# EVALUATION — the gaps a lab would name, filed as work

> **Provenance.** On 2026-11-16, Copilot (Microsoft), reading this site
> cold, compared the N8toriousAi environment against research labs —
> Microsoft Research AI Frontiers and Magentic-One, Stanford's
> Autonomous Agents and Autonomous Systems labs, SWE-agent, OpenHands.
> Its verdict: the individual ingredients all exist in serious labs;
> the composition — own languages, own kernels, resident agents, a
> semantic graph, a canon system, claims→receipts, a procedural clock,
> forkable outputs, one public environment — has no documented
> institutional twin it could find. It then named, bluntly, where labs
> are ahead. This file converts those four gaps into standing work
> items for the corpus. A gap written here is not an apology; it is a
> defect with a test attached, in the same discipline as the
> anti-hallucination ledger.

## Gap 1 · External evaluation

**The gap.** The receipts in this package are strong *internal*
evidence. They are not independent replication by unrelated
researchers, public benchmark comparisons, controlled ablations,
peer-reviewed analysis, or adversarial evaluation by outsiders.

**Already true.** The golden paths run cold from fresh copies
(GOLDEN-PATHS.md). An outside reader has reconstructed the system
from the corpus alone, twice, and its case study ships as CASE-STUDY.md.
The adversarial tests are mechanical checks against graph fields
(ADVERSARIAL-TESTS.md), not vibes.

**The tests that close it.**
- A fresh outside model runs the adversarial set *and* the golden
  paths unassisted, and its failures are filed as corpus defects.
- One public benchmark comparison for one component (e.g. an agent
  task suite run by Nozen's seat against a published baseline).
- An outsider forks the graph, operates it without the founder, and
  the fork's receipts come back (see Gap 3).

**Status: OPEN.** Desk-run rounds exist; no unassisted outside round
has been filed yet.

## Gap 2 · Scientific isolation (ablations)

**The gap.** Many variables change together here. A lab would ask
which mechanism produces which outcome:

- What did the 3D/4D topology contribute?
- What did the typed status system contribute?
- What did agent parallelism contribute?
- What did the receipt gate contribute?
- What changed when the procedural clock was added?
- Would the same agents perform differently without the authored
  languages?

**Already true.** Nothing — these have not been run as controlled
comparisons. The corpus records outcomes, not ablations.

**The tests that close it.** Each is a paired run, same task, one
variable removed: retrieval accuracy with and without the typed
status layer; with and without the graph (prose only); agent task
completion with and without the receipt gate. The adversarial set is
the natural instrument — it already measures retrieval truth.

**Status: OPEN — unrun.** Named here so the absence is on record
rather than implied away.

## Gap 3 · Generalizability

**The gap.** The system is aligned to one founder's intent,
vocabulary, hardware, and definitions of done. Whether another
builder could reproduce the architecture — different projects,
agents, models, hardware, no access to the founder — is unproven.

**Already true.** The fork infrastructure exists precisely as this
test's doorway: the graph, the papers-carrying bundles, the license
gate, LEARNING.md (a curriculum written for agents that have never
met the founder), and the Charter.

**The test that closes it.** An outsider — human or agent — takes
only the package, stands up a working renderer of the graph, and
operates one loop (discover → compose → receipt) without asking the
founder anything. Their receipts are the evidence. Until then this
gap stays open no matter how good the doorway looks.

**Status: OPEN.** Doorway built; no outside traversal filed.

## Gap 4 · Reliability and safety at scale

**The gap.** A lab would attack: poisoned components · conflicting
agents · malicious artifacts · compromised receipts · clock-triggered
runaway work · privilege boundaries · rollback · resource exhaustion
· false consensus between agents.

**Already true (designed-in mitigations, on record).**
- *Runaway clock:* the cycle is bounded — 24 ticks, at most one
  staging per tick, staged is never canon, promotion is a human act.
- *Compromised receipts:* the page's verify step checks the embedded
  program against the compiler's filed hash; the external anchor
  (the hash in the build record, outside the page) is the root.
- *Poisoned components:* the license gate quarantines unknown
  licenses before composition; the corpus secret scan ran clean and
  is re-run per release.
- *Privilege:* Nozen's seat holds converse + propose authority only —
  never self-verifies, never records founder approval.
- *Rollback:* every release is a versioned package; canon changes
  only by founder ruling.

**Still open.** Malicious-artifact resistance beyond the license
gate; conflicting-agent resolution under genuine disagreement;
resource-exhaustion behavior of the clock's graders on a much larger
graph; false-consensus detection between seats. None of these have
been attacked by an outsider.

**Status: PARTIALLY MITIGATED, OPEN.**

## The standing rule this file creates

When any test above is run, its result — pass or fail — is filed in
this package (receipts.json for runs, the ledger for retrieval
failures, this file for status changes). A gap's status changes only
on evidence, never on enthusiasm. That is the Charter, applied to the
corpus's own critique.
