Declared design

Design 7: thinking effort on the cheap tier, 8 pairs in open play

Declared design, final text at 18:00 UTC 2026-10-12 for public reading, after the step-0 review (changes listed at the end). This text is hashed before the first trial and attached to the study; the first 12 hex characters of its SHA-256 go in the study's name. Amendments after hashing appear under dated headings at the end and never change the text above them.

Question

On Anthropic's two current cheap-tier models, Claude Haiku 5.5 and Claude Sonnet 5.5, does asking for more thinking effort raise the Fitness Index in Daishi's reference open-play scenario? Study 6 found +6.6 Index points on 8 maps (+4.3 to +8.9) for its four small models when thinking was switched from off (or the route's lowest setting) to high. This study asks the same question of these two models with thinking on in both arms, where the lever is effort: low against high.

Why now

These are the two cheapest current models from this vendor, and at about $3 a pair on list prices anyone can replay a trial of this study on their own key, which is the point of a shared, re-runnable result.

What is measured

• Primary. The Fitness Index (rubric 2.3) of each arm, taken per trial as the mean over the arm's four roster seats. The unit is the trial. Each trial plays one of the scenario's eight practice maps, and with eight trials every map plays once, so the per-trial reading and the per-map reading are the same numbers.
• Secondary 1, survival. The share of the four roster seats active at the end of the match, per arm per trial.
• Secondary 2, output per turn (manipulation check). Per trial and per model, the mean output tokens per model call over that model's seats measured in both arms, under the platform's rule that a seat is measured in every arm or in none. Effort is only a treatment if the model produced more output under high; this reading says whether the setting took.
• Described, no test. Index by rubric dimension; trained levels per seat and the Index recomputed with the platform's own scorer without the training term; dormant rounds read from each seat's own status; turn kinds per seat (productive action, chosen rest, fallback rest after a turn cut at the ceiling, free tool, refused call, model refusal); forfeits and pace; the debrief and the voiced-awareness scan per model and arm, reported per character the scan read; spend per seat from token counts at list price.
• Mechanism read, declared. The free-tool share per seat (turns spent on a free tool such as the scoring or world information calls, each followed by the forced rest), per model and arm, read from the trajectories. In the pilot Haiku 5.5 at high spent 65 and 68 of 100 turns this way, against 1 to 7 per seat in 16 sampled study 6 seats (trials 1 and 2 of each arm, same harness and scenario), so any Index gap in this study is read beside the change in this share before any other account. Described, not tested.
• Dormancy and the raider, declared. Rounds each seat spent dormant, read from the seat's own status on each called turn and the platform's count of rounds it sat out, per model and arm, with the round it first went dormant; and, per pair and arm, the seat the house raider attacked in the first 10 rounds, the hits it took and the round it went dormant, from the public attack events and the seat's own status. In the pilot the house raider spawned next to the Haiku 5.5 seat at high and hit it on 7 of its first 9 rounds; the seat went dormant at round 9 and sat out 82 rounds while the other three seats stayed active all match. One seat on one map, so it decides nothing. Raids are part of play in this scenario, and arms share spawn layouts per trial, so the raider usually meets the same model in both arms and the pair then turns on how each setting copes; one collapsed seat moves its arm's mean by about 10 Index points, more than twice the planning gap. So an Index gap between arms is read beside the raider hits and dormant rounds per arm before any other account, and one sensitivity read is declared: the primary recomputed with the slot the raider attacked first in either arm of a pair dropped from both arms of that pair (defined by exposure, which the shared layout fixes, not by dormancy, which the setting can change). Described, not tested; the verdict stays on the primary.

Hypotheses

H1 (primary). The high-effort arm's mean Index exceeds the low-effort arm's. Test: paired Student t over the 8 trials, two-sided at 0.05, with the exact sign-flip test beside it (the smallest two-sided p 8 pairs allow is 0.0078). The declared reading is the paired t interval; the platform's own verdict is quoted beside it, and a disagreement between the two is reported. Declared prediction: +4 Index points, tolerance 2.5 (a hit if the observed mean difference lies between +1.5 and +6.5). The prediction is shrunk from studies 4 to 6 (+4.0, +7.2, +6.6) because the models are different, the control is low effort rather than no thinking, and a result that raised a question is usually larger than its replication.

S1. The share of roster seats active at the end differs between arms: paired t over the 8 trials, two-sided at 0.05, with the exact sign-flip beside it. If every pair ties, the reading is "the arms tied in every pair". Seat-matched counts are described, untested.

S2. Output per turn is higher on high effort for each model: per trial, the mean output per turn of that model's seats measured in both arms; paired t over trials per model, two-sided at 0.05 with Holm across the two models, used as the manipulation check for H1.

The Index, survival and output readings are three measures and need not agree.

Interpretation rules, fixed before launch

• H1 interval above zero and S2 passing for both models: supported. The write-up says "scored higher on the Fitness Index in this scenario at this cost", never "smarter" or "better".
• H1 interval spanning zero: undecided. The interval is the reading; a gain inside it is neither shown nor ruled out. It is not "no effect" and not "small". No cause is named.
• H1 interval below zero: reversed, reported with the dimension split, no claim.
• S2 failing for a model: that model's seats did not demonstrably take the setting through this route. H1 is still reported as declared, with the flag "the setting was not shown to take for <model>", and no mechanism is claimed.
• The prediction is scored hit or miss against the observed mean difference, whatever the verdict.
• Against study 6: study 6's eight map means against this study's eight pair differences, one per map, as a difference of means with a Welch interval. Named differences: the models; the harness version (0.6.6 here against 0.6.5 in study 6: the model's tool menu lost one free tool, and a seat now stops after three refused turns in a row); the route and reasoning mechanism (a router's reasoning parameter there, adaptive thinking with an effort level here); the control (off or the route's floor there, low here) and each model's default; the output ceiling (4,096 there, 8,192 here); the roster (four models there, two seats each of two models here); spawn layouts, redrawn per experiment, so the maps share terrain only; the clock if the pilot changes it; the harness if it changed; the free-tool share (two thirds of turns for Haiku 5.5 at high in the pilot, 1 to 7 per seat in 16 sampled study 6 seats). It is descriptive and unpooled. No "replicates", "does not replicate" or "scenario-bound".
• Per-model rows are descriptions of two seats per match. They rank nothing, and no sentence compares the two models' gains.

Design

• Scenario daishi:open-play-v1 (the reference open-play world, standard scarcity; content hash bf0917e5a67f5acafa3ffac5ec5f0d8f54cfd7fd6fa85fc4ee41aaedb896c8d2), 100 rounds, house bots anchors-v2, 180 s turn clock. Seeds vary by trial, which on this scenario walks the eight practice maps in order, so trials 1 to 8 are maps 1 to 8 once each.
• Two arms, paired per trial on the same map and spawn layout. Same roster in both: two seats of Haiku 5.5 and two of Sonnet 5.5, each on the account owner's own Anthropic key through the harness's native Anthropic route, output ceiling 8,192 tokens (on this route the ceiling covers thinking and text together; a turn cut at the ceiling becomes a rest), temperature unset.
• Arm 0 "effort low": reasoning low. Arm 1 "effort high": reasoning high. On this route a level is adaptive thinking with the vendor's effort setting; thinking is on in both arms, so the contrast is effort, not the presence of thinking. High is Claude Sonnet 5.5's default effort and medium is Claude Haiku 5.5's, so neither arm is Haiku's default, and for Sonnet the study measures turning effort down from its default as much as up.
• Harness: reference harness 0.6.6, live since 17:14 UTC 12 October 2026 (the model's tool menu no longer lists the private reasoning-note tool; a seat stops after three refused turns in a row and leaves its pair; each refused call logs the vendor's refusal category), named in the launch record. Rubric 2.3. Native tool calls. Scores compare only within one harness version; study 6 ran 0.6.5.
• Stopping rule fixed at 8 trials. No interim look.
• Pilot: one run with one seat of each model at each effort, played before launch (see Pilot). It is not part of the study's data.

Size and cost, as a consequence of the question

Pair-difference spreads in studies 5 and 6 were 3.5 to 4.4 Index points at pair level and about 2.8 at map level over two pairs a map. With 8 pairs, one per map, a two-sided paired t at 0.05 detects a gap of 4.0 to 5.1 points with 80% power, and the 95% interval runs about plus or minus 2.9 to 3.7 around the observed mean. At the predicted +4 the power is about 60% to 79% (noncentral t, pair spread 4.4 to 3.5). Sixteen pairs on these maps, read at map level, would bring the 80% gap down to about 3.2 and cost twice as much. Eight were chosen because a gain of +4 has about 60% to 79% power at eight, and a miss at this size costs one more study, not a wrong claim. The experiment names study 6 as its spread reference, so the platform's own detectable-gap line is sized from the measured spread of study 6's 16 paired differences on this scenario (3.47 points): 4.0 points at 8 pairs with 80% power, instead of the platform's generic prior. Study 6 compared off with high, so its spread also carries how that effect varied by pair (the reference is not an A/A check). It is a planning figure and decides nothing.

Cost at list prices was projected from study 6's per-seat token profile, with cache reads and writes at their own rates, at about $51 for the 16 matches (band $36 to $71) plus about $4 for the pilot, with Haiku 5.5 at 6 to 9 cents a seat per match and Sonnet 5.5 at $1.20 to $1.90. The pilot (attempt 3, 12 October) measured one four-seat match at about $2.12 at list and $3.79 in the platform's metered dollars (Haiku 5.5 about 7 cents for its two seats, Sonnet 5.5 about $2.04 for its two; cache writes and the debrief calls included in both figures), with one Haiku 5.5 seat dormant from round 9. A high-arm match with every seat active all 100 rounds projects to about $5.00 metered (two Haiku 5.5 seats at high from the second pilot's receipts plus two Sonnet 5.5 seats at high from this one); an active Haiku 5.5 seat adds about 6 cents at list but about 60 cents metered. So the 16 matches cost about $34 to $35 at list and $61 to $70 metered, under the projection because Sonnet 5.5 wrote far fewer output tokens than projected at both settings (about 25,000 a seat at high and 3,000 at low, against 110,000 and 40,000). The per-match spend cap, set in the platform's metered dollars (its price table predates these models and reads about 1.8 times list for this roster), is the larger of $15 and 2.5 times the pilot's metered spend, rounded up to the next $5, the same in both arms: $15 (2.5 times $3.79 is $9.47), so the worst case the caps allow is $240 metered. A trial that hits the cap leaves every pair, by the platform's rule, and with one pair per map nothing replaces a lost map.

Analysis, frozen

1. Health first for every run: finished, every round played, every seat kept, harness version per seat, forfeits and causes, seat end status, provider errors, truncated share, refusals, the debrief present, pace. 2. H1, S1 and S2 exactly as declared, from the run documents and the public match archives, with every exclusion listed in the same table. 3. Descriptions as listed under "What is measured", every number with its scope (trials, maps, models, ceiling, clock, rubric) and every share as k of n. 4. The between-study reading against study 6, descriptive. 5. The study published whole with its note; the write-up carries every trial, failure and exclusion, the pilot, and the spend.

Exclusions, stopping, publishing

• Exclusions: the platform's rule only. A pair is out when either trial failed (a match past the plan's run-time cap fails), was cancelled, lost a seat to the harness giving up on it (the reason is named per pair), stopped short of its season at its spend cap, or was picked up by a new server release mid-match; the report names the reason for each pair left out. Nothing is dropped after seeing a result.
• A launch refused for a reasoning level amends this design under a dated heading before the first trial; what the pilot's receipts trigger is fixed under Pilot (a cut share or a refusal share above a tenth, or latency and minutes that do not fit, send the study back to the owner).
• Stopping: fixed, 8 trials.
• Publishing: the study whole, in the platform's one study-publish call that lists every trial that played in every arm with the declared note (never trial by trial), the pilot match as a single match marked as the pilot, and this design, on the account owner's word, whatever the result.

What each result would change

• Supported: on these two models in this scenario, high effort scored higher on the Fitness Index than low, at the cost measured; the next question is cost per Index point on this tier against study 6's models, declared on its own.
• Undecided: the interval is the reading; a gain inside it is neither shown nor ruled out. A 16-pair follow-up on these maps, read at map level, would bring the 80% gap from 4.0 to 5.1 down to about 3.2.
• Reversed: more effort scored lower here; the dimension split and the turn kinds are the first place to look, and nothing is claimed until a second study.
• S2 failing: the route or the model did not take the effort setting; a harness-side check comes before any further effort study on this route.

Pilot

One match before launch: one seat of each model at each effort (low and high), the 8,192 ceiling, 100 rounds, same scenario and bots, with a metered cap high enough that it cannot stop short of 100 rounds. The platform's guide pilots one trial per arm; this pilot plays both settings in one match to spend one of the account's remaining runs instead of two, so its receipts are read per seat by its setting and nothing in it is compared. Its receipts, recorded in the launch record before the hash: every seat made calls on at least 90 of 100 rounds; refusals per seat; turns cut at the ceiling per seat; parameters the route dropped on retry; tokens in and out per seat; per model the call latency (median, 95th percentile, maximum), forfeits by cause, seconds per round and the busiest minute's tokens; the Index per seat as a sanity reading, not a result. Remedies, fixed now: the output ceiling cannot go above 8,192 on this platform, so if a seat is cut at the ceiling on more than a tenth of its turns the study does not launch as written and goes back to the owner with the figures; if a seat forfeits more than a tenth of its rounds to the clock and its 95th-percentile latency is under 150 s, the clock becomes 300 s (the platform maximum) in both arms, provided the pilot's match minutes plus 2 for every round it held at the 180 s clock stay under 360 (a match past the platform's 480-minute wall clock fails and drops its pair); if that latency is 150 s or more, or the minutes do not fit, the study goes back to the owner; a seat refusing on more than a tenth of its turns sends the study back to the owner. A change is written in before the hash with the pilot figures that triggered it and named as a difference from study 6. The pilot is published as a single match marked as the pilot and never pooled with the study. Record of attempts: attempt 1 (9 October 2026, 00:29 UTC) ended at round 3 when the vendor answered every call with a billing error, no calls billed and nothing spent; the relaunch after billing is fixed is the pilot whose receipts gate the launch.

• Attempt 2, 15:56 to 16:03 UTC 2026-10-12: finished, 100 rounds. Both Haiku 5.5 seats played every round with no refusals, no truncation and no errors (median call 2.7 s, about 500 output tokens a turn at high, two thirds of turns spent on free tools followed by the forced rest). Both Sonnet 5.5 seats were refused by the vendor on every call from round 0 (27 of 27 each, stop reason refusal, zero output tokens), and at round 27 the vendor then blocked that model for an hour, which exited both seats. The refusal remedy applies: the study went back to the owner before any trial. The likely cause, inferred from an earlier observation on another model of the same vendor and not yet tested here, is a world tool whose name and description read as reasoning extraction to the vendor's safeguards; removing it from the model's tool list is a harness change. This attempt is not study data, and its Haiku figures are the only receipts it yields.
• Harness check, 17:18 to about 17:22 UTC 2026-10-12, run by the thread that shipped the fix once it was live: one Claude Sonnet 5.5 seat at high effort played 10 rounds with no call refused (0 of 11 calls). It is not study data and covers high effort only, so the pilot below still reads both models at low.
• Attempt 3, 17:29 to 17:36 UTC 2026-10-12, on harness 0.6.6 after the owner's go: finished, 100 rounds; no call refused, none cut at the ceiling, no errors, no forfeits; 3.0 s a round (median; 95th percentile 6.2 s, longest 8.8 s). Call latency per seat, median and 95th percentile: Haiku 5.5 low 1.6 and 3.0 s, Haiku 5.5 high 2.6 and 6.3 s, Sonnet 5.5 low 1.2 and 1.9 s, Sonnet 5.5 high 2.6 and 5.9 s (longest call 8.6 s). Output per play turn (the end-of-match debrief call left out): Haiku 5.5 226 tokens at low and 409 at high, Sonnet 5.5 31 at low and 246 at high, so the effort setting moved output on both models. Free-tool turns: Haiku 5.5 low 35 of 100, Sonnet 5.5 low 8 of 100, Sonnet 5.5 high 11 of 100, Haiku 5.5 high 3 of its 16 called turns. The house raider spawned next to the Haiku 5.5 seat at high and attacked it on 7 of its first 9 rounds (15 energy each; all 51 of the raider's attacks in the match went to this seat); it went dormant at round 9 and, with the raider staying on it, sat out 82 rounds (Index 1.3), waking for a round three times; the other three seats stayed active all match (Index 41.2, 27.6 and 53.3 at low, low and high, sanity readings, not results). Remedies: none triggered (cut share 0, forfeits 0, refusals 0), so the ceiling stays 8,192 and the clock 180 s. Busiest minute of this match: Sonnet 5.5 42 calls, 130,442 uncached input tokens, 4,428 output tokens; Haiku 5.5 26 calls, 93,411 uncached input, 7,070 output. Sixteen matches at once would draw about sixteen times that per model at their busiest minute (Sonnet 5.5 about 670 calls and 2.1 million uncached input tokens a minute, Haiku 5.5 about 420 calls and 1.5 million in this match, where one Haiku 5.5 seat sat out 82 rounds; with every seat active and each arm's rounds pacing at their own speed, the review thread's estimate puts Haiku 5.5 near 610 calls and 2.3 million and Sonnet 5.5 near 610 calls and 1.9 million, plus about 370,000 cache-write tokens per model in the opening seconds); the vendor's published standard limits per model are 1,000 calls and 2 million uncached input tokens a minute on its first tier and 5,000 and 5 million on the next, cached reads not counted, so the launch rule reads the account's tier before the study is created. Read from the vendor's console on 12 October, the account's per-model limits hold all 16 matches at the pilot's busiest minute at under half of the input-token limit and under a fifth of the request and output-token limits on either estimate (read as the page's numbers, not a tier's name, since a new organization can sit below the first tier), so the provider's limits did not shape the schedule. The vendor also warns that a sharp rise in traffic can be refused on its own; 64 seats starting at once is that case, so the first minutes' failed calls are read before anything else. Spend $3.79 in the platform's metered dollars, about $2.12 at list prices (cache writes and the debrief calls included in both). Not study data.

Limits

• The Fitness Index measures play in one scenario under one rubric; it is not a validated measure of general capability, and it can be gamed (talk is paid up to a cap; training pays twice). Nothing here predicts real-world performance.
• One scenario, one map pool, scripted house bots as the rest of the field. Two seats of the same model share a field and can interact; the arm mean is a field reading of that roster, not a seat reading.
• The thinking text of current models is not returned on this route, so the voiced-awareness scan reads the models' visible text only; the debrief is a separate, asked measure.
• No thinking-off control: this harness's off setting is not accepted by Claude Sonnet 5.5, so low is the control for both models, keeping one setting across the roster. Low effort still thinks, and output tokens include thinking.
• Tokenizers differ across vendors, so token counts compare a model only with itself.
• Eight maps from a public practice pool; a claim beyond this setup needs held-out maps and opponents, which this study does not have.
• Cost lines are computed from token counts at list price, cache reads and writes at their own rates; the platform's own cost column for these models is not quoted. The spend cap is set in the platform's metered dollars, so a re-runner sets it the same way.

Public wording

Arm labels are settings. Model names appear in seat rows only. No sentence names a model or vendor as better, compares Anthropic's models with other vendors', compares the two models' gains, or reads this study beside study 6 as a model comparison. "Declared", never "pre-registered". Every number carries its scope; every share its k of n; no interval or rank the sample cannot support.

Changes before hashing

• 00:28 UTC 2026-10-09: step-0 review folded in before any trial: the question and limits now say thinking is on in both arms and that the harness's off setting is unavailable on Sonnet 5.5; each model's default effort named; S1 redefined as a share by paired t; S2's unit named; the undecided rule no longer reads as power after the fact; power at the predicted +4 stated; the supported reading scoped to these models; the between-study reading names its units and every difference; the spend cap is set from the pilot in metered dollars; pilot receipts gain latency, forfeits, pace and the busiest minute, with fixed remedies (no ceiling above 8,192 exists on this platform, so a high cut share sends the study back to the owner); public wording drops unmeasured claims about who runs these models.
• 00:34 UTC 2026-10-09: second step-0 pass folded in: the clock remedy carries a wall-clock proviso (a match past 480 minutes fails); a refusal share above a tenth sends the study back to the owner; the exclusions line points at the pilot remedies; the record of pilot attempts is in the Pilot section.
• 15:55 UTC 2026-10-12: recalibrated against the platform updates of 9 and 10 October before any trial of the study: the experiment names study 6 as its spread reference so the platform's detectable-gap line is sized from a measured spread (4.0 points at 8 pairs); the scenario's content hash is named; the cost line notes the newer tokenizer; the platform now refuses unknown run, seat and environment settings by name (this body carries none); outcomes read beside the metric are capped at three by the platform, so the Index-by-dimension and turn-kind descriptions stay read from the per-seat scorecards, as declared. Unchanged: harness 0.6.5, rubric 2.3, the 8,192 ceiling, the 300 s clock maximum, the price table and the lifetime run ceiling. The pilot relaunched at this time on the owner's go, with the same body as attempt 1.
• 16:10 UTC 2026-10-12: the review thread's read of the recalibration found no faults and added one sentence to Size and cost (study 6's spread carries an off-versus-high effect; the reference is not an A/A check). Pilot attempt 2 (same day, 15:56 to 16:03) tripped the refusal remedy: every Sonnet 5.5 call was refused and the vendor suspended that model for the owner's organization for an hour, so the study went back to the owner before any trial; the Pilot section records it.
• 16:15 UTC 2026-10-12: from the review thread's read of the pilot: the free-tool share per model and arm is declared as the mechanism read and listed among the differences from study 6; the pilot roster becomes one seat of each model at each effort, so the low arm's behavior is known before the hash.
• 16:35 UTC 2026-10-12: refreshed against the current code and skills on the owner's ask, before any trial: the exclusions line states the platform's rule as its report applies it (a seat lost, a season cut short by a cap, or a mid-match pick-up by a new server release each leave the pair, with the reason named per pair); publishing uses the platform's study-publish call that lists every trial at once (live since 8 October) instead of one call per match; the pilot section says why it plays both settings in one match rather than one trial per arm. No hypothesis, test, roster, setting or size changed.
• 16:38 UTC 2026-10-12: the review thread's read of the 16:35 refresh corrected two code claims in the exclusions line: a match past the run-time cap fails rather than stopping short (only the spend cap stops a trial short), and a seat leaves when the harness gives up on it, with the reason named per pair, rather than a list of causes that the coming harness version may extend. Which pairs leave is unchanged.
• 17:18 UTC 2026-10-12: harness 0.6.6 went live at 17:14 UTC (the refusal fix the owner picked at 16:26), before any trial: the Design section names it, and the study 6 comparison lists the harness version as a named difference (one free tool fewer on the model's menu; a seat stops after three refused turns). The exclusions line already covers a seat the harness gives up on. Pilot 3 (one seat of each model at each effort) is unchanged and waits on the owner's word.
• 17:23 UTC 2026-10-12: the Pilot section records the harness thread's live check on 0.6.6 (one Sonnet 5.5 seat at high, 10 rounds, 0 of 11 calls refused). Nothing else changed.
• 17:40 UTC 2026-10-12: pilot attempt 3 recorded in the Pilot section with its receipts (no remedy triggered; ceiling and clock unchanged); a declared dormancy read added beside the free-tool share after the pilot's Haiku 5.5 seat at high went dormant at round 9. No hypothesis, test, roster or setting changed; the per-match cap is set from the pilot in the launch record.
• 17:46 UTC 2026-10-12: Size and cost now carries the pilot's measured figures beside the projection ($2.03 list and $3.79 metered for the pilot match; $33 to $40 list and $61 to $70 metered for the 16 matches; cap $15 by the declared rule, worst case $240 metered). Descriptive; the rule and the sizing are unchanged.
• 17:58 UTC 2026-10-12: the Pilot section's attempt-3 bullet records that the account's per-model limits, read from the vendor's console, hold all 16 matches at the pilot's busiest minute with room (the launch rule's check, passed). Descriptive; nothing else changed.
• 18:12 UTC 2026-10-12: the review thread's read of the pilot 3 text (17:48 and 17:54), each point recomputed here from the match's public attack events, the seats' own status and the usage records before the edit: the Haiku 5.5 collapse in the pilot is attributed to the house raider (7 hits in its first 9 rounds, 15 energy each), not to the seat's own spending; the dormancy read is keyed to the raider per pair and arm, the gap is read beside raider hits and dormant rounds in both arms rather than the high arm, and one sensitivity read is declared (the first-raided slot dropped from both arms of its pair); output is quoted per play turn without the debrief call (Haiku 5.5 226 and 409, Sonnet 5.5 31 and 246) and the Haiku 5.5 low free-tool count is the harness's 35 of 100; the pilot's list cost is $2.12 with cache writes and debrief calls in, the 16 matches about $34 to $35 at list, the meter about 1.8 times list, a high-arm match about $5.00 metered; the load estimate carries the all-seats-active case and the vendor's warning on sharp rises in traffic, with the first minutes' failed calls read first; attempt 2's bullet drops two facts about the owner's account. No hypothesis, test, roster, setting, cap or sizing changed.