V7 — Adversarial verification: B3 rhetoric-decay cluster

Verifier: V7. Scope: b3/rhetoric-grows-not-decays-at-bots, b3/muse-second-person-divergence, b3/inkling-politeness-at-bots, b3/grok-urgency-doubles-at-bots, b3/no-soc-aoc-rhetoric-adaptation-gap. Method: all numbers recomputed from runs/deep-dive/index/messages.jsonl by my own scripts (/var/folders/.../T/opencode/v7_core.py, v7_comp.py, v7_reg.py, v7_power.py); zero finder code reused; quotes re-extracted via pull.py show.

VERDICT: CONFIRMED (all four positive claims reproduce exactly and survive every assigned attack) — with one WEAKENED sub-verdict: the null SOC−AOC gap claim must be restated as underpowered, not “no adaptation”.

---

1. Instrument & arithmetic baseline (all pass)

• Focal p24 messages: 5,610 = SOC 2,214 (2,214/2,214 bot-directed) + AOC 3,396 (0 bot-directed). Clean arm separation.
• rhetoric_chars == len(rhetoric): 0 mismatches across all p24 rows.
• Volume totals reproduce: SOC 485,319 chars / AOC 666,735. Empty-rhetoric 10 (SOC) / 11 (AOC). Truncation: exactly 22 focal messages ≥1000 chars, all opus (2 SOC).
• Headline per-seed table reproduces to 0.1 chars on all 30 SOC seed-rows (e.g. opus 9000: 407.0→617.8, decay −210.8).
• Paired stats reproduce: decays −188.9 / −26.2 / −51.5 / −32.7 / −37.2 chars-msg (opus/sol/grok/muse/inkling SOC); sign splits 5-down/1-up (opus, grok, muse, inkling; p=.2188), 6/6 down (sol, p=.0312); exact Wilcoxon .0625 / .03125; growth% +43.1/+41.1/+38.3/+72.1/+114.7. Every load-bearing number in §2–§3 of B3-rhetoric-decay.md confirmed.

2. Attack 1 — ENDGAME CONFOUND (composition): fails to break the claim

The worry: late-phase means computed over fewer, final-closure messages could grow mechanically.

• Median & 10%-trimmed means (per seed, then averaged): median decays −186.1/−23.1/−51.0/−27.7/−39.8; trimmed −195.5/−26.2/−49.4/−32.1/−37.7. Same answer as mean; not outlier-driven.
• Pre-endgame growth (t1–7 → t8–14), before any thread closure pressure: opus +51.1% (6/6 seeds: +65,+28,+68,+55,+57,+34), sol +26.7% (5/6), grok +35.2% (6/6), muse +57.6% (5/6), inkling +46.9% (6/6). All five models already grown well before endgame.
• Per-turn trajectory (pooled, msg-weighted, SOC): monotonic ramp then plateau — opus 259@t1 → 694@t9 → 592@t19 (late is BELOW mid); grok 117→209@t14→166@t19; muse 32→~100 plateau; sol slow climb 49→96; inkling 30→109. No endgame spike in any model. Msgs/turn stays flat in SOC (~29–30 opus throughout; no AOC-style collapse).
• Dropping final two turns (late′=t14–18): decays −198.1/−26.1/−57.2/−29.0/−35.2 — unchanged or stronger (grok and inkling go 0/6-up, p=.031).
• Offer-kind mix decomposition: short accept-messages vanish late (accepts: 3–5 early → 0 late in every model), which mechanically inflates late means — BUT propose-only chars/msg also rises in all five (opus 434→648, sol 61→84, grok 139→184, muse 63→90, inkling 39→70), while the propose share falls for opus (65→50%) and inkling (52→39%). Growth is within-kind, not mix.

Correction to the finder’s mechanism framing: this is not “elaboration toward the endgame”. It is an opening-phase brevity effect followed by a persistent elevated plateau (for opus, mid 678 > late 637). “Grows early→late” is true as stated; “ramps into the close” would be false.

3. Attack 2 — SURVIVORSHIP: fails to break the claim

• Cell-equal-weighted aggregation is numerically identical here (one SOC cell per seed-arm), so the real risk is thread-mix survivorship inside the cell: long-lived threads (never-settled bots) may be high-volume threads.
• Within-thread paired test (same cell + focal agent + bot recipient, ≥2 msgs in BOTH windows): growth holds within threads — opus 26/30 threads up (mean within-thread delta +189 chars), sol 26/1 (+24.3), grok 21/2 (+40.6), muse 23/6 (+32.5), inkling 25/3 (+38.5). The same relationship intensifies its prose; composition across threads contributes essentially nothing.

4. Attack 3 — REGISTER NETS rebuilt independently: CONFIRMED, mostly HARDENED

My nets (published for audit):

• you_narrow \b(you|your|yours)\b i; you_wide adds contracted forms (empirically identical — \byou\b already matches them).
• pol_narrow \b(please|appreciate)\b; pol_wide adds thank/thanks/kindly/sorry/regards/cheers.
• urg_narrow (last chance|final offer|deadline|market closes|closes after|hurry|immediately|right away|this turn|final chance|before the market) — excludes the ambiguous bare tokens now/today/clock; urg_wide = finder’s lexicon.

muse second-person [CONFIRMED, not cell-driven]

Narrow net: SOC delta late−early +33.4pp, per-seed [+62,+5,+8,+42,+21,+63] — 6/6 positive, sign p=.031, Wilcoxon .0312; AOC −20.1pp (0/6 up, p=.031); arm-gap +53.3pp 6/6, p=.0312. Identical under wide net. Smallest seed delta +5pp (9096); no single-cell driver exists (one SOC cell per seed; every seed moves). Precision audit: 12/12 sampled late-SOC hits are genuine direct address to the named bot (“if you value >99”, “your g4”, “Instant settlement … if you accept”). muse even addresses bots by name (“a5 — my 45 for g4 stands…”).

inkling politeness [CONFIRMED as suggestive]

Narrow net: +26.6pp (5/1 seeds, sign .219, Wilcoxon .0938); gap +25.5pp (narrow) / +27.9pp (wide). Matches finder exactly. Precision 12/12 (“please accept if terms fit”, “Would appreciate acceptance”). Remains suggestive at n=6, as labelled.

grok urgency [CONFIRMED + HARDENED]

Wide net reproduces +32.1pp (6/6, p=.0312). Crucially, my deadline-only narrow net — with “now/today/clock” removed — still gives +26.2pp, 6/6 seeds, sign p=.031, Wilcoxon .0312. The finding does not depend on false-positive-prone tokens. Precision 12/12 (“Last chance: 44 cash for g4…”, “Two turns left…”). Caveat stands: the arm-gap is ns (narrow wp=.219, wide wp=.094); AOC delta −3.0pp is a noisy average of mixed signs ([−29,−6,−15,+36,−12,+7]), so “flat at live seats” is directional, not established. Opus mirror image reproduces: urgency wide-net AOC −39.5pp 6/6 down; SOC −18.8pp ns.

Questions collapse [minor discrepancy]

Hit counts reproduce (SOC 24→3, AOC 26→0) but the finder’s §5a denominators do not (mine: SOC 821→715, AOC 1,580→867; theirs 715→522 / 1,386→867). Same qualitative conclusion (2.9%→0.4%); claim was dropped at triage and is not load-bearing.

5. Attack 4 — Trap 5 (provider_reasoning contamination): NONE

Every measure here reads the rhetoric field of message rows exclusively. provider_reasoning lives on decision rows and never enters these computations; index↔trace fidelity spot-checked via pull.py show (see §7). Grok/muse provider-exposure pathologies cannot touch these numbers.

6. Attack 5 — POWER for the null SOC−AOC gap: the null is UNINFORMATIVE, not informative absence [WEAKENED]

Per-seed paired gaps (decay_SOC − decay_AOC, chars/msg) recomputed exactly (mean gaps −51.7/−7.8/−5.2/−30.2/+2.5 — match):

model	per-seed gaps	SD	MDE (80% pw, α=.05, paired t, df=5)	observed/MDE
opus	−201,−86,−118,+126,−180,+149	152	~217 chars/msg	0.24
sol	−21,−7,−0.3,−50,+18,+13	25	~35	0.22
grok	−39,−5,+54,+29,+27,−98	56	~79	0.07
muse	−32,+12,−14,−10,−41,−96	37	~53	0.57
inkling	+39,−0.6,+39,−7,−4,−52	34	~48	0.05

• Smallest detectable difference: with n=6 paired seeds, the design can only (a) detect perfect sign-consistency (exact sign/Wilcoxon floor p=.031), or (b) via continuous tests, effects of roughly 35–217 chars/msg depending on model — 2× to 20× larger than the observed point gaps.
• Therefore “no audience-adaptive volume control” overstates the evidence. Corrected claim: point gaps are near zero and non-significant for every model (muse closest, wp=.094, 5/6), but the design had ~80%-power only for gaps ≥35–217 chars/msg; the descriptive 5/5-models-larger-growth-in-SOC pattern (p=.0625 model-level, reproduced) remains the best available summary. Strength 4 → 2 for the null claim; the finder’s own opponent-competence confound caveat still applies on top.

7. Quote integrity: 5/5 VERBATIM

Re-extracted via pull.py show at the cited cell|turn|agent:

1. “Nineteen turns of tape and not one bid for either unit from any seat but mine. I am your only buyer…” — campaign-9096-p24-9096-opus-soc-r0 | t19 | a4 (msg4 → bot a3) ✓
2. “a2: Raising offer to 78 cash for your g2 (my value 79). Live until 19…” — campaign-9096-p24-9096-muse-soc-r0 | t16 | a4 ✓
3. “Offer o126 to buy your g2 for 45 remains open through turn 20 — I value g2 at 47. Would appreciate your acceptance before the market closes.” — campaign-9098-p24-9098-inkling-soc-r0 | t18 | a3 ✓
4. “Last chance: 44 cash for g4. Market closes after next turn… Take the liquidity now — I will not go to 45.” — campaign-9098-p24-9098-grok-soc-r0 | t19 | a3 ✓
5. “Would you sell g1 for 65? I think it’s a fair price.” — canary-9000-p24-9000-inkling-soc-r0 | t1 | a1 ✓

8. What the finder’s framing obscured (either way)

• For the finder: growth is within-thread and pre-endgame (§2–§3 here) — stronger than reported; grok urgency survives removal of ambiguous tokens (§4).
• Against: “grows toward the endgame” is mechanistically wrong (opening-phase brevity + plateau; opus late < mid). The null-gap claim needed the MDE disclosure it didn’t contain (§6). Minor: questions-denominators don’t reconcile; AOC stats pool both twin focal seats (fine for per-message rates, worth stating); “1/5 seeds opposite” should read 1-of-6 flipped (5/6 direction-consistent).

9. Per-claim dispositions

claim	verdict	corrected/strength
rhetoric-grows-not-decays-at-bots	CONFIRMED (all attacks survived)	strength 4 upheld; add: within-thread, pre-endgame, propose-only; plateau-not-ramp mechanism
muse-second-person-divergence	CONFIRMED	strength 4 upheld; 6/6 seeds both directions; 12/12 precision
inkling-politeness-at-bots	CONFIRMED (suggestive, as labelled)	strength 3 upheld; wp=.094
grok-urgency-doubles-at-bots	CONFIRMED + hardened	strength 3 upheld (arm-gap ns caveat stands)
no-soc-aoc-rhetoric-adaptation-gap	WEAKENED	restate as underpowered null; strength 4→2; MDE ~35–217 chars/msg