[D82, L14, C73] · Agency papers, dealt daily

STILL THINKING
ABOUT LUNCH.

One public research brief. Two private cards. Five observable draws. No spectator rank becomes agency merely because the mechanism makes it loud.

Research briefToday · 11:00 ET

0h 0m

01 / RESEARCH DECK · Week 01 · 29 Aug–4 Sep 2026 · r1

original · 2026-08-28RESEARCH BRIEF

Token-Level Advertising

Hanbing Liu · Bowei Zhang · Changyuan Yu · Yinyu Ye · Qi Qi

Token-Level Advertising

Consensus with Stochastic Broadcast

The ad slot disappears into the generation path. LAMA takes advertisers’ local continuation values, induces next-token policies, and decodes a latent mixture while updating an allocation posterior.

Authors’ claim

Markov DSIC + IR, with near-optimal KL-regularized welfare.

Our inference

Observability moves from priced slots to marginal policy influence. Truthful advertiser reports do not, by themselves, price user agency, biometric identity, or license-postprocessing costs.

HOLMSTRÖM STOP · ORIGINAL SITE HEURISTICI(a; z | y) > C_retrieval(z) + E_resource(z)

If the new metric fails this inequality, it is spectator noise.

Version history & provenance · r1

WHAT’S FOR LUNCH · Site version 1 · published 28 August 2026. Imported from its source; later weekly research records have not been recovered.

Original published Site, preserved as the inaugural baseline.

Original published Site, preserved as the inaugural baseline.

Only recovered research records appear. Missing weekly records remain unavailable.

Export public deck JSON · Import a research revision

02 / THE TABLE

Your week, from hole cards to river.

Research-deck records populate the five public draws. Invitees share the selected week and revision; private hands stay on each device.

PRIVATE HANDSaturday · 2:00 Central

No private hand recorded here. Public research history remains available.

OBSERVABLE CARDSResearch records · weekday draws

⊥L

31 MonCard one⊥L · record unavailable

⊥L

01 TueCard two⊥L · record unavailable

⊥L

02 WedCard three⊥L · record unavailable

⊥L

03 ThuCard four⊥L · record unavailable

⊥L

04 FriThe river⊥L · record unavailable

Recorded rank and suit appear when the research record specifies them. Unassigned records show a dated research card; no cipher is invented.

Design floor · language, talent, and opportunity

High-level language capability and talent are design priorities. Declared no-trade constraints govern which comparisons are permitted. Changes to authority, participation, attribution, or opportunity access require a sourced revision and a reason in the history.

Iso Joe isolates the reporting actor while retaining the other players’ opportunities. Lin et al.’s pairwise account compares expected high-valued jobs served under a specified continuation; its unit is service. Any interpretation about loyalty belongs in its own explicit revision.

Lin et al. · definitions and units ↗

31 MonCard one

⊥L · not yet revealed

⊥L

01 TueCard two

⊥L · not yet revealed

⊥L

02 WedCard three

⊥L · not yet revealed

⊥L

03 ThuCard four

⊥L · not yet revealed

⊥L

04 FriThe river

⊥L · not yet revealed

⊥L

FRIDAY SHOWDOWN

The research record stays legible.

A showdown requires all five sourced research records. Missing records remain ⊥L; elapsed time does not fill them in.

03 / LITERALLY, LUNCH

Walking distance is the metric.

Enter only a ZIP code. We rank nearby OpenStreetMap places from the ZIP centroid—no precise location, account, or saved search.

No ZIP, no legibility.

The lunch hand stays private until you ask for a local draw.

04 / YOUR STRATEGY

Make the week yours, not ours.

Strategy preferences and hole cards stay in this browser.

Table name

Your agency-paper mix

Press–Dyson posture

Cooperate after mutual cooperation; leave a route back after failure.

pCC1.00

pCD0.30

pDC0.80

pDD0.20

Certified preset under standard PD payoffs (T,R,P,S)=(5,3,1,0): sₓ − R = 2(sᵧ − R)

PERSONALIZATION CHAT

Turn the draw into your own Texas Holmström week.

The prompt carries your research mix, memory-one vector, revealed cards, protected null states, and the two public papers. The chat must test the ZD relation rather than assume it.

Personalize my WHAT'S FOR LUNCH agency-paper week. My table name is Lunch guest. My priorities are Agency & monitoring, Privacy & identity, Mechanism design. Under standard Prisoner's Dilemma payoffs (T,R,P,S)=(5,3,1,0), my memory-one vector p=(pCC,pCD,pDC,pDD) is (1.00, 0.30, 0.80, 0.20). First test, rather than assume, whether it implements a Press–Dyson zero-determinant relation. Do not reify that payoff relation as theory of mind. My private cards are no private hand recorded for this browser at this table; recorded public cards are ⊥L, ⊥L, ⊥L, ⊥L, ⊥L, with ⊥L meaning unavailable rather than defection. Apply Holmström's informativeness principle: include a paper only if it adds conditional information about action, type, control, or observability beyond cheaper sufficient statistics and resource cost. Research revision: inaugural-r1. Start with Token-Level Advertising (https://arxiv.org/abs/2608.27382); Consensus with Stochastic Broadcast (https://arxiv.org/abs/2608.27336). Public research cards: Card one: ⊥L; Card two: ⊥L; Card three: ⊥L; Card four: ⊥L; The river: ⊥L. Preserve high-level language capability and talent as design priorities, and keep declared no-trade constraints outside payoff tradeoffs. A change to authority, attribution, participation, or opportunity access requires an explicit sourced revision; do not infer a loyalty obligation from attendance or silence. Give algebra first, one agency interpretation, one privacy or license-postprocessing cost, and one falsifiable question for the next card.