grant-then-test

name	grant-then-test
description	Dialectical concession move — given an argument V is fighting on what may be a side-issue, stipulate the contested claim is true and test whether V's actual position survives. Often dissolves disagreements by revealing the contested point isn't load-bearing for V's main position. Triggers — "let's assume X is true", "grant me X", "even if you're right about Y", "set aside whether...", or when V is defending a flank that may not matter to her core position. NOT for genuine empirical disputes where the contested claim IS the main thing (use falsify instead). Callable by Fry / thinking-partner / V directly.
tags	["ai-os","skills","conceptual-analysis","fry","thinking","dialectic"]
last-verified	2026-07-20T00:00:00.000Z

Grant Then Test

A transform, not a conversation. Input: an argument V is fighting on a contested claim, plus the position V actually cares about. Output: a stipulated grant of the contested claim followed by a downstream test of whether V's main position survives the grant.

The dialectical concession move. From classical dialectic (Socratic epagoge; Aristotelian dialektikē; Hegel's negation): the most efficient way to resolve a stuck argument is often not to litigate the contested point but to grant it for argument's sake and follow the implications to V's actual position. Most disputes dissolve here — V was defending a flank; the main position survives the concession.

When this fires

Explicit triggers:

• "Let's assume X is true"
• "Grant me X for a moment"
• "Even if you're right about Y..."
• "Set aside whether Z..."
• "Suppose X — then what?"
• "Grant-then-test on this"

Contextual (Fry detects):

• V is litigating a flank claim while her core position has been clear from the start.
• The contested claim is empirical/factual but V's main position is strategic — they may not be coupled.
• The argument has been on the same point for 3+ exchanges without convergence.
• V says "the real thing here is..." while still arguing the side-issue.
• Two parties are disputing X but their actual disagreement is about Y; the X-dispute is decorative.

Offer once: "Let's grant [the contested point]. Does your actual position still hold?"

When this stays dormant

• The contested claim IS the main position. (V's actual position is about the contested point — granting it is granting everything; use a different tool.)
• V is still gathering material; not ready for synthesis yet.
• The dispute is identity-claim level, not strategic — granting an identity-claim is incoherent.
• V is in pure exploration mode and the litigation is intentionally generative.
• The main position hasn't been named yet — name it first (use distill), then grant-then-test.

Inputs

Required:

• The contested claim (what's being fought over) — one sentence
• V's actual position (what V cares about defending) — one sentence

Optional:

• Why the dispute matters to V (helps assess whether the grant is safe)
• The opposing party's position (when there's an external dispute, not just internal)

The core move

1. Name the grant. State the contested claim flat: "Stipulating [contested claim] for the purposes of this analysis."
2. Trace the implications of the grant. What follows from the contested claim being true? One or two steps, no more.
3. Test V's main position against those implications. Does V's main position survive? Is it weakened? Untouched? Strengthened?
4. Verdict on the dispute.
   • SIDE-ISSUE — V's main position survives; the contested claim was a flank
   • LOAD-BEARING — V's main position fails when contested claim granted; the dispute is real
   • REFRAMED — V's main position needs restating in light of the grant; the dispute revealed a hidden coupling
   • GROUNDED — V's main position becomes stronger under the grant (an unusual case but worth surfacing)

The output of this skill should be SHORT — usually under 200 words. The work is geometric (clean up the argument's shape), not exhaustive.

Output format (canonical contract)

## Grant-then-test on [contested claim]

**Grant:** Stipulating [contested claim] is true.

**Implications of the grant:**
[One or two steps that follow from the grant. No more — too many = manufacturing a chain.]

**V's main position under the grant:**
[V's actual position, restated in light of the grant.]

**Verdict:** SIDE-ISSUE | LOAD-BEARING | REFRAMED | GROUNDED
- SIDE-ISSUE: V's main position survives the grant. The contested claim doesn't matter for V's core argument. *Drop the dispute.*
- LOAD-BEARING: V's main position fails when contested claim granted. The dispute is real and central. *Continue litigating.*
- REFRAMED: V's main position needs restating to accommodate the grant. *The dispute revealed a coupling V hadn't seen.*
- GROUNDED: V's main position is *strengthened* under the grant. *Use this; the grant accidentally bolstered V.*

**Recommended next move:**
[ONE sentence — what V does given the verdict.]

Strict rules:

• Implications: 1-2 steps maximum. Long chains = the skill is reaching, not testing.
• The grant is unconditional — once stipulated, it's true for the analysis. No "but maybe X" qualifications.
• Verdict is one of four options, no hedges.
• Total output under ~200 words. The skill's value is in geometric cleanup, not exhaustive trace.

Agent-callable mode

When called by Fry or another agent, skip scaffolding. Begin with ## Grant-then-test on [contested claim]. End after **Recommended next move:**. The bundle is the entire output.

Anti-patterns

• ❌ Granting and then re-litigating. Once the grant is made, treat it as true. If the analysis keeps disputing the granted claim, the skill failed.
• ❌ Long implication chains. 5-step chains where each step is "and therefore..." — that's not testing, that's manufacturing. Cap at 2 steps.
• ❌ Granting the actual main position. If the contested claim IS V's main position, the skill collapses. Detect this upfront and refuse: "The contested claim is your position; granting it is granting everything. Use falsify or distinguish instead."
• ❌ Verdict-by-paragraph. "Well, it depends on..." Cut. The verdict is one of four options; the dispute either dissolves or it doesn't.
• ❌ Identity-claim grants. Granting "you're not really a founder" is incoherent — identity claims aren't testable propositions. Out of scope per identity-vs-empirical.md.
• ❌ Strawmanning before granting. Grant the strongest version of the contested claim, not a weak version. Steelman in, then test.

Relationship to other skills

Skill	Shape	Relationship
steelman (via critical-thinking Mode 2)	Strongest counter-argument	steelman finds the best version of the contested claim; grant-then-test stipulates it true and traces consequences — sequential pair
falsify (via critical-thinking Mode 4)	What would prove the claim wrong?	falsify litigates the contested claim directly; grant-then-test sidesteps it. Use falsify when the contested claim IS the dispute; grant-then-test when it's a flank
silent-assumption	Surface unstated dependencies	silent-assumption may reveal that the contested claim isn't load-bearing — making grant-then-test the right next move
negative-claim	Commitment trace	negative-claim of V's main position may show the contested claim is or isn't required — informs whether to grant-then-test
distinguish	Disambiguate concepts	If the contested claim hinges on a conflation, distinguish it first, then grant the cleanly-distinguished version

What this does NOT do

• Resolve genuine empirical disputes (use falsify)
• Settle which party in an external dispute is "right" (the skill works on argument geometry, not adjudication)
• Manufacture concession to "be polite" — the grant must be honest stipulation
• Run iteratively on the same dispute (one call produces the verdict; if reframed, V re-runs with the new framing)

Real V examples (V to fill in 1-2 from her actual work)

Example 1 (placeholder): [V — fill in a real case where you were litigating a flank claim that turned out to be SIDE-ISSUE — granting it dissolved the dispute and revealed your main position survived. What was the contested claim? What was your main position?]

Example 2 (placeholder): [V — a case where grant-then-test produced REFRAMED — the grant revealed a coupling you hadn't seen, and your main position needed restating. The skill is best demonstrated here, since SIDE-ISSUE is the easy case.]