Skip to content

Commit 4ab1db6

Browse files
claudeclaude
committed
plan(mul-ewa): add F-MEP-0a (capability gate) and seam to the quadrant plan
Operator ruling: 262e22d was a good emergency brake but stopped one layer early. Covariance-vs-precision fixes the OPERATOR; it does not fix what the operator is PERMITTED TO DO. Get the sign right, normalise hop count, bound the inflation, obtain a gorgeous AUC -- and still have built a flawlessly calibrated machine for performing the wrong epistemic operation. F-MEP-0a now gates F-MEP-0b: D-MEP-1 must declare the EWA sandwich's capabilities before specifying it. Redistribute/locate tension: yes. Rank where counterfactual exploration is worth compute: yes. Generate candidate completions: only marked hypothetical. Eliminate worlds by itself: no. Mint evidence: never. Raise empirical trust from its own output: never. Separates three things that were collapsing into the word "uncertainty": epistemic state (provenance), geometric uncertainty (Sigma/EWA/ Oberflaechenspannung), and counterfactual ambiguity (mutually incompatible completions). Sigma represents the second, cannot manufacture the third, and the third must never silently overwrite the first. Sharpens the anti-circularity invariant: no operation whose inputs are entirely graph-internal may convert a hypothesis into epistemic evidence merely by reducing its uncertainty. Internal reasoning CAN create knowledge about the hypothesis space -- constraint elimination and contradiction are real -- so the ban is on laundering that into a claim about external reality. Three mechanisms kept apart: elimination by logical contradiction (may raise logical certainty), discrimination by observation (may raise empirical certainty), preference by propagated geometry (raises neither; changes only where to look). Scope-corrects 0b's own phrase. "Lower trust yields higher suspicion" is a sign check on the operator, NOT an ontology: low trust != uncertainty, high covariance != epistemic ignorance, high entropy != intelligent doubt. MUL compares epistemic posture against epistemic support, so Overconfident means the graph collapsed onto one explanation while viable alternatives were never discriminated. THE SEAM, which is the real finding. This is not new architecture -- epistemic-quadrant-materialization-v1.md (709 lines) and probe_sudoku_teacher.rs (1859 lines) already implement the membrane. Its G3: bifurcation clones the slab as a counterfactual world, propagates to contradiction, and ONLY THE ELIMINATION RETURNS; the fork's positive guess never enters the main slab. G4 measures the cost of refusing to fork. Live surfaces: contract::scenario::ScenarioBranch, contract::pearl_junction. That plan carries MUL x9 and ZERO EWA/TrustTexture/Oberflaechenspannung; this plan carried zero quadrant vocabulary; neither cited the other. Two halves of one architecture in two plans with no seam -- the rediscovery tax CLAUDE.md "Consult before you guess" exists to prevent. D-MEP-1 must read it before specifying Sigma. Adds F-MEP-0d, the closed-graph monotonicity falsifier: no new observations, no oracle calls, run MUL -> EWA -> thinking -> MUL for 1/10/1000 cycles. Reorganised attention, hypotheses, sharper questions and latent contradictions are all permitted; aggregate epistemic authority rising with cycle count is not. A STOP gate rather than a diagnostic, run on synthetic closed graphs before W1's cohort since it needs no S4 events and can fail early and cheap. Supersession index regenerated.
1 parent 262e22d commit 4ab1db6

1 file changed

Lines changed: 146 additions & 0 deletions

File tree

‎.claude/plans/mul-ewa-trust-propagation-v1.md‎

Lines changed: 146 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -29,6 +29,152 @@ anything: **does sandwich-propagated uncertainty rank multi-hop derived
2929
beliefs differently — and better — than the naive scalar decay every stack
3030
defaults to?** If not, NO-BUY, numbers banked.
3131

32+
## §0a The permission gate — what epistemic operation is this operator ALLOWED to perform?
33+
34+
**(Operator ruling, 2026-08-29. This gate comes BEFORE F-MEP-0b. Settling
35+
covariance-vs-precision fixes the OPERATOR; it does not fix what the operator
36+
is permitted to DO. Get the sign right, normalise hop count, bound the
37+
inflation, obtain a gorgeous AUC — and you can still have built a flawlessly
38+
calibrated machine for performing the wrong epistemic operation.)**
39+
40+
### Three things currently collapsing into one word, "uncertainty"
41+
42+
| # | layer | carrier | what it answers |
43+
|---|---|---|---|
44+
| 1 | **Epistemic state** | provenance: observed / inherited / absent | what is actually known, and *how* it came to be known |
45+
| 2 | **Geometric uncertainty** | Σ, EWA, *Oberflächenspannung* | where tension propagates and concentrates |
46+
| 3 | **Counterfactual ambiguity** | `{W₁ … Wₙ}` — mutually incompatible completions | which admissible worlds still explain the evidence |
47+
48+
**Σ can represent #2 beautifully. It cannot manufacture #3. And #3 must never
49+
silently overwrite #1.** Everything below follows from those two sentences.
50+
51+
### F-MEP-0a — the capability declaration (gates F-MEP-0b)
52+
53+
D-MEP-1 must state, before the operator is specified at all, which of these
54+
the EWA sandwich is permitted to perform:
55+
56+
| operation | permitted? |
57+
|---|---|
58+
| redistribute / locate epistemic tension | **yes** |
59+
| rank places where counterfactual exploration is worth compute | **yes** |
60+
| generate candidate completions | **maybe — marked HYPOTHETICAL, never merged into #1** |
61+
| eliminate worlds by itself | **no** |
62+
| mint evidence | **never** |
63+
| raise empirical trust from its own output alone | **never** |
64+
65+
EWA may say *"given what we currently know, unresolved tension concentrates
66+
HERE."* It may **not** say *"therefore this region is now better known."*
67+
68+
### The anti-circularity invariant, in its sharp form
69+
70+
> **No operation whose inputs are entirely graph-internal may convert a
71+
> hypothesis into epistemic evidence merely by reducing its uncertainty.**
72+
73+
This SUPERSEDES the softer §0b wording. Internal reasoning genuinely *can*
74+
create knowledge about the hypothesis SPACE — constraint elimination can prove
75+
only two worlds remain, and a contradiction can kill a world with no new
76+
observation. That is real and must not be forbidden. What is forbidden is
77+
laundering it into a claim about external reality. Three kinds, kept apart:
78+
79+
| mechanism | may increase |
80+
|---|---|
81+
| hypothesis **elimination by logical contradiction** | logical certainty |
82+
| hypothesis **discrimination by observation** | empirical certainty |
83+
| hypothesis **preference by propagated geometry** | **neither** — it changes only where to look |
84+
85+
### ⊘ Scope correction to F-MEP-0b's own wording
86+
87+
§0b's gate phrase *"lower per-hop trust yields strictly HIGHER suspicion"* is
88+
**a sign check on the operator, NOT an ontology.** It is necessary and it is
89+
not a definition. Explicitly:
90+
91+
- low trust **≠** uncertainty
92+
- high covariance **≠** epistemic ignorance
93+
- high entropy **≠** intelligent doubt
94+
95+
Those are signals MUL may CONSUME. They are not what MUL IS. MUL's actual job
96+
is comparing epistemic **posture** against epistemic **support** — so
97+
`Overconfident` means *the graph collapsed onto one explanation while viable
98+
alternatives were never discriminated*, and a high-uncertainty reading can be
99+
perfectly **`Calibrated`** when several coherent worlds genuinely remain.
100+
101+
### The ordering this implies — geometry decides WHERE, never WHAT
102+
103+
```
104+
EWA "there is epistemic tension here" (attention geometry)
105+
HHTL "these completions are admissible" (constraint field)
106+
thinking style "here are competing explanations" (the intellectual act)
107+
counterfactual "they disagree at S*" (differential probe)
108+
REALITY "S* = x" ← the only evidence arrow
109+
MUL "NOW you may update trust"
110+
```
111+
112+
Thinking styles belong **after** EWA exposes tension, never inside the
113+
interpolation. A renderer is rewarded when an interpolation looks right; a
114+
reasoning system is in deep trouble when a plausible interpolation silently
115+
graduates into a fact.
116+
117+
### ★ This is NOT new — the membrane is already implemented (seam, not invention)
118+
119+
**`.claude/plans/epistemic-quadrant-materialization-v1.md` (709 lines) and
120+
`lance-graph-planner/examples/probe_sudoku_teacher.rs` (1859 lines) already
121+
build the counterfactual half of this**, and #1074 cited them **zero** times
122+
until now. That plan's **G3** is exactly the membrane above:
123+
124+
> *"Bifurcation clones the slab as a counterfactual world, propagates to
125+
> contradiction, and **ONLY the elimination returns** — the main slab changes
126+
> at exactly the sanctioned cell, and the fork's positive (wrong) guess never
127+
> appears in it."*
128+
129+
The fork's NEGATIVE result (an elimination) may cross back into reality; its
130+
POSITIVE guess may not. Its **G4** then measures the cost of refusing to fork
131+
(a fork-refusing policy on a bifurcation-required puzzle does not fully
132+
migrate toward `Wisdom`) — the can-it-fire twin. Related live surfaces:
133+
`contract::scenario::ScenarioBranch`, `contract::pearl_junction`.
134+
135+
**The split is the actual defect.** That plan carries `MUL` (9 mentions) and
136+
**zero** `EWA` / `ewa_sandwich` / `TrustTexture` / *Oberflächenspannung*; this
137+
plan carried zero of the quadrant vocabulary. Two halves of ONE architecture,
138+
in two plans, with no seam — the rediscovery tax `CLAUDE.md` § "Consult before
139+
you guess" exists to prevent. **D-MEP-1 must read that plan before specifying
140+
Σ**, and any counterfactual construction here routes through its existing
141+
fork/elimination discipline rather than a second mechanism.
142+
143+
### The pothole is a TYPED ABSENCE, never a filled cell
144+
145+
An epistemic hole is not missing data awaiting interpolation. It is a
146+
first-class object carrying its own structure:
147+
148+
```
149+
UNKNOWN {
150+
alternatives = {A, B, C} // admissible completions
151+
constraints = {...} // what narrows them
152+
discriminators= {...} // signals on which they disagree
153+
evidence_needed = {...} // what reality would have to supply
154+
}
155+
```
156+
157+
*"Reality is known up to this equivalence class of worlds"* is **knowledge**,
158+
not a gap. Filling it with a smooth field destroys the very structure that
159+
makes it answerable.
160+
161+
### F-MEP-0d — the closed-graph monotonicity falsifier (testable, and it bites)
162+
163+
Take a graph with **no new observations, no oracle calls, no evidence
164+
receipts.** Run MUL → EWA → thinking-style → MUL for `1`, `10`, `1000` cycles.
165+
166+
**Permitted:** reorganised attention, generated hypotheses, sharpened
167+
questions, contradictions discovered that were already latent.
168+
169+
**FORBIDDEN:** aggregate epistemic authority rising monotonically with cycle
170+
count. If trust inflates merely because the system thought longer, the seam is
171+
a philosophical perpetual-motion machine and the answer is NO-BUY regardless
172+
of every other number in this plan.
173+
174+
This is the testable form of the §0b circularity worry, and it is a **STOP
175+
gate, not a diagnostic**: it runs on synthetic closed graphs before W1's real
176+
cohort, because it needs no S4 events and can therefore fail early and cheap.
177+
32178
## §0b The two hinges, and the danger of handing the epistemic frontier to math
33179

34180
**(Operator framing, 2026-08-29 — recorded because §0-§4 specified a

0 commit comments

Comments
 (0)