not establish a better teaching policy; policy comparisons require evaluating resulting predictions and human work.The review distinguishes response predictability, information acquisition, and learning benefit rather than treating them as one efficiency score.Source: §9.1 · p. 22
9.2 Observed, estimated, and latent quantities
- Interaction logs do not necessarily identify human attention or selection, while predictive improvement can be tested without recovering exact target gain.An audit can estimate log-loss improvement without recovering the entropy floor, provided its reference labels represent the declared target and checkpoint selection is handled separately.Source: §9.2 · pp. 22, 24
9.3 Evidence for learning and efficiency
- A benefit claim must specify its outcome, reference population, cost unit, allowed actions, and training resources.Current-mask completion, later-image prediction, and response prediction are distinct endpoints; human time is distinct from cost proxies.Source: §9.3 · p. 24
9.4 What does selection reveal?
- Observed marking frequency alone cannot identify both error prevalence and the probability that an error is detected and marked.The proposed audit samples displayed but unmarked regions using an independent reference or separate inspection protocol.Source: §9.4 · p. 24
9.5 Does using selection improve assimilation?
- To test whether selection improves assimilation, compare addressed-label learning with a learner that also models selection on the same correction stream and fitting resources.Independent audit risk measures the outcome without recovering the source posterior; replaying a fixed correction multiset in different orders can isolate learner-side sequencing effects.Source: §9.5 · p. 24
9.6 Does the interaction policy improve outcomes at matched effort?
- A closed-loop policy test requires actual interaction because changing views or updates also changes what teachers see and supply.Compare against strong baselines under common action, initialization, and fitting budgets, and record interaction, timing, and evaluation details.Source: §9.6 · p. 25
9.7 Extensions: teaching stages, exposure, and stopping
- Progression from clear examples through imbalance correction to boundary refinement remains a hypothesis, motivating tests of stage-dependent features and interventions.The review also cautions that formal active-learning guarantees require specific conditions and do not automatically apply to arbitrary neural updates or human corrections.Source: §9.7 · p. 25
- A low predicted gain or empty correction does not certify low remaining distortion; stopping claims require representative observations and a declared risk for a changing predictor.Confidence sequences apply under specified sampling and estimand assumptions.Source: §9.7 · p. 25
10 Discussion and conclusion
The review concludes that response protocols, assimilation, and effort denominators shape what empirical comparisons establish. Its mathematical account links information to expected log-loss improvement but leaves open whether modelling selection and sequencing improves segmentation at matched effort.
10 Discussion and conclusion
- Heterogeneous empirical outcomes support distinctions among response protocols, assimilation, and effort measures, not a ranking of systems.Fewer annotations do not establish less human effort, and completion quality does not establish reusable-model performance.Source: §10 · p. 25
- The mathematical account explains when information yields expected log-loss improvement and how an approximate update can lose it.The open question is whether modelling selection and sequencing improves assimilation and, in actual interaction, segmentation at matched effort.Source: §10 · p. 25
- The formal core holds the target fixed within a session; changing teacher interpretations or interpretations influenced by model exposure would require an additional state model.That extension is not required merely because RootPainter trains asynchronously.Source: §10 · p. 25
A Standard identities used in the comparison
The appendix records standard probability and information identities used in the comparison, emphasizing that application-specific choices concern random objects, conditioning information, and learner updates. Its derivations rely on the stated finite-entropy or finite-loss conditions.
A Standard identities used in the comparison
- The appendix presents established probability and information identities rather than new results, with finite-entropy or finite-loss conditions where differences are taken.The application-specific choices are the random objects, available conditioning information, and learner update.Source: A Standard identities used in the… · p. 26
A.1 Log-loss risk, entropy, and distortion
- For a fixed target and common evaluation weights, the entropy term cancels from expected risk differences between predictors.The cancellation does not establish a relation to target entropy and does not extend to a changed target or evaluation population.Source: §A.1 · p. 26
A.2 Predictive information and assimilation
- The exact Bayesian update turns evaluation-label information into expected log-loss improvement; an approximate learner's excess risks account for possible loss of that improvement.The information about fresh evaluation labels is bounded by information about the target under the stated conditioning assumptions.Source: §A.2 · p. 26
A.3 BALD as expected KL disagreement
- Expected KL disagreement is information about the label under the joint model; identifying it with source-target information needs additional belief and target-relevance conditions.Those conditions are specified in the main text rather than supplied by the identity alone.Source: §A.3 · p. 26
A.4 Entropy, code length, and coding-model mismatch
- Ideal response code length separates response information from coding-model mismatch, while practical integer code lengths incur less than one bit above entropy.The main text uses ideal probability-based lengths and treats implementation overhead separately.Source: §A.4 · p. 27
A.5 Posterior-entropy averaging and telescoping gains
- Averaging posterior entropy over memory yields conditional target entropy, and successive conditional-entropy reductions telescope across rounds.This expectation is over the joint source-history distribution, not the fixed-target average.Source: §A.5 · p. 27
A.6 Support and label chain rules
- The response code factors into support and conditional-label terms, and the corresponding information decomposition does not assume support-label independence.A label term per marked location does not make locations independent observations or imply one optimizer update per term.Source: §A.6 · p. 28
B The parametric MDL benchmark and its qualifications
The parametric MDL benchmark is a qualified coding construction under regular statistical assumptions, not a theorem about adaptive human correction streams. The review’s purposive comparisons illustrate mechanisms and protocol-specific outcomes rather than rank systems or estimate pooled effects.
B The parametric MDL benchmark and its qualifications
- Under regularity conditions, a fixed regular k-parameter model’s two-part code uses approximately (k/2) log2 N bits for parameter precision, with bounded-order loss from rounding.The benchmark expansion requires the stated model, precision, and code construction; boundedness in probability alone does not justify its expected form.Source: B The parametric MDL benchmark and its… · p. 28
- The Bayesian mixture code has the same leading likelihood and (k/2) log2 N terms under regular Laplace conditions; an expected expansion additionally requires uniformly bounded absolute expected remainder.The review retains this benchmark to clarify interpretation, not as a theorem about adaptive human correction streams.Source: B The parametric MDL benchmark and its… · p. 28
C.1 Question, design, and scope
- The review is a purposive critical narrative synthesis, not a systematic review or meta-analysis, and does not establish comprehensive coverage of 2024–2026.The search checked recent counterexamples and missing mechanisms but was not prospectively sampled under a registered search.Source: §C.1 · p. 29
- The review’s source selection sought mechanisms spanning feedback units, response protocols, updates, and evaluation, while excluding general architecture variants without a relevant interaction or evaluation mechanism.Targeted searches and citation tracing supplemented prior reviews and primary sources; coverage is not exhaustive or independently double-screened.Source: §C.1 · p. 29, §C.3 · p. 30, Other screened hits · p. 31
D.1 Comparison and extraction rule
- Comparisons were selected to illustrate links among selection, assimilation, and effort, not to maximize effects or rank systems; heterogeneous outcomes are not pooled.The appraisal distinguishes missing reporting from design limitations, and uncertainty measures such as SD, SE, and confidence intervals are not interchangeable.Source: §D.1 · p. 32
D.2 Appraisal of the main comparisons
- RootPainter’s duration-matched comparison does not establish superiority: its mixed-effects strategy test is nonsignificant, with results close for roots and dense annotation winning for one nodule user.The reported means and SEs aggregate six user–dataset runs; order was uncontrolled.Source: §D.2 · p. 32
- 14.68% of annotation cost: oracle-cost CEREALS acquisition reaches the target, exposing the gap introduced by cost prediction in five Cityscapes simulations.The target is 95% of full-data mIoU 0.605; this uses recorded polygon-click cost, not measured human time.Source: §D.2 · p. 33
- Size–entropy selection reports fewer loops to strong performance than random selection in the reported XpertLoop organ tasks, but suitability depends on the dataset.Plain entropy can underperform in ablations; size–entropy adds inference over unlabelled cases, and the test set informs stopping rather than serving as untouched evaluation.Source: §D.2 · p. 33
D.3 Adjacent cases and source checks
The adjacent studies distinguish simulated interaction from measured human evidence and separate annotation completion, reusable learning, and evaluation outcomes. Their contrasts support protocol-specific conclusions, not pooled effects or a claim that no complete demonstration exists elsewhere.
D.3 Adjacent cases and source checks
- 34 measured seconds per corrected mask versus 108 estimated seconds for historical COCO polygons is not a matched-user timing trial.The historical speed is scaled linearly by polygon vertex count; the separate downstream experiment trains on completed masks without testing whether the ranker improves reusable learning.Source: §D.3 · p. 34
- SPA transfers user-adapted GMM parameters across images, but its five-participant user study does not isolate transfer benefits.Its latent-Gaussian/i.i.d. convergence claim is not a guarantee for general spatial correction behaviour.Source: §D.3 · p. 34
- 91.14% versus 87.21% domain-preference DSC supports adaptation under HiTTA’s simulated protocol, not demonstrated human-effort efficiency or a validated head-selection rule.Existing ophthalmologists’ masks assigned to different domains simulate preferences; live timed corrections were not measured.Source: §D.3 · p. 34
- SegAgent’s successful imitation under an automated click simulator does not establish realistic human-response probabilities or better reusable learning at matched human effort.The simulator generates trajectories from reference-mask errors rather than recorded human trajectories.Source: §D.3 · p. 34
- The adjacent cases include timed human workflows, simulated sequencing, preference transfer, and within-annotator feasibility evidence, each with distinct inferential limits.For ExiL, one expert timed 360 frames from three volunteers; the evidence is live feasibility evidence with limited transfer across users.Source: §D.3 · pp. 34–35
- The included studies provide protocol-specific contrasts, but the review calculates no aggregate effect and does not establish that no complete selection-to-live-efficiency demonstration exists elsewhere.Coverage is not exhaustive, and independent appraisal of consequential evidence classifications remains desirable.Source: §D.3 · p. 35