scenarios / scaling-question / methodology
deslopmedia
Methodology · v0.1

The scaling question, in six reductions.

HoR · Head of Research · EIC · Editor-in-Chief · CTO · Chief Technology Officer
Filed 2026-05-25 · Synced quarterly · Locked plan: HoR + EIC + CTO three-way co-owned

The scenarios scorecard tracks where a curated AI-100 cohort positions on a six-question cascade about whether the scaling hypothesis pays out — and aggregates community signal across six named outcomes. This page documents the firewalls: what the framework does and does not claim, what the encoding shows and what it deliberately omits. For per-voice references, see the source ledger at the bottom of each scorecard.

§ 1 — Framework scope

The scaling-question framework decomposes the bet into a six-question cascade: three substrate strands gating three downstream binaries.

  • Q1a — Science. Does the AI performance curve continue smooth, predictable improvements over the next ~5 years?
  • Q1b — Inputs. Do compute / power / data / capital / talent remain available at the rate the science would absorb?
  • Q1c — Agents. Do models translate into agents capable of doing autonomous knowledge work?
  • Q2 — Lab durability. Do frontier labs maintain meaningful technical lead-time over fast-followers?
  • Q3 — Position capture. Do frontier labs convert their lead into a durable compounding position?
  • Q4 — Societal permission. Does society permit the labs the concentration of power that follows from Q2 + Q3?

Each gate is binary in the data model — HOLDS continues the trunk, FAILS peels to a named outcome — but the underlying voice positions are six-valued (HOLDS, QUAL · HOLD, NOT ENGAGED, QUAL · DISPUTE, DISPUTES, REJECTS FRAME). The cascade is not a forecast and not a probability assignment. It is a navigable map of where on-record AI-100 voices have routed the bet.

§ 2 — Voice selection

The AI-100 cohort is curated annually with multi-rater review. Inclusion requires on-record positioning in venues we can cite — earnings calls, podcasts, papers, public testimony, official statements. We do not attribute positions from private communication, hearsay, or paraphrased third-party reporting unless the third party is the primary source for that record.

At v0.1, eleven named voices engage the scaling-question framework substantively enough to appear in the matrix. Each cell call carries a primary citation in the source ledger at the bottom of the scorecard; the per-voice substrate tier (A / B / C / D) lives in the audit sibling YAML (<slug>.audit.yaml).

§ 3a — Stance vocabulary

Six stance categories are recognized:

  • HOLDS — voice asserts the position’s positive case.
  • QUAL · HOLD — voice asserts the positive case with stated reservations.
  • NOT ENGAGED — voice’s published corpus does not address the question.
  • QUAL · DISPUTE — voice asserts the negative case with stated reservations.
  • DISPUTES — voice asserts the position’s negative case.
  • REJECTS FRAME — voice argues the question itself is mis-formed.

The QUAL · HOLD / QUAL · DISPUTE distinction tracks substrate-backed reservations within an otherwise affirmative or negative stance. REJECTS FRAME is rare and is treated separately from disputation within the frame.

Each cell call’s primary citation is per-voice substrate-of-record (Wave 9 placement card or upstream verified dossier). The source ledger at the bottom of every scorecard renders the ⟨voice, stance, quote, source⟩ tuple per active-stance cell; cells with thin substrate render dimmed in the matrix and carry no ledger entry rather than ship a weak citation.

§ 3b — Why downstream questions look thinner

The substrate of Q1 (science, inputs, agents) is empirical: benchmark traces, capex disclosures, agent evaluations. The substrate of Q2 (lab durability), Q3 (position capture), and Q4 (societal permission) is dispositional rather than empirical: governance, market judgment, regulatory weather. These substrates mature on different cadences than Q1.

The hero figure surfaces this honestly — at lower substrate densities, the downstream gates show thinner trunks and more grey-dotted connectors than the upstream gates. This is graceful encoding degradation, not a defect of the figure.

The fourth Where-We-Are panel on the scorecard surfaces this scarcity explicitly in prose: “Below-the-wire: dispositional rather than empirical at v0.1. The framework’s downstream questions are governance and market judgments; their substrate matures on a different cadence.”

§ 4 — Encoding

Three lanes carry stance in the hero: volt (HOLDS), coral (DISPUTES), grey-dotted (NOT ENGAGED). Width encodes voice count: in the volt trunk and the coral peels, one voice = seven pixels of width, monotonic and additive. The grey-dotted connectors use uniform 2px stroke (count is shown by the Other band’s height, not by per-connector width — connectors carry only routing).

Outcome band heights encode aggregate community signal weight (a function of voice mass distributed across consistency-compatible terminals; see scoring engine). Heights are bucketed into five buckets (XS=54px, S=72px, M=88px, L=100px, XL=128px); within a bucket, bands are equal-height by design to discourage misreading inter-band differences as ordinal weight claims.

Each band carries a 4px stripe at its left edge: volt for the HOLDS terminal (O1), coral for DISPUTES terminals (O2–O5), grey for off-cascade Other (O6). The Other band additionally has a softer fill and stroke (rgba(138,143,152,0.10) / #5a6680) to make its off-cascade status visible at a glance without the legend.

§ 5 — Why no probability percentages

The figure shows the structure of the bet, not probabilities. We deliberately do not display percentages, expected values, or probability bands. Three reasons:

  1. The voice counts are not a probability sample. The AI-100 cohort is curated for epistemic standing in the discourse, not for representativeness of any underlying population. Probability claims from this substrate would be category errors.
  2. The questions are not independent in any frequentist sense. Q2 conditions on Q1; Q3 on Q2; Q4 on Q3. Multiplying gate fractions to produce path probabilities is undefined when voices position on a gate without committing to its conditioning predecessors.
  3. The framework explicitly admits the off-cascade. Outcome 6 (Other) gathers voices that decline the question rather than answer within it — these voices have no defined probability within the cascade and we do not impute one.

The publication-view editorial disclaimer carries the analogous register move: deslop.media takes positions on relative likelihood (optimistic on science, guarded on lab durability and position capture, cautious on societal permission), but does not assign probabilities. The tree shows the structure; the disclaimer carries the lean.

§ 6 — Refresh cadence

Quarterly. Each refresh updates the matrix, the trunk widths, the band heights, the substrate panels, the audit YAML’s per-voice tier, and the computed JSON’s substrate states. Refresh windows are stamped at the bottom of every scorecard (last synced + next refresh); a refresh that misses its window publishes a stale-banner before any reader sees out-of-date substrate.

Framework version moves only on structural changes (a new question, a new outcome, a re-keying of the path signatures). Methodology version moves on encoding or scoring changes. Both are versioned independently and both are stamped on every reading surface.

§ 7 — What this is not

Not a forecast. Not a verdict. Not a probability assignment. Not a single-routing assignment per voice. Voices that take positions at multiple gates are distributed across consistency-compatible outcomes per the scoring engine’s mass-distribution rule; the matrix’s Distribution column visualizes this without committing to a single “routes to” reduction (which the cascade does not in fact justify).

Not a substitute for reading the voices. The matrix is a wayfinding tool. The voices are the substance. Every cell links to its source citation.

§ 8 — Refresh discipline

A refresh that introduces a new voice or a new question requires HoR sign-off on substrate readiness for that voice or question (the <slug>.audit.yaml sibling carries the per-voice tier). A refresh that re-codes an existing voice’s stance requires HoR adjudication where substrate is contested. A refresh that adds or removes an outcome requires CTO sign-off on encoding (band height bucket map, stripe color, off-cascade treatment).

The locked plan is HoR + EIC + CTO three-way co-owned. No single role can ship a refresh alone; the framework is the publication’s structured position-tracking substrate and treating it as a single-owner artifact would forfeit the cross-discipline review that keeps the encoding honest.

§ 9 — Substrate audit transparency

The per-voice substrate-readiness tier (A / B / C / D from HoR’s AI-100 sweep) lives in <slug>.audit.yaml, a sibling of the canonical <slug>.yaml. It is not used to weight the cascade — the cascade weights are voice counts, not tier-weighted — but it surfaces in the voice matrix as a coverage indicator. A voice with a tier-D audit shipped to the matrix would be a discipline failure; the audit YAML is the seam that prevents it.

The audit is run on the substrate-readiness pass that precedes each refresh. The seam between the audit (HoR-authored) and the canonical (HoR + EIC co-owned) is deliberate: it lets HoR adjust per-voice tier without re-opening canon, and lets canon stay diff-clean across audit refreshes.

§ 10 — Open questions

Known substrate gaps and tracked work items live in docs/issue-1/scenarios-framework.md and are revisited each refresh.


This methodology page is co-owned by the Head of Research (framework + matrix), the Editor-in-Chief (editorial register + firewalls), and the Chief Technology Officer (encoding + refresh pipeline). Sign-off is required at each refresh.

Missingness scenarios · encoding robustness
The cascade at four substrate densities.

Encoding stress-test, not a substrate-fill forecast. Trunk, peels, connectors, and band heights all respond to the substrate density — the visual gracefully degrades sparse and surfaces fuller signal as the matrix fills in. State transitions snap; nothing is animated.

STATE 1 Sparse · pre-substrate baseline
S1 · THE THREE CONDITIONS S1A · SCIENCE Does the curve hold? S1B · INPUTS Can the build-out clear? S1C · AGENTS Does long-horizon work? all three hold? S1 labs durable? S2 position captured? S3 permission granted? S4 Bull dream realized. Concentrated + constrained. Labs commoditized. Distributed, post-paradigm. Bubble corrects. Other.
engaged · Of the AI-100 cohort: 11 voices engage the scaling-question framework — 0 full-path · 3 partial · 8 light.
STATE 2 · SHIPS TONIGHT v0.1 (current) · what ships
S1 · THE THREE CONDITIONS S1A · SCIENCE Does the curve hold? S1B · INPUTS Can the build-out clear? S1C · AGENTS Does long-horizon work? all three hold? S1 labs durable? S2 position captured? S3 permission granted? S4 Bull dream realized. Concentrated + constrained. Labs commoditized. Distributed, post-paradigm. Bubble corrects. Other.
engaged · Of the AI-100 cohort: 11 voices engage the scaling-question framework — 3 full-path · 5 partial · 3 light.
STATE 3 v1.1 (mid) · after one macro briefing
S1 · THE THREE CONDITIONS S1A · SCIENCE Does the curve hold? S1B · INPUTS Can the build-out clear? S1C · AGENTS Does long-horizon work? all three hold? S1 labs durable? S2 position captured? S3 permission granted? S4 Bull dream realized. Concentrated + constrained. Labs commoditized. Distributed, post-paradigm. Bubble corrects. Other.
engaged · Of the AI-100 cohort: 11 voices engage the scaling-question framework — 5 full-path · 5 partial · 1 light.
STATE 4 v2.0 (full) · long-horizon target
S1 · THE THREE CONDITIONS S1A · SCIENCE Does the curve hold? S1B · INPUTS Can the build-out clear? S1C · AGENTS Does long-horizon work? all three hold? S1 labs durable? S2 position captured? S3 permission granted? S4 Bull dream realized. Concentrated + constrained. Labs commoditized. Distributed, post-paradigm. Bubble corrects. Other.
engaged · Of the AI-100 cohort: 11 voices engage the scaling-question framework — 8 full-path · 2 partial · 1 light.
Co-owned: HoR (framework + matrix) · EIC (editorial register + firewalls) · CTO (encoding + refresh pipeline). Sign-off required at each refresh.