Claims and evidence ledger

Current frozen system · HydroCore-v5 · v0.2.1 · source snapshot 4bbf6fa3ff9f

This ledger gives reviewers a compact mapping from a claim to its evidence and the wording boundary that keeps the claim scientifically honest.

Claim	Evidence	Safe wording	Do not claim
V5 is the final model	M11.2 freeze, runtime manifest	“HydroCore-v5 M10 frozen release is the final serving identity”	that V4 is still current
Exact model identity is frozen	same + M10.5 completion	quote seed/hash/parameter count exactly	approximate/alternate checkpoint identity
Only Sentinel is valid trained task	finalist freeze / runtime manifest	“final trained task family is Sentinel”	“Scout/Strategist/OOD are learned authorities”
Five learned outputs are runtime-enabled	finalist freeze / V5 loader	list the exact five	include next_step
OOD/Scout/planning remain deterministic	finalist freeze, safety counters	name OODDetector, rank_sample_locations, generate_response_plans	imply neural heads select authoritative samples/plans
V5 does not silently fall back to V4	V5 loader, safety counters	“invalid V5 assets fail closed without V4 fallback”	that every degraded state remains full hybrid
Exact simulation is required for verification	architecture/runtime + safety counters	“configured WNTR/EPANET verification is required for VERIFIED”	“WNTR proves real-world safety”
Human approval is required	finalist freeze / safety counters	“separate human approval event required”	“approval executes infrastructure”
M11.6 final lock passed	gate, closure	“locked-final and locked-topology gates passed”	“all predictive metrics passed thresholds”
Final population was complete	gate / post-run governance	“125/125 complete: 105 final + 20 topology”	extrapolate beyond tested population
Lock was opened once	opened record, governance	“one authorized opening; no rerun/post-lock tuning”	claim the lock remains unopened
All hard safety counters were zero	safety counters	“0 violations across all 15 frozen counters in 125 locked incidents”	“guaranteed safe”
Aggregate applicable final coverage passed	metrics, gate	“88.6% aggregate locked-final coverage; frozen floor 85%”	imply every condition slice passed 85%
Nominal predictive result	metrics	“73.3% Top-1 on nominal locked-final, n=15”	present 73.3% as all-condition final accuracy
Aggregate stress predictive result	metrics	“55.2% Top-1 across all locked-final, n=105”	hide stress degradation
Novel-topology signal exists	metrics	“55% Top-1 / 70% Top-3 descriptive on n=20 novel-topology cases”	call it calibrated topology-transfer performance
Novel topology failed closed	gate / metrics	“0% calibrated/actionable/approved; fail-closed gate passed”	imply 60% raw set inclusion is calibrated coverage
Data are synthetic	dataset artifacts	“all model/evaluation evidence is WNTR/EPANET-generated synthetic evidence”	field-validation claim
Larger M model was not promoted	M9 final summary	“larger M did not meet predeclared meaningful capacity-gain threshold”	“small is universally superior”
M10.4 development trajectories passed	M10.4 closure	“full development trajectory gate passed”	treat M10.4 as final held-out evidence
Reference incident demonstrates workflow	Reference demo	“checksummed WNTR-backed replay of a frozen workflow”	use it as final V5 benchmark evidence
Live Example demonstrates current v0.2.1 governed runtime behavior	Reference demo, current runtime	“Current frozen runtime on bundled scenario inputs may abstain rather than reproduce the Reference replay”	“Reference Incident is a live V5 run” or “Live Example is benchmark evidence”
Current source app serves V5	source runtime / finalist freeze	“current source app defaults to V5; the published v0.2.1 release-compose image also serves V5”	say a pre-v0.2.0 release-compose tag is V5

Numeric claim rules

When quoting a result:

1. name the population;
2. include n;
3. state calibration applicability when reporting coverage;
4. state whether the metric was gating or descriptive when relevant;
5. avoid converting marginal conformal coverage into per-incident probability;
6. avoid comparing populations as if they were matched if their generation/protocol differs.

Safety claim rules

Preferred:

“All 15 preregistered hard authority/safety counters were zero across the complete 125-incident locked evaluation.”

Not supported:

“HydroSwarm is proven safe for water utilities.”

Preferred:

“Novel-topology incidents retained descriptive localization signal while the runtime withheld calibrated/actionable authority.”

Not supported:

“HydroSwarm generalizes safely to unseen utility networks.”

Historical claim rules

Historical V3/V4/M9/M10 artifacts can be cited when discussing research evolution. Their metrics, model identities, and “locked unopened” statements must be labelled with their historical generation/milestone. They do not override the current M11.6 terminal state.

Source hierarchy

For current facts, prefer in this order:

1. frozen/generated identity/governance artifacts;
2. generated final metrics/gate/safety artifacts;
3. runtime code for actual serving/authority wiring;
4. this current documentation;
5. historical prose only for historical claims.

This prevents documentation drift from becoming scientific truth.