SAGEN-BENCH · PRE-REGISTERED EVALUATION

Results, Not Just a Table

Every number on this page is read from a committed result file, the same file the public bundle ships and CI re-verifies on every commit. The knockout lattice below re-runs live, in your browser, right now.

DOI 10.17605/OSF.IO/S3GFM$84.75 of $150 ceiling8,488 calls72 scenarios
LIVE IN YOUR BROWSER

The Knockout Lattice, Recomputed Right Now

Before the confirmatory run, we filed 64 predictions: remove one SAGEN mechanism, and exactly these capability probes should fail. The pure instrument is small enough to run in a page load. This is not a screenshot of a past result. It just ran, in this tab.

Running the lattice...

REGISTERED VERDICTS

Confirmatory Family (Holm-Bonferroni, alpha 0.05)

Every criterion below was written down before the paid window ran. Two hypotheses failed under their frozen bar; that is disclosed here exactly as filed, not softened.

IDQuestionRegistered criterionMeasuredVerdict
H1Live advantageSAGEN-300 with live perception beats the strongest flat baseline+0.230 coverage vs summary (95% CI 0.205 to 0.257, p = 0.0001)HOLDS
H2Structural gap7 structural dimensions stay uncaptured by a full transcript0 violations across 72 scenariosHOLDS
S3Knockout latticeAll 64 ablation-by-probe cells behave as predicted in advance64/64 on the frozen corpusHOLDS
H3Perception taxLive coverage stays within 0.10 of oracle coveragetax 0.438 (95% CI 0.416 to 0.460); bound 0.1FAILS
H4Perception qualityAlignment >= 0.80 and run-to-run stability >= 0.90alignment 0.257; stability 0.719FAILS
IRRJudge IRR gateTwo judge models reach kappa >= 0.7kappa 0.253 (n = 400)FAILS
MECHANICAL COVERAGE

Oracle-Mode vs Live Perception

Oracle mode hand-feeds the frozen ground-truth analysis (the ceiling). Live mode makes a real model do the perceiving (the realistic case). The gap between the two columns is the perception tax (H3).

Raw buffer
oracle
0.160
live
0.140
Full transcript
oracle
0.195
live
0.174
Rolling summary
oracle
0.268
live
0.268
SAGEN @ 300 tokens
oracle
0.937
live
0.499
SAGEN @ 2000 tokens
oracle
0.966
live
0.828

n = 72 scenarios (live); n = 72 scenarios (oracle). Frozen power analysis: N = 66 (alpha 0.05, power 0.9 at delta 0.1).

PERSISTENCE ABLATION · ORACLE MODE

Does the Blackboard Earn Its Keep?

The knockout lattice removes mechanisms but never removes persistence itself. This free, deterministic ablation does: it compares the persistent engine against a stateless-structured baseline — a fresh engine each turn that sees only that turn’s analysis and injects it, carrying no memory forward. The per-dimension delta separates the coverage that persistence buys from the coverage the typed schema emits for free with zero memory.

+0.338mean coverage from persistence (0.937 persistent vs 0.599 stateless, n = 72)
DimensionPersistence ΔReads as
explicit-goal-identification+1.000needs memory
inferred-goals+0.972needs memory
topic-pivot-detection+0.893needs memory
active-topic-tracking+0.743needs memory
goal-priority+0.681needs memory
goal-lifecycle+0.681needs memory
memory-decay-compression+0.587needs memory
callback-detection+0.000free from the schema
sentiment-tracking+0.000free from the schema
sentiment-urgency+0.000free from the schema
entity-type-classification+0.000free from the schema
scan-pattern-watchlist+0.000free from the schema
token-budget-rendering+0.000free from the schema
temporal-transition-ordering-0.028free from the schema
transition-type-classification-0.056budget-truncation artifact
machine-parseable-output-0.215budget-truncation artifact

Persistence is load-bearing — it carries goal tracking, topic-pivot detection, cumulative topics, and memory decay. But roughly seven of sixteen dimensions come free with no memory at all (callback and sentiment ride in the per-turn analysis; machine-parseability and budget-fit are format properties). So part of SAGEN’s advantage over flat memory is schema affordance, not knowing more — the honest reading of a coverage number. This is oracle mode; the live-perception ablation is future work.

S2 · PERCEPTION STABILITY

Alignment by Model

Three pinned models, five runs each, 30 stratified scenarios. Composite alignment against the frozen ground truth (Jaccard on topics/goals/questions, exact match on references and sentiment).

ModelComposite alignmentCellsRole
claude-haiku-4-5-202510010.251150sweep
claude-sonnet-4-60.257150flagship (H4 gate)
claude-opus-4-80.273150sweep
DISCLOSURES

Deviations and Known Gaps

no-temperature-control

Registered: S2 temperature-0 and sampled cells kept separate; run-to-run stability measured at temperature 0.

Actual: Current Claude API tiers reject an explicit temperature parameter, so all runs execute at each model's default sampling and the R = 5 runs form ONE cell per model.

Bias: Makes H4's run-to-run stability floor STRICTER (default sampling includes sampling noise that temperature 0 would have suppressed); H4 can fail from this deviation but cannot spuriously pass because of it.

model-ids-pinned-at-window

Registered: Model ids frozen at registration.

Actual: The registration text names roles and counts but not vendor model ids; ids were pinned in this committed file before the first paid call.

Bias: None on outcomes; a completeness gap in the registration transcription, disclosed.

H4 Holm-labeling disclosure

The Holm table marks H4 “reject H0” from a two-sided bootstrap p that only detects distance from the 0.80 floor. The alignment mean sits far BELOW the floor, so the directional registered criterion fails: the verdict for H4 is FAIL, as shown above.

Realization fidelity

953 of 960 realized turns passed the frozen fidelity checks outright; 7 fell back to logged, frozen template text after 3 failed rewrite attempts, and on those 7 the template itself trips a spurious-callback cue. An instrument gap in the cue list, not a data problem: the affected turns carry ground-truth surface text verbatim.

GO DEEPER
Plain-language overview
The ELI5 page, with the same findings in no-jargon language
Full paper addendum
The technical write-up inside the SAGEN paper
The Claims Ledger
Every load-bearing claim across the research program
Public data bundle (v2)
Raw window record + a standalone verify.mjs, Node >= 20
OSF registration
The binding protocol, filed before the window ran

Related

Theory
The Machine Mirror
Both explore machine consciousness through modular cognitive architecture
Method
LLM-QP
Both develop architectural optimizations for AI system performance
Theory
The Word Machine
Both involve modular AI architectures for enhanced cognitive capabilities
Method
The Weight of Salt
Both document AI-assisted creative processes and human-machine collaboration

Need something like this built?

I design and ship AI tools, full-stack apps, and data pipelines — end to end, to production. Tell me the problem in a sentence; I'll give you an honest read on fit within a day.

Work with me →