cs.CVOct 6, 2026
SaveHave I Seen Enough? Frozen Video-Language Models Encode Evidence Readiness
Organizations: INSIGHT Lab, Ben-Gurion University of the Negev, Israel · Ben-Gurion University of the Negev, Israel
Abstract
Streaming video-language models must decide not only what to answer, but whether the evidence needed for the current question has arrived. Existing systems learn that decision as a separate trigger; we ask whether an unmodified model already computes it. We show that frozen VideoLLMs carry a linearly readable evidence-readiness signal, labelled from timestamped evidence rather than from model output. It decodes in all seven models of a shared byte-identical evaluation (AUROC 0.733-0.905 under the strictest not-ready sampling, where a fitted clock is near chance), and a probe fitted without any of a benchmark family's footage still reads that family. It is question-conditioned: on byte-identical windows, changing only the question reverses the readout on 66.1% of pairs, while every question-blind control is at chance by construction. The model can answer incorrectly and still encode readiness: AUROC remains 0.722 among wrong answers. Readiness also beats uncertainty estimators and their supervised combination on latency-matched answer selection, and tracks independent human judgments more closely than confidence. Released streaming triggers are also linear readouts, yet a trained trigger read on its own base model's activations is approximately orthogonal to readiness and decodes it far less accurately than a probe. We turn the readout into Readiness Gating, an answer-timing policy that improves accuracy by up to +9.75 pp at matched video duration with negligible computational overhead. How much it gains varies with the accuracy headroom the task makes available: across 26 configurations the gain tracks that headroom, and an intervention that moves it over identical pixels moves the gain with it.
Figures & tables
Figure 1: A model can be confident and unready at the same time. Asked an object’s colour before it enters the stream, the model answers confidently from prior alone; a linear readout of its frozen residual stream reports that the evidence has not arrived. (Schematic; measured in section 4 .)
| source | geometry | label | units | shuffle control | confidence | visual layer- | readiness AUROC |
|---|---|---|---|---|---|---|---|
| HD-EPIC † | composite | cue present | 520 prefixes | ||||
| HD-EPIC † | matched pair | availability | 4,800 windows | ||||
| NExT-GQA | quad | containment | 1,208 quads | ||||
| NExT-GQA | prefix | onset | 5,580 windows | ||||
| ST-Evidence | prefix | onset | 6,546 windows | ||||
| ST-Evidence-Gen | prefix | onset | 4,346 prefixes |
Table 1: Core representation results. Labels come from evidence timestamps, not model correctness. Rows marked † pose questions we construct rather than questions the source ships (Appendix D.2 ). The column is nine separate tests, not one scale: rows differ in geometry and strictness, so each AUROC is read against the three columns beside it rather than against the rows above it (Appendix A ). Read that way the verdict is uniform, with readiness above all three columns on every row by to . Only the shuffle is a control, and it is measured on every row rather than by construction; the other two are competing readouts that may legitimately carry signal. Visual layer- reads raw patch embeddings before any decoder layer, against a language-position layer- readout forced to everywhere ( section 3 ); it is forced by the design on the crossed quads, and on the controlled composites is quoted under a pair-preserving grouping ( subsection C.4 ).
| measurement | Oops! | ActivityNet-Captions |
|---|---|---|
| readiness AUROC on evaluated windows | ||
| elapsed-time AUROC on the same windows | ||
| selection accuracy, answering at end of video | ||
| selection accuracy, confidence | ||
| selection accuracy, Readiness Gating (ours) | ||
| gating advantage over supervised combination | pp | pp |
Table 2: Uncertainty baselines do not recover readiness on the two tested high-conversion settings. All arms answer at an identical deadline; the middle block gives the accuracies the advantages come from (paired margins and pp over confidence). The two AUROCs are read on the windows the policy chooses among, not the Table 1 pools of the same names. Every estimator loses on both, and adding a supervised combination back to gating gives no detectable gain.
Figure 2: Readiness accrues with depth, and it is not the model’s own confidence. Seven models, four language-backbone families, on one frozen Oops! manifest of windows over videos with not-ready rows drawn from both sides of the evidence. (a) Held-out AUROC at the final language position against fraction of decoder depth, cluster-bootstrap bands over videos; layer is the raw input embedding, exactly by construction, and every model peaks between and of depth rather than at the output. (b) Readiness at each model’s best depth, rows ordered by parameter count, against the model’s own answer probability ( – ), the measured layer- visual readout ( – ) and the label-shuffle null, all on the same rows. Parameter count does not order readiness: the B model reads above four larger ones.
| HERBench shot-ID | ActivityNet, same-action | |||
| monitor | readiness AUROC | selection acc. | readiness AUROC | selection acc. |
| answer at end (no selection) | — | — | ||
| evidence oracle | — | — | ||
| Readiness Gating (ours) | ||||
| confidence | ||||
| elapsed time | — | — | ||
Table 3: Released streaming triggers, scored against evidence-readiness labels on two substrates. These are the policy ladders, and windows, not the Table 1 probing pools of the same names ( Appendix I ). At a common deadline every arm emits the answer cached from its own highest-scoring window, so selection is latency-matched by construction. Our probe is fitted on each dataset’s training folds and evaluated on held-out folds; the released triggers are applied without fitting. Removing that asymmetry does not change the ordering: on MMDuet’s own activations a probe reads where its trained head reads ( subsection 4.5 ). The comparison tests transfer to readiness labels, not which complete streaming system performs best. Dispider is a secondary baseline and decides once per s, too coarse for the ActivityNet ladder, so those cells are empty; every rival selection arm is significantly worse than plain confidence ( to ).
Appendix figures & tables26 assets
Supplementary material from the paper’s appendix.
Appendix
| geometry | shape | frames | label | what it is for |
|---|---|---|---|---|
| growing prefix | , evidence-relative ends | onset, | graded accumulation; earliest evidence | |
| recent window | , fixed span, overlap | containment | first evidence-agnostic ladder | |
| containment window | s span, free s stride | @ fps | containment | the reference format; model axis, utility |
| short window | s span, s stride | @ fps | containment | point-onset footage |
| matched pair | rows per item, all else fixed | @ fps | availability | availability vs. correctness; causal work |
| crossed quad | 2 questions 2 windows | @ fps | containment | question-conditioning; visual controls at |
Table A.1: Every streaming geometry, its shape, its frame budget, and its label. The containment window (bold) is the primary geometry: constant span and an evidence-agnostic checkpoint grid. Sampled frame count is constant within each geometry, so no readout can substitute “how many frames” for “which frames”. Reported AUROCs are not comparable across rows: the rows differ both in how strict the label is and in how the not-ready rows are drawn (Appendix A.3 ), so each result in section 4 names the row it belongs to.
| label | definition | used for |
|---|---|---|
| containment | coverage contains the evidence hull; null on partial | probe fitting |
| deployed | the same, partial folded into not-ready | policy evaluation |
| onset | any delivered frame inside an evidence segment | threshold sensitivity |
Table A.2: Three labels from one cache. Rows with partial coverage are excluded from binary probe fitting, because their state is genuinely intermediate; a deployed policy may not drop a row, so it treats them as not ready.
| scored against the benchmark’s evidence label | NExT-GQA | ActivityNet-Cap. | Oops! |
|---|---|---|---|
| AUROC, readiness probe | |||
| AUROC, independent rater verdict | |||
| rater accuracy | |||
| rater | |||
| rater sensitivity / specificity | |||
| screens ( ) |
Table B.1: A person shown the delivered frames recovers the benchmark’s own evidence label about as well as the probe does. Intervals overlap on all three evaluations, with the raters ahead on NExT-GQA and behind on the other two. The on Oops! is because of its screens sit at partial coverage and carry no binary label there; the rater-verdict comparisons in Table B.2 use all .
| scored against the raters’ readiness verdict | NExT-GQA | ActivityNet-Cap. | Oops! |
| AUROC, readiness probe | |||
| AUROC, the model’s own confidence | |||
| AUROC, an external LLM asked directly | |||
| model accuracy, rater-ready screens | |||
| model accuracy, rater-not-ready screens | |||
| screens ( ), probe and confidence |
Table B.2: The readout tracks a human readiness verdict; the model’s own confidence and an external LLM asked the question directly do not. All three rows are continuous scores on the same screens. The lower block shows the raters’ verdict is not a private opinion: the frozen model answers correctly , and pp more often on the screens they called ready.
| rater answer accuracy | pre (far) | pre (near) | partial | first ready | post | overall |
|---|---|---|---|---|---|---|
| NExT-GQA | ||||||
| ActivityNet-Cap. | ||||||
| Oops! |
Table B.3: The windows sit where the design says they do. Accuracy rises into the evidence-containing windows on all three evaluations. It stays high in the post window on NExT-GQA, where the delivered frames still contain the evidence, and falls on the other two, where by that window the evidence has left the span. This reproduces the non-monotonicity of the labels themselves.
| evaluation | held-out unit | fitting and scoring |
|---|---|---|
| controlled-composite headline | cue object | Nested within-pool fitting; evaluation on held-out objects. |
| controlled-composite visual null | pair-preserving group | Both questions on identical frames share a fold; distinct from the headline protocol. |
| natural within-benchmark probes, including the model axis | video | Video-grouped inner selection and outer evaluation; matched pairs and crossed quads stay together. |
| held-out-family transfer | footage family | Selection on source families only; the target family’s footage is excluded from fitting. |
| external transfer | disjoint source and target videos | Source scaler, layer and probe are frozen on the target; a target-fitted probe is a separate comparison. |
Table C.1: Which groups are held out, and where fitting occurs. The geometry, label and scored row counts for the within-benchmark representation results are listed together in Table E.1 . Table D.1 inventories the source and builder pools, whose counts can differ from the scored subset. Transfer details are in Appendices F and G .
| layer | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| evaluation | probe | final | vision | option | shuffle | clock | conf. | resid. drop | over conf. |
| NExT-GQA | |||||||||
| ST-Evidence-MCQ | |||||||||
| ST-Evidence-Gen | |||||||||
| the streaming cohort | |||||||||
| HERBench shot-ID | |||||||||
Table C.2: The dissociation battery. One row per evaluation, each read against its own comparisons; values are not comparable down a column, because the rows differ in geometry and in label strictness (Appendix A.1 ). The three layer- cells are different objects (Appendix C.3 ): at the final language position, and at the option tokens wherever the option count is constant, layer is the same embedded state on every row and is forced to , whereas the vision cell is a baseline that can carry real signal. The two option cells that depart from belong to the pools that mix two- and four-option items, where that cell becomes an indicator of the option count rather than a control. A dash marks a cell not available on that pool; such a cell is never assumed to be . For shot ordering, the activation cache was not retained. The clock column names a different pool in each row: it is high wherever evidence onset is strongly time-correlated, which is why those rows carry the qualification stated in Appendix E.1 and are not read as readiness results on their own. “Conf.”, “resid. drop” and the last column, which gives the AUROC gain of the readiness score over a confidence-only baseline, all come from the pooled confidence-separation estimator on the same rows ( Table E.3 ), which is why the confidence figure here can differ in the third decimal from the fitted baseline in Table 1 . Every row uses the pixel-cleaned pool that section 4 reports on. The indented row is the same graded shot-boundary prefix ladder as the row above it, restricted to the streaming eligibility cohort; it is a growing prefix, which is why its clock sits with the other prefixes.
| pool | footage | evidence timestamps | items | rows | source groups | readiness | model acc. |
|---|---|---|---|---|---|---|---|
| controlled composites, crossed | HD-EPIC egocentric, P01–P09 | cue presence by construction, 17-object library | 260 (520 prefixes) | 1,040 | 31 videos, 90 clips | 0.5 by constr. | – |
| matched present / absent windows | HD-EPIC egocentric, P01–P09 | official fine-grained action segments | 2,400 pairs | 4,800 | 142 videos | 0.5 by constr. | – |
| fine-grained action VQA | HD-EPIC egocentric | the benchmark’s own items | 1,200 | – | 86 videos | – | – |
| NExT-GQA prefixes | NExT-QA | human temporal grounding | 1,000 | 5,695 | 473 videos | 3,600 / 2,095 | 0.729 five-opt., 0.695 six-opt. |
| NExT-GQA crossed quads | NExT-QA | human temporal grounding | 1,208 quads (1,196 intact) | 4,808 | 277 videos | 0.5 by constr. | – |
| ST-Evidence-MCQ prefixes | NExT-QA | independent human evidence | 1,298 | 6,756 | 348 videos | 4,400 / 2,356 | 0.704 |
Table D.1: Source and builder pools, evidence annotations, and sizes. “Rows” counts delivered checkpoints, not items. Readiness gives ready / not-ready counts where recorded and marks the balanced matched and crossed designs. Source groups describe the listed pool, not necessarily the split or bootstrap key. A dash marks a value not recorded. Pixel filtering and its pool-specific effects are detailed in Appendix D.5 ; scored representation pools are identified in Table E.1 .
| evaluation | questions | source of the items |
|---|---|---|
| HD-EPIC controlled composites | constructed | synthetic cues composited onto real footage |
| HD-EPIC matched present / absent | constructed | from the official fine-grained action segments |
| HD-EPIC fine-grained action VQA | shipped | the benchmark’s own items, read verbatim |
| NExT-GQA | shipped | NExT-QA’s questions with human grounding |
| ST-Evidence-MCQ | shipped | the benchmark’s own QA and evidence |
| ST-Evidence-Gen | shipped | the benchmark’s own items, two item filters applied |
Table D.2: Task provenance. “Constructed” means we authored the question, the answer and the distractors from the source’s annotation; the footage and the evidence timestamps are always the source’s. Readiness labels are derived from source timestamps throughout.
| model | HuggingFace id | vision tower | language backbone | dec. lay. | how frames are fed | probed layers |
|---|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct | Qwen/Qwen3-VL-8B-Instruct | Qwen-ViT | Qwen3-8B | 36 | native video path, video metadata | 0, 14, 18, 23, 29, 32 |
| Qwen3-VL-2B-Instruct | Qwen/Qwen3-VL-2B-Instruct | Qwen-ViT | Qwen3-2B | – | native video path | 0, 11, 14, 18, 22, 25 |
| Qwen3-VL-32B-Instruct | Qwen/Qwen3-VL-32B-Instruct | Qwen-ViT | Qwen3-32B | – | native video path | 0, 26, 32, 42, 51, 58 |
| InternVL3.5-8B [ Wang et al., 2025b ] | OpenGVLab/InternVL3_5-8B-HF | InternViT | Qwen3-8B | 36 | multi-image, one px tile per frame | 0, 14, 18, 23, 29, 32 |
| LLaVA-OneVision-7B [ Li et al., 2025 ] | llava-hf/llava-onevision-qwen2-7b-ov-hf | SigLIP [ Zhai et al., 2023 ] | Qwen2-7B | 28 | native video path, no metadata | by depth fraction |
| LLaVA-OneVision-1.5-8B [ An et al., 2025 ] | lmms-lab/LLaVA-OneVision-1.5-8B-Instruct | – | Qwen3-8B | – | – | 0, 14, 18, 23, 29, 32 |
Table D.3: Every model evaluated. Layers are probed at fractions of each model’s own depth rather than at transcribed indices, because decoder depth differs across architectures; the indices shown are the resolved ones where they are recorded, and a dash marks a value not recorded. Qwen3-VL-8B-Instruct [ Qwen Team, 2025 ] is the default everywhere unless a result names another.
| geometry | frames | rate | span | stride | decode resolution |
|---|---|---|---|---|---|
| controlled composites | 16 | 2 fps | fixed 8 s | – | 448 px long side |
| growing prefixes | 32 | – | evidence-relative ends | 448 px long side | |
| containment windows | 32 | 2 fps | 16 s | free 2 s | – |
| short windows | 8 | 4 fps | 2 s | 0.5 s | – |
| sparse bursts | 16 | 2–3 fps | 3–6 bursts of 1–2 s | one block | – |
Table D.4: Frame budgets per geometry. Frame count is constant across every row of a geometry. Prefixes sample uniformly across , except on the segment-aligned pools, where sampling guarantees at least two frames per revealed evidence segment. A dash marks a value not recorded.
| pool | dropped | of | rate |
|---|---|---|---|
| NExT-GQA prefixes | 115 prefixes (21 items entirely) | 5,695 | 2.0% |
| ST-Evidence-MCQ prefixes | 210 prefixes | 6,756 | 3.1% |
| ST-Evidence-Gen prefixes | 5 prefixes | 4,346 | 0.1% |
| matched present / absent windows | 0 | 4,800 | 0% |
| model-axis short windows | 0 | – | 0% |
| Oops! short windows | – | – | 0.3% |
Table D.5: Pixel quality-control drop counts. The filter is model-independent and runs before any outcome exists. Only the last row has a rate large enough to change what a pool can support.
| pool | geometry | label | rows / clusters | probe AUROC | shuffle | L @vision | clock |
|---|---|---|---|---|---|---|---|
| controlled composites | controlled composite | cue presence | 1,040 / 17 obj. | – | |||
| matched windows | matched pair | availability | 4,800 / 142 | ||||
| the same, InternVL3.5-8B | matched pair | availability | 4,800 / 142 | ||||
| NExT-GQA prefixes | growing prefix | onset | 5,580 / 458 | ||||
| ST-Evidence-MCQ prefixes | growing prefix | onset | 6,546 / 338 | ||||
| ST-Evidence-Gen prefixes | growing prefix | 4,346 / 1,529 |
Table E.1: Every decodability cell with its own comparisons. All AUROCs are nested, cluster-held-out, and read at the final prompt token. The column is not a ranking: the rows differ in geometry ( Table A.1 ), in label strictness ( Table A.2 ) and in where the not-ready rows are drawn from (Appendix A.3 ), so each cell is interpretable only against the comparisons printed beside it. The shuffle is the cluster-grouped label shuffle measured on that pool, not assumed to be ; on the graded prefixes it is , a base-rate artefact, so that row reads “ against ”. The clock is a fitted elapsed-time baseline on the same rows: it is informative by construction on a growing prefix, exactly on a balanced crossed pool, and on the shot-ordering prefixes , where evidence onset is strongly time-correlated and the readiness reading is therefore not creditable on its own. Layer at the final language position is exactly in every row by construction; the visual layer- baseline, which can carry real signal, is the adjacent column. It is forced to on the crossed quads by the design and is quoted on the controlled composites under a pair-preserving grouping (Appendix C.4 ); the dash on the shot-ordering prefixes marks a cell that is not measurable, its activation cache not having been retained, rather than one assumed to be , and the composited stimuli carry no elapsed-time column at all. Row counts identify the reported decoding pools; pixel quality control is documented in Appendix D.5 , and grouping in Table C.1 . Oops! is quoted in the both-sided negative regime; the before-only regime reads but carries a clock at .
| paired drop on byte-identical frames, pairs | |||||
|---|---|---|---|---|---|
| model | language backbone | identical question | paraphrase | distinct question | register-matched |
| Qwen3-VL-2B | Qwen3 | ||||
| Qwen3-VL-8B | Qwen3 | ||||
| Qwen3-VL-32B | Qwen3 | ||||
| LLaVA-OV-1.5-8B | Qwen3 | ||||
| GLM-4.1V-9B | GLM-4 | ||||
Table E.2: Replacing the question collapses the readout in every model tested. Each cell is the mean paired drop in readiness score between a row and the same row re-asked with a donor question, on the same evidence-carrying pairs and byte-identical frames. Distance increases left to right: another item’s identical question, a paraphrase from a different corpus, and a specific question from a pool carrying one question per item, with a register-matched build of that last donor in the final column. Under a genuinely distinct question every model drops by to , on – of pairs. Register is not what drives it: Pixtral’s raw and register-matched drops agree to four decimals ( , ) and Idefics3’s register-matched drop is the larger of its two, so the two builds should be read as a range rather than a point.
| pool | readiness AUROC | residualized on conf. | drop | confidence alone | increment |
|---|---|---|---|---|---|
| controlled composites, factorial | |||||
| controlled composites, crossed | |||||
| matched windows | |||||
| the same, InternVL3.5-8B | |||||
| NExT-GQA prefixes | |||||
| ST-Evidence-MCQ prefixes |
Table E.3: Confidence explains none of the readout. Each row residualizes the readiness score on the model’s output confidence over the same examples and re-scores; the drop is raw minus residualized, and the largest movement in either direction is . These cells use a pooled estimator, and the ActivityNet row is the containment-label build, so the levels are not comparable with Table E.1 and only the drop and the increment are read here. The increment is the AUROC gain of the activation score over a confidence-only baseline on the same rows.
| Oops! windows | ActivityNet | ||||
| signal | what it is | acc. | probe this | acc. | probe this |
| sc_agree | share of sampled answers, each with a rationale, agreeing with the mode | ||||
| sc_neg_entropy | negative entropy of that sampled answer distribution | ||||
| se_neg | semantic entropy: open-ended answers clustered by bidirectional entailment | ||||
| se_neg_lennorm | the same, length-normalized | ||||
| se_neg_loose | the same, one-direction clustering | ||||
Table E.4: Every alternative signal, and the probe’s margin over it. Selection accuracy is measured at an identical deadline, so the comparison is latency-matched by construction; the probe reads and on these substrates and answering only at the end reads and . Intervals are video-cluster bootstraps of the paired difference. Every alternative loses on both substrates; adding the supervised head back to the probe moves selection by and , both intervals spanning zero. Signal names match 1(b) .
| rows | ready | not ready | AUROC | |
|---|---|---|---|---|
| 2,309 | 1,877 | 432 | ||
| 1,121 | 667 | 454 | ||
| 379 | 174 | 205 | ||
| 211 | 81 | 130 | ||
| 156 | 51 | 105 | ||
| 75 | 24 | 51 |
Table E.5: Per- strata, each with its minority-class count. Read on the pixel quality-controlled build of the graded prefixes, which reads overall. The minority class is in bold and falls from rows at to at ; the cell rests on positives and is printed for completeness rather than read as a decline. Strata above are not tabulated for the same reason. Within each the mean score rises monotonically with the number of segments seen for – , which is of the rows, with within- Spearman correlations of , , , and .
| backbone | model | regime | AUROC | CI | L @final | L @vision | shuffle | clock track |
|---|---|---|---|---|---|---|---|---|
| Qwen3 | Qwen3-VL-2B | before | ||||||
| near | ||||||||
| after | ||||||||
| both | ||||||||
| Qwen3 | Qwen3-VL-8B | before | ||||||
| near |
Table F.1: The full model axis: seven models four negative-sampling regimes. Readiness AUROC is nested cross-validated with video-disjoint folds; the interval is a video-cluster bootstrap. The final-language-position layer- column is in all cells and is forced to be: at layer a fixed text position holds the same embedded token on every row, so the readout is degenerate before any frame is consulted. The visual layer- column can carry real signal because it reads raw patch embeddings. It spans – , and the language-position readout clears it in every cell, by at least . The within-video shuffle stays inside – throughout. The fitted elapsed-time clock is a property of the regime rather than the model: ( before ), ( near ), ( after ), ( both ). Clock tracking is one number per model, on the before both contrast.
Figure F.1: Readiness is neither a stopwatch nor an uncertainty estimate. (a) Readiness AUROC against the AUROC of a fitted elapsed-time clock, one point per model negative regime. Colours identify Qwen3-VL-2B (blue), Qwen3-VL-8B (red), Qwen3-VL-32B (green), LLaVA-OV-1.5-8B (orange), GLM-4.1V-9B (violet), Idefics3-8B (brown) and Pixtral-12B (teal). Before-only sampling makes the clock ( ) stronger than every probe. With both-sided sampling the clock falls to , while all seven probes remain informative: their AUROCs change much less than the clock’s. (b) Eleven modern uncertainty estimators, plus a supervised head fitted over all of them ( composite ), compared on fixed-deadline answer selection. Points are Readiness Gating’s margin over each baseline with a cluster-bootstrap interval, using earliest-window tie-breaking; Appendix E.4 reports the favourable-tie-break comparison. Every estimator loses on both substrates. The last row compares the combined head ( composite+probe ) with Readiness Gating alone: adding the baseline scores yields no detectable selection gain, with both intervals spanning zero.
| held-out family | transfer AUROC | CI | own AUROC | retention | L @vision | shuffle | clock |
| NExT-GQA | |||||||
| ST-Evidence-MCQ | |||||||
| ST-Evidence-Gen | |||||||
| ActivityNet | |||||||
| Oops! | |||||||
| mean retention, footage-group hold-out | |||||||
Table F.2: One probe, fitted without a family’s footage, read on that family. Rows are the footage-group hold-out, in which NExT-GQA and ST-Evidence-MCQ are removed together because they share videos. Retention is transfer AUROC over the family’s own fitted AUROC, the comparable quantity given that the family ceilings themselves run – . Holding out benchmarks singly raises the mean to , and the whole of that difference is the two NExT-QA families, whose retention rises by and once their shared footage is allowed back into the training union. Every cell sits fully inside the source union’s delivered frame-density envelope.
| rank | aligned (median) | random-rotation null | row-shuffled alignment | pairs beating own null |
|---|---|---|---|---|
Table F.3: The cross-model rotation, with both controls read at the same rank. Medians over the ordered model pairs on the frozen Oops! manifest. Rank is the primary rank; reading an aligned value against a null taken from a different rank is exactly the error these columns exist to prevent. In un-aligned transfer, the source probe is applied to the target’s activations with no map at all. It reads over the ordered pairs whose two models share , i.e. at chance. Recovering the basis therefore needs paired activations on shared inputs, and this is not a shippable probe.
| cell | base rate | carried-over | at | Platt at | oracle threshold | constant predictor |
|---|---|---|---|---|---|---|
| OVO-Bench (window) | ||||||
| Fine-grained Action | ||||||
| NExT-GQA ladder | ||||||
| ActivityNet ladder |
Table G.1: Accuracy under a threshold refitted on eight labelled target videos. Direction, layer, and scaler are frozen throughout; only the threshold moves. The two refitting arms differ in objective: our intercept-only variant of Platt scaling [ Platt, 2000 ] moves the crossing point, whereas Youden’s [ Youden, 1950 ] maximizes TPR FPR, which is not an accuracy objective and loses by roughly on the same eight labels ( against on the first row). The last column is the accuracy of always predicting the majority class, and it is the number that bounds how much this repair can be worth.
Figure G.1: The ranking transfers; the operating point does not. Reliability of the carried-over probe on the external benchmark over scored rows, against the diagonal. The probe concentrates almost all mass above , placing of rows there despite the pool’s far lower positive rate. At the carried-over threshold of , almost every row is therefore called ready. The curve is nonetheless monotone at the top end, which is why the ordering survives while the calibration does not (ECE ).
Figure H.1: Changing distractors moves headroom and policy gain together. (a) ActivityNet-Captions conditions hold items, videos, windows, labels, frames and answer positions fixed, changing only the distractors. The attainable band moves from to pp and from to pp; error bars are the reported confidence intervals. (b) The same two conditions (red) enter the -configuration comparison over ten substrates, each configuration entering once on its full item pool. The plotted quantity is the seconds-matched of gated stopping, not the selection edge quoted in subsection 4.4 , which is why some configurations sit below zero. The fitted slope is pp per band pp. Three high-band configurations carry much of the pooled relationship; the remaining give . These are configuration-level observations, not independent dataset replications. The band tracks gain but does not determine it.
| system | head | shape | base model |
|---|---|---|---|
| MMDuet | relevance , informative | ea. | LLaVA-OneVision (Qwen2-7B) |
| Dispider | silent | Qwen2 | |
| VideoLLM-online | streaming-EOS logit | unembedding row | Llama-3-8B |
Table I.1: The released triggers are linear residual-stream readouts, the same kind of object as our probe. MMDuet ships as a GB adapter and publishes a fps rate; Dispider’s head is embedded in its checkpoint’s root shards and decides once per s, and its sign convention is that a high score means respond . VideoLLM-online was obtained but never run, so no number is asserted for it anywhere in this paper.
| quantity | Qwen3-VL-8B | InternVL3.5-8B |
|---|---|---|
| commitment excess over a magnitude-matched control | ||
| deferral-to-answer flips, treatment vs. control | vs. | vs. |
| exceeds all alternative directions | yes, | yes, same |
| dose–response Spearman over four doses | ||
| answer-content equivalence test | passes | passes |
| replicates on evidence-present-but-wrong | yes | yes |
Table J.1: Adding the readiness direction changes answer commitment. Nothing was re-selected for the second model: it ran through the locked cell chosen on the first. The second model’s dose curve is not fully monotone, which is the one qualitative difference between them.