# Conceptual guardrails

Points that are easy to get wrong when reviewing outside work on IIT or
answering an objection. Each guardrail is stated by Giulio Tononi (2026),
followed by where the corpus supports it.

**Evidence comes from ourselves.** IIT is consciousness-first: it is grounded in
our own experience and validated on its substrate in us. Never describe the
source of evidence as "systems that can report" or "report".
*Sources:* [Tononi & Boly 2025](https://learniit.org/references/tononi-2025b), "Introduction:
experience as intrinsic existence" (the empirical validation of IIT is
"established in ourselves"; its attribution to others "is necessarily a matter
of inference") and "Inferring consciousness beyond adult humans".

**Inference, not prediction.** IIT's verdict on a system we cannot validate
against (a computer, a feedforward equivalent, an organoid) is an inference
from a good explanation validated in ourselves. A behaviorally matched system
without integrated information is neither a behavioral "prediction" to be
tested nor a counterexample. Treating it as either accepts the premise of the
unfolding argument.
*Sources:* [Tononi & Boly 2025](https://learniit.org/references/tononi-2025b), "Some implications
of IIT: meaning, perception, and matching" ("inferences from a good
explanation") and "Inferring consciousness beyond adult humans"; the ledger
entry on [the unfolding argument](https://learniit.org/ledger/unfolding-argument).

**Large Φ in a regular grid is expected, not a reductio.** IIT concluded this
already in 2.0, and the 2014 reply to Aaronson and the extendedness papers
(Haun & Tononi 2019; Grasso et al. 2021, 2026) treat lattice Φ-structures as
what accounts for spatial experience. A realistic large complex also needs
near-determinism with some indeterminism, the right grain, and the high
fan-in/fan-out that physical grids lack.
*Sources:* [Balduzzi & Tononi 2008](https://learniit.org/references/balduzzi-2008);
[Tononi 2014](https://learniit.org/corpus/tononi-2014a), "Why certain grids may be conscious";
[Haun & Tononi 2019](https://learniit.org/references/haun-2019);
[Grasso et al. 2021](https://learniit.org/references/grasso-2021b);
[Grasso et al. 2026](https://learniit.org/references/grasso-2026);
[Tononi & Boly 2025](https://learniit.org/references/tononi-2025b), "Grids" and "IIT and the
richness of experience" (cooperative but specialized macro units, "a large
fan-in/out to promote integration", "a sufficiently regular lattice");
[Mayner, Marshall & Tononi 2026](https://learniit.org/references/mayner-2026), §4 (the tradeoff
between determinism and indeterminism); [Marshall et al. 2026](https://learniit.org/references/marshall-2026)
(grain).

**The human main complex is posterior-central.** Say "posterior-central
integrated information", not "posterior cortical integration".
*Source:* the term [Tononi & Boly 2025](https://learniit.org/references/tononi-2025b) use
throughout, e.g. "IIT and the richness of experience" ("a large main complex
in the human brain, most likely in posterior-central cortex").

**Inactive vs. inactivated.** Inactive units in the complex contribute to the
Φ-structure; inactivated units do not. Don't conflate them.
*Sources:* the [IIT wiki](https://learniit.org/corpus/iit-wiki), "Validate empirically in humans"
and its footnote 1; [Tononi & Boly 2025](https://learniit.org/references/tononi-2025b), "The
quality of experience: space is extended" (a largely inactive main complex in
a state of causal readiness).

**Boundary and grain are outputs of exclusion**, never chosen by the
investigator before the analysis.
*Sources:* the [exclusion postulate](https://learniit.org/glossary/exclusion);
[Albantakis et al. 2023](https://learniit.org/references/albantakis-2023b), "Exclusion:
Determining maximal substrates (complexes)" and "Determining maximal unit
grains", Eqs (24)–(26); [Marshall et al. 2026](https://learniit.org/references/marshall-2026),
"Intrinsic units".

**Hardware, not software.** Candidate units are physical; weights and
activations are virtual. Changing a network's architecture on fixed hardware
does not change intrinsic causal organization, and perturbational measures on
model activations do not test IIT.
*Sources:* [Tononi & Boly 2025](https://learniit.org/references/tononi-2025b), "Intrinsic
meanings vs. extrinsic referents" (an LLM's units, "being simulated on a
computer, are virtual"); [Findlay et al. 2025](https://learniit.org/references/findlay-2025),
"Cause–effect structures specified by the computer can be dissociated from
those of target systems it is simulating". The last clause, about
perturbational measures, follows from the units being virtual; no source states
it in those words.

**Know which φ_s an outside paper uses.** Under the current φ_s (Mayner,
Marshall & Tononi 2026) a purely deterministic system has φ_s = 0. Many outside
results still use 2023 φ_s.
*Sources:* [φ_s](https://learniit.org/glossary/phi-s), including "Before 2026";
[Mayner, Marshall & Tononi 2026](https://learniit.org/references/mayner-2026), §2.2–2.3;
[what is current](https://learniit.org/current).

**Check before calling something an objection.** Much outside criticism
restates what IIT already holds; search the corpus and the ledger first.