Proposed Experiments Register (Hypotheses & Designs)

The Experiment Ledger records experiments we have run. This page records experiments we intend to run, written down before any results exist, so that the hypothesis, the metric, and the pass/fail line can’t quietly move after the data comes in.

Naming note (2026-10-03): entries written before this date say “IDC” for the in-pass order corrections (null-prior, dual-mirror). Those corrections are retired; the production approach is hesitation-gating, with order bias monitored per release.

Lifecycle: Proposed → Ready (dependencies met, design reviewed) → Running → Done (or Closed / Parked when superseded). When an entry starts running, give it the next free EXP-XX number, add it to the ledger, and link it back here. Keep the PROP-XX entry, but mark it Done → EXP-XX and note any change to the design along with the reason for it.

Numbering: claim the EXP-XX number by adding the ledger row (status 🧪 Running) when the run starts, in the same change that marks the entry Running, so two teams never pick the same number. “Free” means unused on main and in every open pull request (gh pr list, then check each branch’s ledger and this page). The same applies to PROP-XX.

Where entries come from: besides our own open questions, every model comparison ends with a “Findings for dgem” table; each hypothesis in it becomes an entry here. The Source column records every study that raised it (Order = the order-bias research behind EXP-13–EXP-17, formerly called IDC; Laya, Strands, Clef = model comparisons; DI = Decision Index runs). When a later study raises the same point, add it to Source instead of opening a second entry.

---

1. Ground Rules for Every Entry

1. Pre-register. State the hypothesis, the primary metric, and the decision threshold before running. If you change them afterwards, say so explicitly.
2. Report the sample size next to every number. Include a 95% interval. Use a bootstrap over items for Brier/ECE/accuracy, and Wilson intervals for proportions.
3. Hold data out. Any parameter fitted on labeled data ($T^*$, cascade thresholds, --prior-alpha) must be evaluated on items it was not fitted on, either a held-out split or k-fold cross-validation.
4. Compare like with like. Run the baseline and the treatment in the same session, against the same endpoint revision and model build, over the same items. Record the endpoint URL, model path, and git SHA in the receipt.
5. Measure the noise floor. Where possible, repeat the baseline at least 3 times to see how much results move from run to run (see PROP-10).
6. Leave a receipt. Every run writes a benchmarks/results_*.json receipt with per-item probabilities, so figures can be regenerated from data (as scripts/generate_idc_diagrams.py does).

Small-sample guidance: 10-bin ECE is unreliable below roughly 200 items. Zero errors in n high-confidence items is still consistent with an error rate of up to about 3/n (the “rule of three”). Prefer paired comparisons: the same items under two conditions.

---

2. Register at a Glance

ID	Title	Question it answers	Depends on	Source	Priority	Status
PROP-00	Expose Mirror TVD on all four surfaces (enabler)	Can production use the mirror signal at all?	—	Order	P0	Closed: not worth it after EXP-17
PROP-10	Run-to-run & revision noise floor	How big a difference is real?	—	Order	P0	Done → EXP-14 (noise floor: ±1–2 items, Brier ±0.02)
PROP-01	Held-out temperature scaling	Does the ECE gain from $T^*$ survive out of sample?	—	Order	P0	Done → EXP-14 (met on 231 JevBench items, not on 50)
PROP-02	End-to-end IDC cascade	Does IDC + a mirror-aware gate beat the entropy-only cascade?	PROP-00, PROP-01	Order	P0	Partly → EXP-14 (entropy gates); mirror gate moves to PROP-13
PROP-03	Same-canvas vs separate-pass mirror	Does sharing a canvas hide disagreement?	—	Order	P1	Closed: replaced by PROP-12 (done → EXP-17)
PROP-04	Order sensitivity on real ChaosNLI	Do entropy / Mirror TVD / cyclic JSD track real human disagreement?	—	Order, Laya	P1	Proposed
PROP-05	Mirror merge rule	Is the 70/30 forward-priority merge the right one?	PROP-03 (optional)	Order	P1	Closed → EXP-14 (50/50 beats 70/30; neither beats single slot); mirror not used in production
PROP-06	Null-prior strength & label-aware priors	Can de-biasing improve calibration without raising order flips?	—	Order	P1	Closed: in-pass order corrections retired after EXP-13–EXP-17; order bias is monitored per release instead (overview §4.3)
PROP-07	Wording invariance	How often does rephrasing an option change the decision?	—	Order	P2	Proposed
PROP-08	Missing-context detection	Does uncertainty rise when a decisive fact is removed, or does an explicit evidence slot help?	—	Order, DI	P1	Running → EXP-40
PROP-09	Base-rate & retrieval prior adaptation	Can unlabeled target traffic or retrieval prior logits improve calibration without masking hesitation?	—	Order, DI, Clef	P1	Running → EXP-40
PROP-11	Letter collision in the mirror	Is the mirror’s damage caused by shared letters rather than by a second slot?	—	Order	P0	Done → EXP-15 (supported)
PROP-12	Separate-pass mirror (replaces PROP-03)	Does a mirror read in its own pass give a clean order signal?	PROP-11	Order	P1	Done → EXP-17 (supported, small effect)
PROP-13	Mirror-aware cascade (unblocks PROP-02)	Does an uncoupled mirror improve the hand-off gate?	PROP-12	Order	P1	Closed: not worth it after EXP-17
PROP-14	In-context vs blank-question prior	Does a prior estimated from real items fix null-prior’s overcorrection?	—	Order	P1	Parked (research only): production monitors order rather than correcting it (overview §4.3)
PROP-15	Correction strength by question type (extends PROP-06)	Do yes/no and lettered choices need different correction strengths?	PROP-14	Order	P2	Parked (research only, depends on PROP-14); see overview §4.3
PROP-16	Slot names are part of the prompt	How much do slot ids change answers?	—	Order	P1	Done → EXP-16 (single slot: no; second slot: yes)
PROP-17	Calibrated agent context pre-compiler (internal pilot)	Can one multi-slot pass decide which context blocks an agent turn needs, dropping little that matters?	—	Pilot	P2	Running (internal pilot)
PROP-18	Validate image metrics beyond EXP-09	Which image aspects (location, presence/count, attributes/relations, quality/occlusion, domain checks) does one pass read reliably, judged against ground truth, Gemini 3.x and classical detectors?	—	DI	P1	Phases 1–4 (pilot) done → EXP-09 re-baseline, EXP-22; real-image bench-vision, matrix image gate and live image cascade done; guided boxes and masks done → EXP-24
PROP-19	Cross-slot coupling on mixed-polarity yes/no questions	Why do two yes/no answers in one request collapse to one label, and which serving change removes it?	—	Strands	P0	Proposed (bug #68)
PROP-20	Shipped temperature per question type and domain	Does a held-out temperature per type (and per policy) as a served default fix dgem’s raw overconfidence without hurting error detection?	PROP-01	Laya, Strands, DI	P0	Closed (per-policy only): leave-one-suite-out on 8 dev suites, a global T and a per-type T each made ECE worse on 4 of 8; no served default. Per-policy temperature on all four surfaces instead
PROP-21	Fine-tuning DiffusionGemma: LoRA on the letter readout	Does a LoRA trained with a masked-denoising loss on answer slots raise in-domain accuracy and fix coupling without losing zero-shot breadth?	PROP-19, PROP-20	Laya, Strands, Clef	P1	Proposed
PROP-22	Distil dgem into a small student	Do dgem’s soft labels train a better 2B decision model than hard labels?	PROP-20	Strands	P2	Proposed
PROP-23	Nondeterminism at samples=1	Where do the 5–7% answer changes between identical requests come from, and can serving remove them?	—	Laya, Strands	P1	Proposed (#82)
PROP-24	Wide-option accuracy (6–26 options and beyond)	Why does accuracy fall with option count even in a single read, and which readout change recovers it?	—	DI, Clef, Strands	P1	Running → EXP-41
PROP-25	NLI “neutral” pull	Why does dgem over-predict “neutral” on XNLI, and does wording or a per-label prior fix it on the validation split?	—	Laya, Strands	P2	Proposed
PROP-26	First-shown bias on small option sets (extends PROP-14)	Does the first listed option win too often at K ≤ 6, and does an in-context prior remove it?	PROP-14	Laya, Strands	P2	Parked (research only, with PROP-14): order bias is monitored per release, not corrected
PROP-37	Contact-sheet multi-crop visual verification and co-denoised evidence slot	Does a 6-panel reference-vs-query contact sheet with a same-canvas evidence slot raise fine-grained visual accuracy over single-image reads?	PROP-08, PROP-18	DI, Clef	P0	Running → EXP-40
PROP-38	Set-of-Marks discrete next-hop UI navigation	Does discrete [A..F] candidate mark selection replace coordinate binning on ScreenSpot and web UI elements?	PROP-18	DI	P1	Running → EXP-42
PROP-39	Multimodal entity resolution and pack-size equivalence	Can one multi-slot read classify duplicate/variant/counterfeit pairs while extracting pack-size evidence and checking triangle consistency?	PROP-08, PROP-24	DI, Clef	P1	Running → EXP-43
PROP-40	Sliding-window live checklist and multi-agent swarm lock arbitration	Can an 8-slot single pass verify live transcript + frame checklist transitions and arbitrate concurrent agent lock collisions in under 150 ms?	—	DI	P1	Running → EXP-43

Numbering note: experiment and proposal numbers are shared with the team’s internal research log; gaps in the sequence are internal entries not published here.

Queued re-runs

Re-runs re-measure a published number after a harness or serving change; they need no hypothesis, but they are listed here so they are not lost. Each lands as a new run in benchmarks/runs/ and a dated note on the affected EXP page (history is annotated, not rewritten).

Re-run	Why	Affects	Source	Status
JevBench through bench-jev with ordinal score items kept as score	bench-jev flattens score items to choice; the /v1/systemone path keeps them ordinal and scores about 4–5 more of the 18 score items	EXP-11, EXP-14 to EXP-17 headline counts	Laya	Done (2026-10-05): totals within noise; score items 14–15 vs 9–14 of 18 (20261005-rerun-jev-score)
--auto-temperature fitted by k-fold, not in-sample	In-sample T is optimistic on 50 items	EXP-04, EXP-05, EXP-11 calibration figures	Laya	Done (2026-10-05): held-out ECE no better than raw on every 50-item receipt; note on EXP-11
Matrix v2 reference run (T1 + TC baseline only)	Sets reference ranges for calib_systemone, intents_systemone, gate_mixed_noul before v2 becomes the default	Regression matrix	Strands	Done (#87, 2026-10-03; re-referenced for v0.3.x)
EXP-23 on matrix v2 tier TC	The EXP-23 numbers came from the study harness that TC replaces; one TC run confirms the tooling reproduces them	EXP-23	Strands	Queued (needs a competitor deployment): #83

---

3. Entries

PROP-00: Expose Mirror TVD on all four surfaces (enabler)

• Closed (2026-09-28), not run. EXP-15–EXP-17 showed the mirror signal adds little over hesitation: a separate-pass mirror improves cross-validated error detection by at most 0.016 AUROC at twice the cost, and the same-canvas mirror costs accuracy unless labels are digits. Exposing it on every surface (PROP-00) or gating the cascade on it (PROP-13) is not worth the complexity; hesitation remains the production signal. Reopen only if a cheaper, uncoupled order signal appears.
• Motivation: --dual-mirror computes Mirror TVD/JSD in pkg/permutation.PostProcessDecisionResponse, but cmd/decide.go discards the return values. dgem serve, MCP, and the Studio don’t offer dual_mirror / null_prior_debias at all, which violates the four-surface parity rule in AGENTS.md. A mirror-aware gate (PROP-02) can’t run in production until this is fixed.
• Deliverable (not a hypothesis):
  • Add diagnostics.mirror_tvd / mirror_jsd per slot to the dgem decide output.
  • Add dual_mirror, null_prior_debias, and prior_alpha to GatewayDecideRequest, DecidePolicyToolInput / DecideCustomToolInput, and the Studio.
  • Add a cascade_mode option that escalates on mirror_tvd ≥ τ (with τ configurable).
  • Wrap the post-processing in an OTel child span (dgem.idc.postprocess).
• Acceptance: The same request returns identical mirror_tvd on all four surfaces. Unit tests cover the merge and TVD computation.

PROP-01: Held-out temperature scaling

• Motivation: The headline ECE gain (0.075 → 0.033, $T^* = 1.35$) was fitted and scored on the same 50 items (--auto-temperature on results_calibration_cloudrun.json).
• Hypothesis (H1): Under 5-fold cross-validation, temperature scaling still lowers ECE relative to $T = 1$, but by less than the in-sample figure. Pre-registered success line: the cross-validated ECE reduction is ≥ 25% and its 95% bootstrap interval excludes 0.
• Design:
  • 5-fold CV on the 50-item calibration suite, fitting $T^*$ on 4 folds and scoring the 5th.
  • Repeat on the 231-item JevBench v1.3.1 public split (benchmarks/jevbench/), which is large enough for a meaningful 10-bin ECE.
  • Also fit on JevBench and test on the calibration suite, to check transfer across suites.
• Metrics: 10-bin ECE, Brier score, NLL. Report the fitted $T^*$ for each fold to show how stable it is.
• Decision: If H1 fails, stop reporting in-sample ECE as a headline result and keep $T = 1$ as the default.
• Cost: No GPU needed. It re-scores existing receipts (--from-receipt), plus a small code change to add a --cv-folds option.

PROP-02: End-to-end IDC cascade

• Motivation: The 98% (49/50) cascade result is entropy-only (results_calibration_cascade_normalized.json). The IDC page describes a combined gate (entropy or mirror disagreement), but it has never been run end to end.
• Hypothesis (H2): At a matched escalation rate (±3 points), a gate using null-prior + Dual-Mirror + (held-out) temperature, escalating on $\tilde{H} \ge \tau_H$ or $\text{TVD}_{\text{mirror}} \ge \tau_{\text{TVD}}$, achieves higher cascade accuracy than the entropy-only gate. Alternatively, at matched accuracy it escalates fewer items.
• Design:
  • Choose $\tau_H$ and $\tau_{\text{TVD}}$ on a tuning split; report on a held-out split.
  • Stage 2 is gemini-3.8-flash with prior forwarding, as in EXP-05b.
  • Use a suite of ≥ 200 items: the 50-item suite plus JevBench, plus the PROP-04 items once available.
  • Compare: (a) no cascade, (b) entropy-only, (c) IDC + entropy gate, (d) IDC + entropy-or-TVD gate.
• Metrics: Accuracy, escalation rate, the accuracy-vs-escalation curve (area under the curve), p50/p95 latency, USD per 1,000 decisions, and wrong answers that exited early.
• Decision: If (d) doesn’t beat (b) on the held-out split, describe Dual-Mirror as a diagnostic tool rather than a production gate.

PROP-03: Same-canvas vs separate-pass mirror

• Motivation: On one canvas, the forward and reversed slots can see each other. In perm_08, adding the reversed slot moved the forward reading from 99.9% to 71.3%. The two readings may therefore be coupled, which could either hide disagreement (the slots copy each other) or manufacture it.
• Hypothesis (H3): Mirror TVD measured on a shared canvas is systematically lower than TVD between two separate single-slot passes (forward and reversed orders). Pre-registered: a paired Wilcoxon test over items gives p < 0.05, and the median ratio is < 0.8.
• Design:
  • For each item, run: (i) forward-only single slot, (ii) reversed-only single slot, (iii) the shared mirror canvas. Compare TVD(i, ii) with TVD(fwd, rev) from (iii).
  • Add a condition where the reversed slot’s instructions forbid looking at the other slot. This tests whether an instruction reduces coupling.
  • Suite: the 16 permutation_suite.jsonl items, plus the PROP-04 items.
• Metrics: Paired TVD difference, agreement between each gate’s decisions (Cohen’s κ), and latency.
• Decision: If the shared canvas masks disagreement, --dual-mirror should document the coupling or offer a two-pass mode for high-stakes slots.

PROP-04: Order sensitivity on real ChaosNLI

• Motivation: The four “ambiguous” items in EXP-13 are authored (“ChaosNLI-style”). The EXP-04 8× entropy figure rests on 3 items per group. Neither is enough to claim that uncertainty signals track human disagreement.
• Hypothesis (H4): Across ≥ 300 real ChaosNLI items (SNLI, MNLI, and αNLI subsets, each with ~100 annotator labels), the entropy of the human label distribution correlates positively with (a) single-pass $\tilde{H}$, (b) Mirror TVD, and (c) cyclic JSD. Pre-registered: Spearman ρ ≥ 0.3 for each, with a 95% interval excluding 0. Secondary: adding (b) or (c) to (a) improves the prediction of human entropy (partial ρ > 0).
• Design:
  • Sample items stratified by human entropy.
  • Run bench-permutation-style evaluation: every cyclic order, the null prior, and the Dual-Mirror canvas.
  • Add the items to a new benchmarks/chaosnli_permutation_suite.jsonl, recording the source dataset and version.
• Metrics: Spearman ρ, soft-label TVD against the human distribution, order-flip rate by human-entropy tercile, and the fraction of items with $\tilde{H} < 0.16$ that still flip under reordering (the “hidden toss-up rate”).
• Decision: Replaces the EXP-04 / EXP-13 n = 3–4 figures in the docs. If ρ is weak, soften the claim that entropy “tracks human disagreement”.

PROP-05: Mirror merge rule

• Motivation: Production merges the forward and reversed readings as 70% / 30% with a forward-priority winner rule. On the 50-item suite, Dual-Mirror gave worse Brier (0.221 vs 0.186) and worse ECE (0.081 vs 0.075) than the baseline.
• Hypothesis (H5): A symmetric merge, either the arithmetic mean $\tfrac12(p^{\text{fwd}} + p^{\text{rev}})$ or the normalized geometric mean, gives a Brier score no worse than the single-slot baseline (non-inferiority margin 0.01) and better than the 70/30 rule.
• Design:
  • Re-score existing Dual-Mirror receipts offline. The raw forward and reversed probabilities must be stored in the receipt, so add them if they’re missing.
  • Compare merges: 70/30 + forward priority, 50/50 arithmetic, geometric, and “forward only” (the mirror used for gating only).
• Metrics: Accuracy, Brier score, ECE (cross-validated where anything is fitted), and flip rate against the single-slot answer.
• Cost: No GPU needed once the receipts store both readings.

PROP-06: Null-prior strength and label-aware priors

• Motivation: Null-prior de-biasing is the strongest IDC component on the 50-item suite (Brier 0.186 → 0.149). But on the ordering suite, at α = 0.75, it doubled the cyclic flip rate (12.5% → 25%). The prior is also measured with neutral Option 1…K labels rather than the real option text.
• Hypotheses:
  • H6a: An intermediate α (chosen by cross-validation from {0, 0.25, 0.5, 0.75, 1.0}) improves Brier over α = 0 without increasing the cyclic flip rate by more than 5 points.
  • H6b: A label-aware null prior (content-free input, but the template’s real option labels) lowers the flip rate compared with the neutral-label prior at the same α.
• Design: Sweep α on the calibration suite and the ordering suite (plus PROP-04 items), and compute a label-aware prior for each template.
• Metrics: Brier score, ECE, accuracy, cyclic flip rate, and cyclic JSD.

PROP-07: Wording invariance

• Motivation: IDC only addresses option order. Rephrasing an option label or description (“benign” vs “authorized business activity”) may shift decisions just as much.
• Hypothesis (H7): On ambiguous items, the paraphrase flip rate (answers changing under a meaning-preserving rewording of the options) is at least as large as the order flip rate. On clear-cut items, both are below 2%.
• Design:
  • Generate 3 paraphrases per option set (Gemini, human-checked), covering the 16 ordering items plus a 100-item sample of the calibration suite and PROP-04 items.
  • Start from bench-jev’s existing paraphrase-consistency support.
  • Optionally test a “paraphrase mirror”: a second slot on the same canvas using the paraphrased labels.
• Metrics: Paraphrase flip rate, paraphrase TVD, and their correlation with order TVD.
• Decision: If wording sensitivity is large, broaden what “invariant” covers in IDC (or say clearly that it doesn’t).

PROP-08: Missing-context detection

• Motivation: The docs use the example of a phishing message whose sender domain is omitted. It’s unknown whether dgem’s uncertainty signals actually rise when a decisive field is removed, or whether the model stays confidently wrong in both orders.
• Hypothesis (H8): When a decisive field is removed, $\tilde{H}$ or Mirror TVD rises above the gate threshold on ≥ 50% of items that were answered correctly with the field present.
• Design:
  • Build ~60 paired items where one field determines the label (phishing with/without sender_domain, refunds with/without order status, agent steps with/without the task goal).
  • The gold label for the “field removed” version is “cannot determine” or a known human split.
  • Evaluate with and without IDC.
• Metrics: Detection rate, false-certainty rate (> 90% confidence on items with the field removed), and change in uncertainty between each pair.
• Decision: If detection is poor, recommend explicit “is required information present?” slots in templates instead of relying on uncertainty.
• Running (EXP-40, 2026-10-21): Extended in pkg/decisionindex with an opt-in co-denoised evidence slot (EvidencePrompt / <q>_evidence) placed ahead of the choice slot on the same diffusion canvas (think: 0, 1 pass), so missing or occluded discriminative evidence is read explicitly alongside raw slot hesitation.

PROP-09: Base-rate (label-shift) adaptation

• Motivation: A score calibrated on a benchmark can be wrong for a deployment where class frequencies differ (1% fraud vs 50% fraud). There’s currently no base-rate option.
• Hypothesis (H9): Using unlabeled target traffic, EM-based prior re-estimation (Saerens et al., 2002) lowers Brier on a skewed target split compared with no adjustment. Pre-registered: ≥ 10% relative Brier reduction on at least 2 of 3 skewed splits.
• Design:
  • Build skewed resamples of banking77_26.jsonl / clinc150_26.jsonl / the guardrail items (e.g., 5%, 50%, and 90% positive class).
  • Run once, then re-weight probabilities offline.
  • If it works, add a --class-prior option (explicit base rates) alongside the EM estimate.
• Metrics: Brier score, ECE, accuracy, and error in the estimated prior.
• Running (EXP-40, 2026-10-21): Implemented in pkg/decisionindex via PriorLogits and PriorWeight ($\alpha \in [0, 1]$) with a pre-fusion hesitation gate: raw normalized slot entropy $\tilde{H}_{\text{raw}}$ is evaluated before log-linear Product-of-Experts prior fusion ($\log P_{\text{fused}}(k) \propto \log P_{\text{token}}(k) + \alpha \cdot s_{\text{prior}}(k)$), and prior fusion is suppressed whenever $\tilde{H}_{\text{raw}} \ge \tau_{\text{gate}}$ so a confident retrieval or base-rate prior cannot mask model hesitation on out-of-distribution inputs.

PROP-10: Run-to-run and revision noise floor

• Motivation: The IDC comparisons in the docs mix runs from different days and Cloud Run revisions (/mnt/gcs/dgemma on 2026-09-20 vs diffgemma-26b-a4b-it-q4 on 2026-09-23). Without a noise floor, a 2-point accuracy difference on 50 items can’t be interpreted.
• Hypothesis (H10): Repeating the 50-item calibration suite 3× against one revision, and 1× against each of two revisions, gives accuracy variation of ≤ 2 points and Brier variation of ≤ 0.01 within a revision. Differences between revisions are reported separately.
• Design: Run the same items, same flags, and -w 1 for determinism where possible, on Vertex AI (<legacy-l4-endpoint>) and Cloud Run.
• Metrics: Standard deviation within a revision and differences between revisions for accuracy, Brier score, ECE, and per-item answer agreement.
• Decision: Sets the minimum effect size that later entries (PROP-01–PROP-09) must exceed to count as real.

---

PROP-11: Letter collision in the mirror

• Motivation: In EXP-14 the forward answer changed 65% of the time when the forward and reversed slots chose the same letter (which then means different options), versus 9% otherwise. See EXP-14, “Why the mirror degrades”.
• Hypothesis (H11): The mirror’s damage to the forward reading comes from the two slots sharing letters with different meanings, not from the presence of a second slot.
• Design: JevBench (231), Vertex G4, baselines at the start, middle and end of the session. Conditions: (a) single slot; (b) two identical copies in the same order; (c) forward + reversed with letters (current); (d′) forward + reversed where the reversed slot is sent as a score question so the server labels it 1, 2, 3… (client-only way to remove shared letters); (e) reversed slot placed before the forward slot. Server-side condition (d), explicit labels that keep each option’s original letter, is deferred until the server accepts caller-supplied labels.
• Metrics: forward-slot accuracy vs. (a); same-letter rate; how often the forward answer changes vs. (a); breakdown by question type and number of options.
• Decision (pre-registered): H11 is supported if (b) and (d′) are within ±2 items of (a) while (c) stays more than 3 items below (a). If (b) is also degraded, a second slot hurts by itself and the same-canvas mirror should be dropped in favour of PROP-12.
• Cost: about 1,200 requests. Client-only (--mirror-mode).
• Amendment (2026-09-25, before any PROP-11 data): the parity probe measured a JevBench noise floor of about ±3 items (baselines 187–192), so the ±2 band would fail on noise alone. Revised rule: run at least 3 baselines in the session; “within noise” means inside their min–max range; “degraded” means more than 3 items below the range minimum. Added a per-item test that is insensitive to run-to-run drift: among items where the two slots pick the same letter for different options, the forward answer should change (vs. the baseline majority answer) significantly more often than on other items, and this gap should vanish in copy and reversed-digits. Questions with more than 9 options fall back to letters in reversed-digits (0 such items in JevBench).

PROP-12: Separate-pass mirror

• Motivation: Replaces PROP-03. A reversed reading in its own request cannot copy the forward slot.
• Hypothesis (H12): Forward vs. reversed disagreement from two separate passes adds error detection beyond hesitation (entropy).
• Design: JevBench and the 50-item suite: one pass in original order, one pass with options reversed. Record per-item disagreement (TVD) and argmax agreement.
• Metrics: partial correlation of disagreement with errors, controlling for hesitation; AUROC for error detection of hesitation alone vs. hesitation + disagreement.
• Decision: Supported if disagreement adds detection (partial correlation > 0 with a 95% bootstrap interval excluding 0). Forward accuracy equals baseline by construction; the cost is a second pass.
• Pre-registered design (2026-09-26, before data): JevBench 231 only (the 50-item suite has too few errors for detection metrics, and bench-calibration has no option-reversal flag). Vertex G4, one session, order: forward F1, reversed R1 (bench-jev --flip-options), forward F2, reversed R2. Primary: partial Spearman correlation between forward/reversed TVD and forward-pass error, controlling for forward hesitation, averaged over the four F×R pairings, 95% bootstrap CI (2,000 resamples). Noise control: the same statistic for F1 vs F2 (two forward passes), since Vertex is not fully deterministic; order disagreement must beat it. Secondary: 5-fold CV AUROC of logistic error models (hesitation vs. hesitation + TVD), and accuracy of the averaged forward+reversed distribution vs. forward alone. About 924 requests.
• Cost: about 460 requests.

PROP-13: Mirror-aware cascade

• Closed (2026-09-28), not run. EXP-15–EXP-17 showed the mirror signal adds little over hesitation: a separate-pass mirror improves cross-validated error detection by at most 0.016 AUROC at twice the cost, and the same-canvas mirror costs accuracy unless labels are digits. Exposing it on every surface (PROP-00) or gating the cascade on it (PROP-13) is not worth the complexity; hesitation remains the production signal. Reopen only if a cheaper, uncoupled order signal appears.
• Motivation: Unblocks PROP-02 once an uncoupled mirror exists.
• Hypothesis (H13): A gate of hesitation ≥ τ or disagreement ≥ τ′ beats entropy alone at the same hand-off rate.
• Design: Offline, using PROP-12 readings and the existing Gemini-on-every-item receipt; choose τ, τ′ on half the items and report on the other half.
• Decision: Supported if accuracy is higher at a matched hand-off rate (±3 points), or hand-offs are fewer at matched accuracy.
• Cost: offline.

PROP-14: In-context vs blank-question prior

• Motivation: On JevBench the baseline picks position A less often than the correct answer is there (62 vs 70), so dividing out a blank-question slot-A habit overcorrects.
• Hypothesis (H14): A position prior estimated from real items with rotated options (PriDe-style, Zheng et al. 2023) avoids null-prior’s overcorrection.
• Design: Rotate options on held-out JevBench items (60) and the 50-item suite; estimate per-position pick rates independent of content; compare blank-question prior, in-context prior, and none, with cross-validation.
• Decision: Supported if the in-context prior is no worse than baseline on JevBench and at least as good as the blank-question prior on the 50-item suite.
• Cost: about 900 requests.

PROP-15: Correction strength by question type

• Motivation: Extends PROP-06. Three yes/no items broke under null-prior on JevBench.
• Hypothesis (H15): Yes/no and lettered choices need different correction strengths (possibly none for yes/no).
• Design: Sweep α ∈ {0, 0.25, 0.5, 0.75, 1} separately by question type and number of options on both suites, with cross-validation (offline re-scoring from receipts’ raw distributions).
• Decision: Supported if a per-type setting is never worse than baseline on either suite.
• Cost: offline.

PROP-16: Slot names are part of the prompt

• Motivation: Renaming the mirror slot from __mirror_rev to __rev changed JevBench from 145 to 163 correct.
• Hypothesis (H16): Slot ids act as instructions; meaningful or loaded words change answers more than neutral ids.
• Design: Single-slot and two-slot schemas with ids: decision, q1, random strings, and loaded words (e.g. “mirror”, “check”).
• Decision: If the effect exceeds 2 items on JevBench, add a template style rule for slot ids.
• Cost: about 1,200 requests.
• Pre-registered design (2026-09-26, before data): JevBench 231, Vertex G4, one session. Single slot: 3 baselines with id decision (noise band = min–max), then ids q1, x7k2q (random), mirror, check. Two slots (--dual-mirror --mirror-mode copy, so no letter collision): suffixes __rev, __mirror_rev, _b. An id “matters” if its correct count is more than 3 items outside the baseline band, or (two-slot) more than 3 items away from copy+__rev. Secondary: per-item agreement with the baseline majority answer. About 2,300 requests (revised up from 1,200).

PROP-17: Calibrated agent context pre-compiler (internal pilot)

• Motivation: Agents re-send large context windows (tool schemas, history, retrieved documents) every turn. A single dgem pass could grade every candidate block at once and use hesitation to downgrade uncertain blocks to a one-line stub instead of dropping them. The pilot’s design and data are internal for now; this entry holds the number and the decision rule.
• Hypothesis (H17): On labelled agent turns, one pass keeps at least 95% of the blocks the gold labels mark as required (full or summary), removes at least 50% of candidate tokens, and flags planted prompt-injection blocks with recall ≥ 0.9; hesitation-driven “stub instead of drop” reduces required-block misses compared with taking the most likely action.
• Design: ~40 labelled turns across several domains, 8 candidate blocks each (tool schemas, history turns, retrieved documents, notes), gold action per block from one annotator with a second model labelling independently to measure agreement; production serving image, samples: 1, neutral slot ids (EXP-16).
• Metrics: required-block recall (primary), token savings, quarantine precision/recall, rescue rate (required blocks kept as stubs only because of hesitation), agreement with the gold labels.
• Decision: If H17 holds, build a larger labelled set from real agent traces and consider publishing the templates; if recall < 95%, test a two-stage variant (coarse keep/drop, then action) before continuing.
• Cost / Dependencies: ~200 requests plus labelling; none.

PROP-18: Validate image metrics beyond EXP-09

• Motivation: The EXP-09 re-baseline (phase 1) found three things:

  • The model localizes on 11 synthetic UI fixtures (mIoU 0.60 vs 0.33 for a fixed box).
  • The sub-bin and occlusion-entropy claims do not hold.
  • Boxes are unstable under flips and padding (consistency IoU 0.34–0.53).

  The suite is too small and too synthetic to say which image readouts can be trusted.

• Hypothesis (H18): For each aspect:

  • Accuracy beats the best image-free baseline, with a paired 95% interval above 0.
  • Per-slot hesitation flags errors with AUROC ≥ 0.75.
  • Results hold within the measured noise floor under transforms that should not change the answer.

• Design (phased; each phase is analyzed before the next is fixed):

  1. Done. Harness fixes, image-free baselines, repeats, probe variants (EXP-09 re-baseline).
  2. Gemini 3.x reference and judge (gemini-3.8-flash, gemini-3.7-flash). Gemini gives its own answers on the same items; it also judges dgem’s annotated overlays edge by edge. First measure the judge against ground truth (agreement, Cohen’s κ) on the synthetic and RefCOCO subsets; only then use it to label images without ground truth.
  3. Data and references.
     • A generated synthetic set of several hundred images, varying: aspect ratio (native, not just padded), horizontal position, size, occluder coverage 0–100% per edge, distractors, blur/noise/JPEG.
     • RefCOCO/COCO and ScreenSpot/RICO subsets: committed manifests; images fetched by script.
     • OWLv2, Grounding DINO and SAM boxes as references, computed on a short-lived GCE GPU VM that the script tears down when it finishes.
     • Test client-side preprocessing (pad or stretch to square) against the aspect-ratio failure.
  4. bench-vision multi-aspect harness. Typed aspects: location (box and 3×3/5×5 grid), presence and count, attributes and relations, quality/occlusion/legibility, and the domain templates (ui_design_review, pcb_defect_triage, kyc_document_audit). Each aspect has its own ground-truth source, metric and Gemini comparison. A small image gate goes into the regression matrix.

• Metrics:

  • Primary: per-aspect accuracy (or IoU) lift over image-free baselines.
  • Then: hesitation→error AUROC, transform consistency, Gemini agreement and κ, latency and cost against Gemini.

• Decision:

  • Aspects that pass get documented as supported, with their measured limits.
  • Aspects that fail are documented as unsupported.
  • If padding/flip instability persists on real images, the localization template gets a preprocessing recommendation or a coarse grid output.

• Cost / Dependencies:

  • About 5k dgem requests and 2k Gemini calls per full pass.
  • About 1 GPU-hour of GCE for the detector references.
  • Phase 2 needs Vertex Gemini access.

PROP-19: Cross-slot coupling on mixed-polarity yes/no questions

• Source: Strands (EXP-23, section 5). Also filed as a bug.
• Motivation: On 51 tool-call review cases with two yes/no questions per request, dgem’s two answers carried the same label in 34–40 cases (gold: 7) and it scored 58% (Strands Decider 86%). Asked one question per request, dgem scored 100 of 102; through dgem systemone serve --noul-mode choice, 95 of 102. typed-decisions (200 yes/no pairs in one request) shows no coupling, so the trigger is related wording with opposite answers.
• Hypothesis (H19): Coupling comes from rendering both nouls as bare yes/no labels on one canvas. Rendering them as two-option choices whose labels carry the true/false descriptions brings equal-answer cases on gate_mixed_noul to ≤ 12 of 51 without moving JevBench or typed-decisions beyond the noise floor.
• Design: Matrix v2 T1 plus gate_mixed_noul_single, on a canary next to production, 3 runs. Arms: current rendering; noul-as-choice server-side; noul-as-choice plus neutral distinct letters per slot; one read per noul (latency reference). Develop on a second, independently written gate suite; score gate_mixed_noul once.
• Metrics: equal-answer cases (primary); gate accuracy; JevBench and typed-decisions agreement vs production; latency.
• Decision: Ship the best arm that passes T1 as the server default; add the coupling count to the matrix gates.
• Cost / Dependencies: about 5k requests per arm; a server change in structured_server.py.

PROP-20: Shipped temperature per question type and domain

• Source: Laya (held-out T ≈ 1.5 global, 2–3 for yes/no; T ≈ 3.3 on typed-decisions), Strands (EXP-23: one T ≈ 1.5 brings pooled ECE from 0.069 to 0.028; yes/no wants 1.6–3.5; raw dgem is 74% correct at ≥ 0.9 on XNLI), DI (improvement plan track T1: overconfidence grows with option count).
• Motivation: dgem’s confidence ranks errors well (higher AUROC than both comparison models on most suites) but its absolute level is too high off the development suites, so a ”≥ 0.9 means act” policy over-automates.
• Hypothesis (H20): Temperatures per question type and option-count bucket, fitted by k-fold on the matrix development suites and served as defaults, cut ECE by ≥ 30% on every T1 suite and on the frozen sets (scored once), with AUROC unchanged (± 0.01).
• Design: Fit on T1 receipts (5-fold by case); evaluate on held-out folds, then on T2 frozen sets after shipping. Compare global T, per-type T, per-type × bucket T. Per-policy override via template field (four-surface parity).
• Metrics: ECE10 (primary), Brier, NLL, AUROC, coverage and accuracy at ≥ 0.9.
• Decision: Ship the simplest variant within 0.005 ECE of the best.
• Cost / Dependencies: offline on existing receipts, then one T1 + one T2 run.

PROP-21: Fine-tuning DiffusionGemma: LoRA on the letter readout

• Source: Laya (a 421M encoder fine-tuned on 6,000 typed-decisions labels beat zero-shot dgem on that domain), Strands (EXP-23: a 1.9B LoRA model matches dgem on its training tasks, 85.9% vs 85.2%, and trails by 10 points elsewhere), Clef (a trained routing head leads on intent and taxonomy benchmarks).
• Motivation: Fine-tuning is the main accuracy lever of every competitor; dgem’s lead is its base model’s breadth. The diffusion objective matches inference exactly (unmask the answer slots), and a training canvas can hold several questions with independent labels, which is the direct fix for PROP-19.
• Hypothesis (H21): A rank-16 LoRA on attention and the shared expert (router frozen), trained with the masked denoising loss on answer slots plus a proper scoring rule and a KL term to the frozen model, on Strands Decider’s public training corpus plus synthetic mixed-polarity canvases: (1) raises accuracy on the train exposure pool from 85% to ≥ 90%; (2) keeps JevBench ≥ 190 of 231; (3) brings gate_mixed_noul equal answers to ≤ 12 of 51; (4) lowers the fitted yes/no temperature below 1.5.
• Design: Same corpus as the competitor so the base model is the only difference. BF16 training (about 52 GB of weights, 80 GB-class GPUs), merge, re-quantize to NVFP4, refit T. Spike first: vLLM structured mode with a merged checkpoint on the G4 profile. Never train on JevBench or frozen sets; differences under 10 JevBench items are unresolved.
• Metrics: the four predictions; full T1 against production; frozen sets once, after the decision.
• Decision: Ship only if (2) and T1 pass; otherwise publish the result and keep zero-shot as default.
• Cost / Dependencies: training compute (hours on 8× H100-class), one canary; PROP-19 and PROP-20 define the baselines.

PROP-22: Distil dgem into a small student

• Source: Strands (its recipe distils from a frozen 4B teacher; about 1 hour on 8× H100).
• Hypothesis (H22): Replacing the teacher in Strands Decider’s recipe with dgem’s temperature-scaled soft labels gives a 2B student that beats v19 on the unseen exposure pool by ≥ 3 points at equal latency.
• Design: One recipe change, same data and seeds; compare with v19 and with dgem on matrix TC.
• Decision: If it holds, publish the recipe and offer the student as a CPU-capable option.
• Cost / Dependencies: dgem labelling of the corpus (one batch job), training compute; PROP-20 first.

PROP-23: Nondeterminism at samples=1

• Source: Laya (5–7% answer changes between identical requests at seed=42), Strands (6.4%, EXP-23). Both competitors are bit-identical across runs.
• Hypothesis (H23): The changes come from batch composition in vLLM (requests batched with different neighbours), not from sampling; serving with batch-invariant kernels or a fixed per-request seed path removes most of them at ≤ 10% throughput cost.
• Design: Identity runs at 1 vs 16 concurrent workers; then the candidate fixes on a canary.
• Metrics: answer agreement between identical runs (primary), throughput.
• Decision: Ship if agreement ≥ 99% and throughput ≥ 90% of current.
• Cost / Dependencies: about 3k requests per arm; vLLM options.

PROP-24: Wide-option accuracy (6–26 options and beyond)

• Source: DI (improvement plan findings F1 and F8: 6–26 options in one read already score 36% on the index benchmarks), Clef (leads intent and taxonomy benchmarks by 20–55 points), Strands (EXP-23: banking77 with 30 options, dgem 80% through the adapter vs 87% natively).
• Motivation: Letter readout caps a slot at 26 options and loses accuracy well before that.
• Hypothesis (H24): On di_wide and intents_systemone, hierarchical routing with calibrated finalists and option descriptions in the prompt recovers ≥ 5 points over the current bracket routing. A pointer readout (PROP-21 route B) is the longer-term fix.
• Design: Owned with the Decision Index improvement plan; develop on validation data only; T1 plus Decision Index dev set.
• Decision: Ship in the adapter if T1 passes.
• Cost / Dependencies: adapter changes; no training.
• Running (EXP-41, 2026-10-21): Added two-stage ShortlistFunc routing in pkg/decisionindex (stage-1 prior/embedding shortlist down to $M \le 6$ candidates with pairwise contrastive option descriptions, reducing $K > 26$ routing from $\lceil K/20 \rceil + 1$ passes to 1 diffusion pass when retrieval priors are present).

PROP-25: NLI “neutral” pull

• Source: Laya (XNLI: dgem predicted “neutral” for 659 of 1,500 contradictions and 714 of 1,500 entailments), Strands (EXP-23: XNLI is dgem’s smallest lead, 70.0% vs 63.7%, with ECE 0.227).
• Hypothesis (H25): Clearer option descriptions or a per-label prior estimated in context raise XNLI validation accuracy by ≥ 3 points without changing the calibration suite’s NLI items.
• Design: XNLI validation only for development; one scored test run after shipping.
• Cost / Dependencies: about 5k requests.

PROP-26: First-shown bias on small option sets

• Source: Laya (emotion, K=6: dgem picks the first-shown option 41.5% of the time vs 30.5% gold), Strands (EXP-23: 45.0% vs 30.5% on the same suite; Strands Decider 47%).
• Hypothesis (H26): The in-context prior of PROP-14 brings first-shown picks within 5 points of the gold rate at K ≤ 6 without raising order flips.
• Design: Option-order suite plus PROP-14’s items; extends PROP-14.
• Status: Parked with PROP-14. Under hesitation-gating, order bias is monitored per release (the matrix’s option-order suite) rather than corrected; reopen only if the monitored first-shown rate rises.
• Cost / Dependencies: PROP-14.

PROP-37: Contact-sheet multi-crop visual verification and co-denoised evidence slot

• Source: DI, Clef (EXP-22 and EXP-26: single-image categorical reads trail on fine-grained sibling classes and dense small text because 896×896 global resizing blurs variant badges and numeric callouts).
• Motivation: When distinguishing visually similar sibling classes (e.g., product variants, component revisions, or document versions), downsampling a single full scene into SigLIP’s 896×896 frame loses sub-centimeter text and badge differences, and text-only option labels lack visual reference anchors.
• Hypothesis (H34): On a held-out 200-item fine-grained visual sibling verification split (paired query crop + top-5 reference crops composited into a single $3 \times 2$ 896×896 contact sheet, with a co-denoised evidence_feature slot on the same canvas (PROP-08) and pre-gated retrieval prior fusion (PROP-09)), single-pass top-1 accuracy exceeds single-image text-option readout by ≥ 10 points (95% paired bootstrap CI excluding 0) and 5-fold cross-validated ECE decreases by ≥ 20% without increasing p50 server denoise time by more than 15 ms.
• Design:
  • Compare 4 arms on the same Vertex G4 revision in one session (with 3 baseline repeats to establish the session noise floor):
    1. (a) Single query image + text-only candidate descriptions (K=6, including none_of_the_above).
    2. (b) $3 \times 2$ contact-sheet composite (Panel 0: query crop; Panels A–E: reference crops) + pairwise contrastive option descriptions (PROP-24).
    3. (c) Arm (b) + same-canvas co-denoised evidence_feature slot (PROP-08) using neutral slot IDs (EXP-16).
    4. (d) Arm (c) + pre-gated retrieval prior fusion (PROP-09, gated off when raw normalized hesitation $\tilde{H}_{\text{raw}} \ge 0.16$).
• Metrics: Top-1 accuracy and abstention recall on out-of-catalog items (primary); 5-fold CV 10-bin ECE, multi-class Brier score, error-detection AUROC of $\tilde{H}_{\text{raw}}$, and p50/p95 server denoise latency (secondary).
• Decision: Ship templates/multimodal/contact_sheet_verify.json.tmpl, EvidencePrompt, and pre-gated PriorLogits in pkg/decisionindex if (d) beats (a) by ≥ 10 accuracy points on the validation split and held-out Brier does not regress when the stage-1 retrieval top-1 candidate is wrong.
• Cost / Dependencies: PROP-08, PROP-09, PROP-18; ~800 Vertex G4 requests; adapter and template changes only (no training).

PROP-38: Set-of-Marks discrete next-hop UI navigation

• Source: DI (EXP-22 and EXP-24: raw [0, 999] coordinate binning on ScreenSpot scores 12% vs Gemini 93%, whereas categorical single-pass image readouts lift +0.40 to +0.52 over baseline with hesitation AUROC 0.61–0.85).
• Motivation: Visual UI agents and workflow pathfinders do not need continuous pixel regression when a lightweight DOM / accessibility tree or upstream detector already enumerates candidate interactive elements; they need a calibrated discrete choice over candidate marks [A..F] plus a loop/stuck detector.
• Hypothesis (H35): Reformulating UI localization on a 200-item ScreenSpot / web-UI slice from 21-bin [ymin, xmin, ymax, xmax] coordinate readout (EXP-09) to Set-of-Marks discrete [A..F] candidate selection (next_target_mark, element_state, loop_or_error_detected, goal_progress_score in 1 pass) raises single-pass target hit rate from 12% (EXP-22) to ≥ 75% on dgem alone, and ≥ 90% with a hesitation-gated cascade ($\tilde{H}_{\text{raw}} \ge 0.16$) handing ≤ 25% of steps to gemini-3.8-flash.
• Design:
  • Overlay up to 6 candidate bounding boxes labelled [A]–[F] on each screenshot (benchmarks/ui_pathfinder_suite.jsonl, dev/test split) and evaluate templates/multimodal/ui_pathfinder_som.json.tmpl in a single 4-slot pass (think: 0, layout: document_first, neutral slot IDs).
  • Compare against EXP-22/EXP-24 coordinate readout and standalone gemini-3.8-flash on the same items.
• Metrics: Target element hit rate (primary); loop/dead-end detection F1, hesitation error-detection AUROC, cascade hand-off rate, and p50 step latency (secondary).
• Decision: Ship dgem bench-ui-pathfinder and recommend Set-of-Marks discrete choice in docs/policies/images.md in place of raw coordinate binning for UI navigation policies.
• Cost / Dependencies: PROP-18; ~600 Vertex G4 requests + ~150 gemini-3.8-flash cascade calls.

PROP-39: Multimodal entity resolution and pack-size equivalence

• Source: DI, Clef (PROP-24 wide-option taxonomy routing; PROP-08 explicit missing/conflicting evidence readout).
• Motivation: High-cardinality catalog deduplication and entity resolution fail when text embeddings conflate multipacks or size variants (1×100ml vs 6×100ml) or miss counterfeit packaging changes, and pairwise LLM verification across $10,000+$ items is prohibitively slow.
• Hypothesis (H36): On a 200-pair multimodal entity-resolution suite (benchmarks/multimodal_dedup_suite.jsonl, stratified across exact duplicates, pack-size/flavor variants, counterfeits, and distinct items), a 4-slot single-pass readout (entity_relation, pack_quantity_match, visual_evidence_cue, merge_action) improves macro-F1 over cosine-similarity thresholding by ≥ 12 points, and gating pairwise merges with a transitive graph triangle-consistency check ($P(A \equiv B), P(B \equiv C), P(A \equiv C)$) reduces false-positive cluster merges by ≥ 50%.
• Design:
  • Stage-1 embedding retrieval selects top-$K$ candidate pairs; Stage-2 dgem evaluates side-by-side image crops + structured attributes in 1 pass (templates/multimodal/entity_resolution_dedup.json.tmpl).
  • Any 3-cycle where $\max_{e} P(\text{duplicate}) > 0.80$ and $\min_{e} P(\text{duplicate}) < 0.25$ is flagged as a transitive contradiction and escalated alongside slots with $\tilde{H}_{\text{raw}} \ge 0.16$.
• Metrics: Pairwise macro-F1 and false-merge rate at 95% precision (primary); pack-size mismatch recall, triangle-contradiction catch rate, and 5-fold CV ECE (secondary).
• Decision: Ship dgem bench-multimodal-dedup and transitive triangle-consistency checking in pkg/decisionindex if false-merge rate drops by ≥ 50% over raw argmax pairing.
• Cost / Dependencies: PROP-08, PROP-24; ~600 Vertex G4 requests.

PROP-40: Sliding-window live checklist and multi-agent swarm lock arbitration

• Source: DI (EXP-01 and EXP-19 multi-slot joint reads up to 8 slots per pass; EXP-10 listwise 12-slot canvas evaluation).
• Motivation: Real-time meeting/inspection assistants and multi-agent coding orchestrators cannot block on multi-second autoregressive LLM turns to track checklist state transitions (unaddressed → in_progress → verified) or arbitrate concurrent file/tool lock collisions across parallel worker agents.
• Hypothesis (H37): Packing 8 checklist or lock-arbitration slots into a single document_first canvas pass (MaxSlotsPerPass = 8, think: 0) over sliding-window transcript + keyframe state (bench-live-checklist, $n=150$ windows) and concurrent 4-agent action proposals (bench-swarm-arbiter, $n=150$ collision scenarios) achieves ≥ 90% slot agreement with standalone gemini-3.8-flash after a $\tilde{H}_{\text{raw}} \ge 0.16$ hesitation gate (≤ 20% escalation), at ≤ 160 ms p50 Stage-1 wall latency on Vertex G4.
• Design:
  • bench-live-checklist: 6 checklist-item state slots (choice: unaddressed, in_progress, completed_verified, contradicted) + speaker_commitment_present (noul) + interrupt_urgency (score 0–3) over 30-second transcript windows and sampled frames (MergeMedia text + image path).
  • bench-swarm-arbiter: 4 per-agent lock/action slots (choice: grant_lock, serialize_after_peer, abort_redundant, escalate_conflict) + file_write_collision (noul) + shared_state_hazard (choice) + winning_agent (choice) + rollback_severity (score 0–3).
  • Compare single-pass 8-slot joint readout against isolated single-slot reads (EXP-19 control) and standalone gemini-3.8-flash.
• Metrics: Per-slot accuracy / macro-F1 against gold annotations and p50/p95 wall latency (primary); false-interruption rate (bench-live-checklist), unflagged write-collision rate (bench-swarm-arbiter), and 5-fold CV ECE (secondary).
• Decision: Ship dgem bench-live-checklist and dgem bench-swarm-arbiter templates and CLI harnesses if Stage-1 p50 stays ≤ 160 ms on Vertex G4 and cascaded accuracy is within 2 points of standalone gemini-3.8-flash.
• Cost / Dependencies: ~900 Vertex G4 requests + ~180 gemini-3.8-flash cascade calls.

---

4. How to Add an Entry

Copy this block, choose the next PROP-XX, and add a row (with its Source) to the table in §2:

### `PROP-XX`: <title>

* **Source:** <studies that raised it: Order, a comparison (e.g. Strands), DI; add to this list instead of duplicating>
* **Motivation:** <which doc claim or open question this tests; link receipts>
* **Hypothesis (HXX):** <falsifiable statement + pre-registered metric and threshold>
* **Design:** <items, conditions, splits/CV, endpoint + revision>
* **Metrics:** <primary first; secondary after>
* **Decision:** <what we change in code/docs if it passes or fails>
* **Cost / Dependencies:** <GPU hours, Gemini calls, code changes, blocking PROPs>