Check the ambition against the published record, not just against this lab

README.md

@@ -37,6 +37,10 @@ autonomous-lab throughput single_cell_genomics  # plates/day, or why that number
 autonomous-lab provenance single_cell_genomics  # what could be proven about a run afterwards
 autonomous-lab lineage single_cell_genomics   # which cell did this read come from?
 autonomous-lab knowledge                      # encoded expert judgment and robot benchmarks
+
+autonomous-lab evidence                       # what the published record shows, and what it does not
+autonomous-lab authority                      # which decisions a model may make, and which need a person
+autonomous-lab cadence 500 --endpoint started --hours 16  # what that target actually costs
 ```
 
 ## What it reports today
@@ -318,6 +322,92 @@ adequacy. The STAR's dry motion is validated; its volumetric accuracy at the wor
 has never been measured, and no camera retires that benchmark. It needs a calibration
 experiment.
 
+## Is the ambition supported by anything published?
+
+Every layer above asks what *this* lab can prove. Three more ask a question none of them
+can, because it is not about this lab: **is the plan supported by anything published, who
+is allowed to decide, and what does a stated throughput target actually cost?**
+
+```
+autonomous-lab evidence                      # what the record shows -- and what it does not
+autonomous-lab authority                     # which decisions a model may make
+autonomous-lab cadence 500 --endpoint started --hours 16
+```
+
+### Evidence, scoped rather than cited
+
+Autonomous-lab claims get justified by pointing at famous papers whose scope is far
+narrower than the claim. A materials-synthesis campaign gets cited to justify a cell-assay
+production line. So every entry carries what it **does not establish**, and support is
+resolved by scope rather than by fame.
+
+Fifteen works, each independently verified against the live record before it was encoded --
+**ten confirmed, five partial, and eleven carrying figures that could not be confirmed**,
+listed rather than rounded away. An entry whose numbers were checked against a preprint
+rather than the version of record is `PARTIAL`, however solid the work is. The verification changed the content:
+
+- The A-Lab paper was **corrected**. Forty-one successes became thirty-six after four were
+  reclassified as inconclusive, and the "43 new materials" figure in news coverage matches
+  no version of the paper.
+- Assay Guidance Manual chapters carry **no DOI**, so any citation supplying one is wrong.
+- PyLabRobot makes an *architectural assertion* about cross-family simulation. It does not
+  demonstrate it, and the entry says so.
+
+`unsupported_claims()` is the standing list of things **no** cited work establishes. It is
+computed from the table, so it shrinks on its own the day real evidence arrives.
+
+### Authority: which decisions a model may make
+
+"AI-native" gets read as "the model drives the instruments". The benchmark evidence is that
+**no evaluated model exceeded 70% hazard-identification accuracy**, and fluent output is
+not evidence of laboratory safety competence. But the honest position is not that models
+are useless -- it is that decision classes differ.
+
+Thirteen classes across four levels: `MODEL` may decide alone, `MODEL_PROPOSES` needs a
+deterministic check or a human to ratify, `DETERMINISTIC` needs a coded interlock with no
+model in the path, and `HUMAN` needs a named accountable person. **Authority may be
+strengthened and never weakened** -- putting a person on a decision a model could make is a
+cost, and the reverse is the defect `violations()` exists to catch.
+
+Each class states *why* it sits where it does, against a specific published finding. Triage
+sits at `MODEL_PROPOSES` because of a documented case where an unexpected intramolecular
+cyclization carried the same molecular weight as the expected product, was indistinguishable
+by chromatogram or MS, and passed the automated rule -- until a human read the NMR.
+
+The qualification ladder runs simulation -> device acceptance -> integrated dry run ->
+volumetric qualification -> assay transfer -> sustained production. **No rung implies the
+one above it.** A passed simulation qualifies nothing physical.
+
+### Cadence: what a throughput target actually costs
+
+A target is the easiest number in a lab to state and the easiest to state meaninglessly.
+`cadence` refuses one until it says what it counts, because plates *started*, *completed*,
+*QC-passed*, and *released* differ by exactly the failure and rerun rate -- a measurement,
+not an assumption. Resolving that silently would pick the loosest reading in the caller's
+favour. It refuses twice: a downstream endpoint is a **return** rate, so reaching 500
+released means admitting more than 500, and without a measured survival fraction the module
+will not convert. Dividing the face value would report the *started* interval under a
+*released* label.
+
+Given an endpoint it is pure arithmetic on the caller's inputs, needing no measured
+durations: 500 plates a day is one every **172.8 s** round the clock, or **115.2 s** inside
+a sixteen-hour window. The ratio between those is the walk-away penalty, and it depends on
+the clock alone -- not on the target, protocol, or instruments, which is the whole
+quantitative case for closing the steps that need a human.
+
+It then names what the interval forces, and refuses where the input does not exist: a
+buffer depth is a makespan divided by an interval, so it inherits every unmeasured second
+and is reported as not computable rather than estimated. `headroom()` is where "just buy a
+faster robot" dies -- an intervention rate of one per two plates at ten minutes each
+consumes **260%** of a 115.2 s interval, and instrument speed is not in that arithmetic.
+
+And `demonstrated_support()` answers the question everybody skips. No cited system
+establishes integrated multi-domain operation at a production rate: the longest integrated
+campaign in the record is seventeen days in a single domain, and the one system running
+continuously at scale across institutions is a cage sensor that observes rather than
+actuates. That emptiness is **computed from the evidence table**, so it flips the day
+somebody publishes a counterexample rather than staying a stale assertion.
+
 ## The RE queue is computed, not argued about
 
 ```
@@ -472,7 +562,7 @@ event receiver and silently steals the first one's callbacks.
 pip install -e '.[dev]' && pytest
 ```
 
-149 device-free tests. The ones that matter most try to make the layer lie:
+292 device-free tests. The ones that matter most try to make the layer lie:
 
 - claim a step is automated when its command is undecoded; claim a decoded command is
   runnable while its siblings are not; claim a federated leg runs when no run card was ever

autonomous_lab/__init__.py

@@ -35,6 +35,20 @@
   lineage       which cell a read came from -- and whether pooling already destroyed the
                 answer, which is a different question from whether the plate was tracked.
   intelligence  the tacit expert judgment and the benchmarks a robot must meet first.
+
+Three further layers ask the question the others cannot, because it is not about this lab
+at all: is the ambition supported by anything published, who is allowed to decide, and what
+does a stated throughput target actually cost?
+
+  literature    what the published record demonstrates -- and, entry by entry, what it
+                explicitly does NOT establish. Scope, not fame, decides whether a paper
+                supports a claim.
+  authority     which decisions a model may make, which need a deterministic interlock, and
+                which need a named accountable human. Authority may be strengthened, never
+                weakened.
+  cadence       a plates-per-day target converted into the admission interval it implies,
+                and refused until it says whether it counts plates started, completed,
+                QC-passed, or released.
 """
 
 from .executor import Executor, Handoff, RunReport, StepResult
@@ -48,7 +62,39 @@
   loop_closure,
   trusted_for,
 )
+from .authority import (
+  Authority,
+  DecisionClass,
+  Plan,
+  QualificationRung,
+  Rung,
+  Violation,
+  qualified_for,
+  violations,
+)
+from .cadence import (
+  Cadence,
+  Demonstration,
+  Endpoint,
+  Target,
+  cadence_for,
+  demonstrated_support,
+  headroom,
+  launch_interval,
+  sustained_rate_demonstrations,
+)
 from .ledger import Ledger, StepVerdict, Unlock, build_ledger, cost_step, rank_unlocks
+from .literature import (
+  EVIDENCE,
+  Confidence,
+  Domain,
+  Evidence,
+  EvidenceKind,
+  Scope,
+  Support,
+  support_for,
+  unsupported_claims,
+)
 from .lineage import (
   MISASSIGNMENT,
   Cohort,
@@ -92,13 +138,23 @@ def loop_closure_for(protocol, workcell=None):
 
 __all__ = [
   "Artifact",
+  "Authority",
+  "Cadence",
   "Cohort",
+  "Confidence",
   "MISASSIGNMENT",
   "Attestation",
   "Basis",
   "Benchmark",
   "BenchmarkStatus",
   "Criterion",
+  "DecisionClass",
+  "Demonstration",
+  "Domain",
+  "EVIDENCE",
+  "Endpoint",
+  "Evidence",
+  "EvidenceKind",
   "CustodyGap",
   "Decision",
   "Detection",
@@ -122,44 +178,60 @@ def loop_closure_for(protocol, workcell=None):
   "Misassignment",
   "LoopClosure",
   "Observable",
+  "Plan",
   "Protocol",
+  "QualificationRung",
   "Readiness",
   "Role",
   "RunRecord",
+  "Rung",
+  "Scope",
   "Separability",
   "RunReport",
   "Severity",
   "Step",
   "StepResult",
+  "Support",
   "StepVerdict",
   "Tier",
+  "Target",
   "TimeBasis",
   "Traceability",
   "Transform",
   "UndeclaredTransform",
   "Unlock",
   "Verdict",
+  "Violation",
   "VisionCapability",
   "VisionRequirement",
   "VisualCheck",
   "Workcell",
   "ZeroDecodeOp",
   "build_ledger",
   "build_lineage",
+  "cadence_for",
   "cost_step",
   "declared",
+  "demonstrated_support",
   "estimate",
   "evaluate",
   "gate_report",
+  "headroom",
   "knowledge_summary",
+  "launch_interval",
   "lineage_report",
   "loop_closure",
   "loop_closure_for",
   "provenance_report",
+  "qualified_for",
   "rank_unlocks",
   "recovery_report",
   "registry",
   "spec",
+  "support_for",
+  "sustained_rate_demonstrations",
   "trusted_for",
   "undeclared_transforms",
+  "unsupported_claims",
+  "violations",
 ]