This note consolidates the operational modes of DIRAC — established piecewise across Prop. 1, Prop. 2 and Prop. 3 — into a single framework. A DIRAC mode is defined in Def. 5; this note enumerates the modes, their transforms, and their status.
Framing
DIRAC compiles a directed causal model into an energy landscape (Prop. 1) and then reshapes that landscape to answer a causal query; each mode (Def. 5) produces one job — the pair
The preceding phases produced three distinct reshaping operations — a clamp penalty (Prop. 2), a mechanism swap (Prop. 3), and its soft/global specialisations — this framework collects them as one interface. Its job is to make the comparison between modes measurable.
The modes
Mode 0 — Joint sampling
- Query: none — the unconditioned joint
. - Transform: none; compile
and sample directly. - Procedure: map each CPT to
; sample the resulting energy model. - Exactness: exact, and
by Prop. 1 — a free correctness check. - Status: reference. The exact baseline, and the natural first implementation target on toy graphs (chain, fork, collider — factor census) before any intervention is added.
Mode 1 — Evidence Clamp
- Query:
— observational conditioning (level 1 in causal hierarchy of The Book of Why by Judea is seeing). - Transform:
, with the evidence indices and target (clamping). - Procedure: keep the original graph; add a diagonal bias of strength
pushing each evidence variable toward ; sample . - Exactness:
as , with finite- leakage and (Prop. 2). Choice of follows the floor/ceiling/stop rule of Phase 5. - Status: reference — proven observational conditioning.
Mode 2 — Hard Intervention Reference
- Query:
— hard intervention (level 2 : doing). - Transform: delete the incoming edges of
and replace its factor by the point mass ; equivalently with (Prop. 3, Def. 8 hard case). - Procedure: mutilate the DAG, install the point-mass CPT, recompile, sample.
- Exactness: exact in the ideal model (
). The point mass reintroduces zero entries, so its implementation inherits the / treatment of note 2. - Status: reference — the standard causal answer against which the physical pathways are judged.
Mode 3 — Soft Intervention Reference
- Query: mechanism replacement
, or the parent-free . - Transform: swap the factor;
with (Prop. 3 soft case). - Procedure: install
, recompile, sample. - Exactness: exact at every scale when
is strictly positive — no limit and no leakage, unlike Mode 1. - Status: reference.
To be tested
Pure clamping on the unmutilated graph reproduces conditioning, not intervention: its distance to the do-reference flattens at the confounding-bias floor that no clamp strength closes. Recovering genuine intervention on the hardware therefore needs a mechanism that removes the incoming factor.
Mode 4 — Physical Clamp Test
- Intuition: realise the intervention physically — reconfigure the backend so the parent-to-
coupling terms vanish ( for ), performing graph mutilation directly rather than recompiling a mutilated model, then sample. - Aim: reproduce the Mode 2 reference by physical surgery.
- Open: whether the backend supports the reconfiguration at all, its latency and cost, and whether the sampled distribution matches Mode 2. Not fixed — intuition only.
- Status: to be tested.
Mode 5 — Virtual Mutilation
- Intuition: leave the couplings untouched and inject a compensating field
that cancels the incoming mechanism, reproducing mutilation virtually, then sample. - Aim: reproduce the modes 2, 3 & 4 without rewiring.
- Open: feasibility within dynamic range, the higher-order/auxiliary cost of the order-
compensating term (note 3), and fidelity to Mode 2. Not fixed — intuition only. - Status: to be tested.
Summary
| Mode | Causal query | Transform on | Exact? | Status | Result |
|---|---|---|---|---|---|
| 0 — Joint sampling | none (baseline) | none; sample | exact, | reference | Prop. 1 |
| 1 — Evidence Clamp | limit-exact, | reference | Prop. 2 | ||
| 2 — Hard Intervention Reference | drop parent edges + point mass | ideal-exact; zeros → | reference | Prop. 3 | |
| 3 — Soft Intervention Reference | mechanism | swap factor for | exact at every scale ( | reference | Prop. 3 |
| 4 —Physical Clamp Test | realise | reconfigure | open — intuition | to be tested | — |
| 5 — Virtual Mutilation | realise | compensating field | open — intuition | to be tested | — |
The reference rows fix what the answer should be, still to be confirmed in a tryout; the two to-be-tested rows are the candidate ways of producing intervention on the hardware.
Deterministic entries:
Any mode whose compiled energy contains a zero CPT entry — Mode 2’s point mass, and deterministic base CPTs generally — carries an infinite barrier that no finite-range backend can hold.
Realisation as a job
Every mode compiles to a single job — the pair
Commentary