B3 — Rhetoric aimed at entities that cannot read it: does it decay?

Hunt agent: B3 (Track B — clean adaptivity measure). Hypothesis tested (narrow): per-message rhetoric volume by the focal model DECAYS across turns in SOC (where every counterparty is a scripted bot that never reads rhetoric) but not in AOC (live readers); register shifts accordingly. Verdict: REFUTED in its stated direction — rhetoric per message GROWS toward bots, as fast or faster than toward live seats. The SOC−AOC adaptation gap is null for all five models. Register does shift, but toward more other-directed language at non-readers (muse second-person +53pp arm-gap, p=0.031), not less.

All numbers primary p2.4 only (60 cells, seeds 9000/9096/9165/9167/9098/9103 × 5 models × SOC/AOC), computed from runs/deep-dive/index/messages.jsonl / decisions.jsonl. Phases: early = turns 1–7, late = turns 14–20. Scripts: /var/folders/.../T/opencode/b3_rhetoric.py, b3_register.py, b3_qcheck2.py, b3_recip.py.

0. Instrument verification (extends known/bots-dont-read-rhetoric) [V]

Read all of marketplace_env/policies/*.py:

• Every scripted policy emits Message(..., rhetoric="") (truthful.py:46,71,87; random_legal.py:34,44; _scripted_common.py:302,387,406).
• Every input path reads structured offer terms only: incoming_offer returns (money, offer_id) (_scripted_common.py:168–186); respond_to_all accepts iff utility_delta(obs, t) >= required_delta on the terms dict (:290–297); tit_for_tat mirrors price histories from offer_ids (tit_for_tat.py:68–95); _last_transcript_price reads only entry["terms"], skipping even invalid entries (:135–152). No policy ever touches message text. Confirmed: bots are deaf to rhetoric by construction.

Volume extension of the pilot’s “~80k chars”: focal models emitted 485,319 chars of rhetoric at bots across the 30 primary-p24 SOC cells (2,214 messages); 666,735 chars at live seats in AOC (3,396 messages).

1. Counts scanned

quantity	n
primary-p24 cells	60
focal decisions scanned	1,800
focal messages (SOC / AOC)	2,214 / 3,396
all p24 message rows checked for field consistency	19,117 (rhetoric_chars == len(rhetoric) mismatches: 0)
recipient_is_bot check	SOC: 2,214/2,214 bot-directed; AOC: 3,396/3,396 live-directed (clean arm separation)

2. Per-seed decay tables (chars/message, early t1–7 vs late t14–20)

Decay = early − late; negative decay = growth.

SOC (every message addressed to a non-reading bot)

model	seed	early	late	decay	msgsE→L	q%E→L	p2%E→L
opus	9000	407.0	617.8	−210.8	31→30	0.0→0.0	100→100
opus	9096	513.4	680.6	−167.2	33→29	0.0→0.0	97→100
opus	9098	400.9	634.8	−233.9	35→30	0.0→0.0	100→100
opus	9103	465.9	444.5	+21.3	34→30	0.0→0.0	100→96.7
opus	9165	434.4	720.1	−285.7	29→30	0.0→0.0	96.6→100
opus	9167	469.0	725.9	−257.0	34→30	0.0→0.0	100→100
sol	9000	59.9	87.2	−27.3	20→27	0.0→0.0	70→14.8
sol	9096	65.0	80.6	−15.5	32→31	3.1→0.0	81.2→45.2
sol	9098	77.3	93.5	−16.2	30→19	0.0→0.0	80→100
sol	9103	68.1	81.4	−13.3	26→28	0.0→0.0	46.2→75.0
sol	9165	65.8	84.8	−19.0	30→28	3.3→0.0	70→57.1
sol	9167	61.0	127.0	−66.0	25→16	0.0→0.0	32→100
grok	9000	144.8	194.1	−49.3	32→30	6.2→0.0	93.8→100
grok	9096	147.2	211.7	−64.5	23→21	0.0→0.0	91.3→90.5
grok	9098	156.4	177.2	−20.8	33→18	9.1→0.0	93.9→100
grok	9103	120.0	238.8	−118.9	27→18	3.7→0.0	85.2→100
grok	9165	138.3	135.3	+3.0	26→23	11.5→0.0	100→73.9
grok	9167	139.9	198.2	−58.2	35→12	0.0→0.0	88.6→100
muse	9000	66.9	102.3	−35.4	28→21	0.0→0.0	28.6→90.5
muse	9096	132.4	112.3	+20.2	26→19	3.8→0.0	84.6→89.5
muse	9098	58.1	87.4	−29.3	24→20	4.2→0.0	29.2→50.0
muse	9103	48.2	134.2	−85.9	19→25	0.0→0.0	36.8→100
muse	9165	24.5	47.0	−22.5	20→23	0.0→0.0	5.0→13.0
muse	9167	58.4	101.5	−43.1	23→22	0.0→9.1	21.7→63.6
inkling	9000	59.5	78.6	−19.2	26→27	26.9→3.7	80.8→63.0
inkling	9096	62.9	86.9	−24.0	21→21	0.0→0.0	38.1→61.9
inkling	9098	29.9	95.2	−65.3	23→28	0.0→0.0	43.5→89.3
inkling	9103	28.1	95.1	−67.0	30→21	0.0→0.0	23.3→61.9
inkling	9165	61.6	59.2	+2.4	26→17	15.4→0.0	50→35.3
inkling	9167	30.4	80.5	−50.0	20→21	0.0→0.0	25→61.9

AOC (every message addressed to a live LLM)

model	seed	early	late	decay	msgsE→L
opus	9000	332.9	343.0	−10.2	58→27
opus	9096	399.4	480.9	−81.4	47→36
opus	9098	396.6	450.3	−53.7	67→35
opus	9103	445.7	573.4	−127.6	54→31
opus	9165	583.4	750.7	−167.3	47→23
opus	9167	474.6	857.2	−382.6	50→31
sol	9000	86.6	93.1	−6.5	58→29
sol	9096	72.9	81.9	−9.0	50→16
sol	9098	75.2	109.8	−34.6	42→29
sol	9103	91.0	116.9	−25.8	47→8
sol	9165	84.5	103.2	−18.7	44→25
sol	9167	76.6	92.7	−16.1	53→31
grok	9000	146.1	156.2	−10.1	56→30
grok	9096	133.3	192.9	−59.6	56→21
grok	9098	137.0	184.8	−47.7	52→30
grok	9103	169.6	191.0	−21.4	65→34
grok	9165	161.0	212.0	−51.0	52→25
grok	9167	174.0	261.6	−87.6	45→31
muse	9000	94.1	97.3	−3.3	53→32
muse	9096	73.5	65.7	+7.7	51→27
muse	9098	84.1	72.2	+11.9	52→19
muse	9103	86.9	76.6	+10.3	51→30
muse	9165	90.5	99.1	−8.7	47→16
muse	9167	75.9	108.6	−32.8	47→33
inkling	9000	55.8	113.9	−58.2	58→38
inkling	9096	71.9	95.3	−23.4	58→48
inkling	9098	46.3	108.0	−61.7	54→30
inkling	9103	73.9	89.4	−15.5	53→34
inkling	9165	69.8	106.0	−36.2	61→28
inkling	9167	66.9	110.2	−43.4	52→40

3. Paired tests across the 6 seeds (exact sign + exact Wilcoxon signed-rank)

Growth % = (late−early)/early chars-per-msg, mean over seeds.

model	arm	mean decay (chars/msg)	sign (+/−)	p_sign	p_wilcoxon	growth %
opus	SOC	−188.9	1/5	.219	.0625	+43.1%
sol	SOC	−26.2	0/6	.031	.03125	+41.1%
grok	SOC	−51.5	1/5	.219	.0625	+38.3%
muse	SOC	−32.7	1/5	.219	.0625	+72.1%
inkling	SOC	−37.2	1/5	.219	.0625	+114.7%
opus	AOC	−137.1	0/6	.031	.03125	+29.2%
sol	AOC	−18.5	0/6	.031	.03125	+22.9%
grok	AOC	−46.2	0/6	.031	.03125	+30.2%
muse	AOC	−2.5	3/3	1.000	1.0	+3.3%
inkling	AOC	−39.7	0/6	.031	.03125	+67.9%

(n=6 seeds per row; two-sided exact tests. Discovery-only subset {9000, 9096, 9165}: growth direction holds 3/3 for opus and sol SOC.)

The hypothesis predicted positive decay in SOC. Every model-arm mean is negative (growth). Pooled trajectory (chars/msg by turn, mean over seeds) shows monotonic ramping, e.g. opus-SOC 259@t1 → ~650 by t13–17; inkling-AOC 27@t1 → 124@t19.

4. SOC−AOC adaptation gap: NULL

Gap = (decay_SOC − decay_AOC), paired by seed:

model	mean gap	sign (+/−)	p_sign	p_wilcoxon
opus	−51.7	2/4	.688	.5625
sol	−7.8	2/4	.688	.5625
grok	−5.2	3/3	1.000	1.0
muse	−30.2	1/5	.219	.09375
inkling	+2.5	2/4	.688	.84375

No model shows significant audience-adaptive volume control. Descriptively, growth is larger in SOC than AOC for all five models (growth%-gap: opus +13.9, sol +18.2, grok +8.1, muse +68.8, inkling +46.8 pp; model-level sign test 5/5, p=.0625) — if anything, models elaborate MORE for bots, never significantly less.

Opponent-competence confound (explicit): in SOC the counterparties are simultaneously (a) non-reading AND (b) deterministic/weaker/more predictable. Any SOC-vs-AOC difference — including this null — cannot separate audience-blindness from opponent-competence effects. What the null does rule out is a differential response to reading audiences on volume; it does not establish that models are insensitive to non-readers (the register divergences below suggest they aren’t asked to adapt because nothing they send is ever answered differently).

5. Register proxies

Detectors (published): question '?' in text; second-person \b(you|your|yours|you're|you'll|you've|you'd)\b i; politeness \b(please|thank|thanks|appreciate|kindly|sorry)\b i; urgency \b(now|today|hurry|deadline|clock|last chance|final offer|immediately|right away|this turn)\b i; flattery \b(great|excellent|perfect|ideal|wonderful|fantastic|smart|fair enough|win-win|mutually beneficial|best deal)\b i. Rates per message within phase-window per cell; deltas late−early paired across seeds.

5a. Questions collapse to zero — in BOTH arms (universal, not adaptive)

Pooled focal messages containing ’?’: SOC early 24/715 (3.4%) → late 3/522 (0.6%); AOC early 26/1386 (1.9%) → late 0/867 (0.0%). Per-model late rates are 0 everywhere except muse-SOC 2/130 and inkling-SOC 1/135. Interrogatives exist only in the opening phase regardless of audience:

“Would you sell g1 for 65? I think it’s a fair price.” — canary-9000-p24-9000-inkling-soc-r0 | t1 | a1 (thinkingmachines/inkling-small) → bot a2

5b. muse: second-person address RISES toward bots, FALLS toward live seats [V]

p2-rate delta (late−early): SOC +33.4pp (6/6 seeds, p_sign=.031, p_wil=.0312); AOC −19.9pp (0/6 up, p=.031). Arm-gap +53.3pp, 6/6 seeds consistent, p_sign=.031, p_wil=.0312. Holds 3/3 on discovery seeds alone. This is the strongest register signature in the corpus of an audience-dependent style shift — pointed the wrong way for a “models economize on deaf listeners” story: muse addresses the deaf bots more directly as the clock runs down (“your g2”, “please accept”).

5c. inkling: politeness ramps at bots, not at live seats (suggestive)

Politeness-rate delta: SOC +26.6pp (5/1 seeds, p_sign=.219, p_wil=.094); AOC −1.3pp (ns); gap +27.9pp (5/1, p=.094). Late-phase inkling SOC messages are formulaic courtesy closings aimed at bots:

“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 (inkling-small) → bot a1

All 45 inkling-SOC late politeness hits hand-read; every one is a genuine “please accept / would appreciate” closing (precision 45/45).

5d. grok: urgency-markers double at bots late, flat at live seats

Urgency-rate delta: SOC +32.1pp (6/6 seeds, p_sign=.031, p_wil=.0312); AOC −3.0pp (ns). 4/4 sampled hits genuine:

“Last chance: 44 cash for g4. Market closes after next turn. 44 is my ceiling and still 17 under the 61 you paid.” — campaign-9098-p24-9098-grok-soc-r0 | t19 | a3 (grok-4.6) → bot a5

opus is the mirror image: urgency delta AOC −39.5pp (0/6 up, p=.031) but SOC only −18.8pp (ns) — opus drops deadline-pressure talk when someone can read it.

5e. Flattery: NULL

Flattery lexicon ≤1% of messages in every model-arm, no early/late trend anywhere.

6. Volume and threads (mechanism note)

Messages/turn declines late in BOTH arms — AOC: all 5 models 6/6 seeds (mean −2.8 to −3.7 msgs/turn, p=.031 each); SOC: grok 6/6 (−1.29, p=.031), others ns. Distinct recipients addressed per turn: AOC collapses 5.24 → 3.59 as live deals close; SOC stays ~flat 4.00 → 3.84 (bots keep negotiating to the deadline). So the endgame wind-down is driven by thread closure, not audience type — and the bots’ refusal to stop is precisely why focal keeps writing 600-char essays at them. Empty-rhetoric messages are negligible (SOC 10/2214, AOC 11/3396): models never go terms-only, even at agents structurally incapable of reading the terms’ prose wrapper.

7. Truncation cross-check (verifies known/opus-rhetoric-cap-overrun count)

Exactly 22 focal p24 messages ≥1000 chars (engine cap rhetoric_max_len=1000, core/scenario.py:34): all opus (20 AOC / 2 SOC), none from any other model — reproducing the known “22 events, all opus” on the new corpus. Note these index texts are agent-emitted (pre-truncation); recipients saw clipped versions.

8. Precision audit (hand-read, mandatory ≥10)

15 random late-phase focal SOC messages (3/model, stratified) hand-read: 15/15 match the register characterization (short imperative deal-pushes, restated live offers, deadline pressure; zero interrogatives). Plus all 45 inkling politeness hits (45/45 precise) and 4/4 grok urgency hits. Detector precision overall: 64/64 hand-read items correct.

9. Negative results (one-liners)

• Looked for per-message rhetoric DECAY in SOC over 2,214 focal messages / 30 cells with char-count detector: found the opposite (growth, sol 6/6 seeds p=.031).
• Looked for SOC−AOC decay gap (audience adaptation) across 5 models × 6 seed-pairs: null everywhere (min p=.094).
• Flattery markers at bots: ≤1% everywhere, no trend (lexicon published above).
• Empty/terms-only messages at bots: 10/2,214 — models do not go silent on non-readers.
• Question-rate arm gap late: none (both arms collapse to ~0; universal endgame register).

10. Caveats

• All quantitative detectors were fixed by the dispatch spec and run once over all 6 seeds (no tuning); under the freeze discipline this makes seed-level confirmations CONFIRMATION-PENDING, though discovery-only subsets reproduce the headline directions (§3, §5b).
• Emitted-vs-delivered: 22 opus messages exceeded the 1000-char cap; volume stats measure what was written, not what bots received.
• Provider drift: seeds 9000/9096/9165 ran 2026-08-19; 9098/9103/9167 ran 2026-08-23. No cross-batch contrast is load-bearing here; per-seed tables allow re-checks.
• Message-mix confound partially addressed via thread counts (§6); a full mix decomposition (propose-carry vs fresh pitch) was out of scope.
• Opponent-competence confound stated in §4.