an uncalibrated power model is an estimate, not ModelBased #220

Merged

dcharlot merged 8 commits from fix/uncalibrated-model-is-estimate into main 2026-10-11 22:20:18 -04:00

Conversation 0 Commits 8 Files changed 35 +750 -318

2026-10-11

Transaction Science

ci(wai-web-wasm): the native unit-test count is 32 with the OTT grade test ...

The wai-web native step pins its test count so a test that stops being
compiled fails instead of passing silently. This branch adds
video_ott::tests::a_page_supplied_figure_is_sealed_as_an_estimate, so the
count is 32; the gate still read 31 and failed the job with every test
green. Name the new test's group in the comment beside the others.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

743641234a

• View at this point in history

2026-10-11 17:01:29 -04:00

Transaction Science

sandbox-exec-microvm: the liveness tick advances the fake counters itself ...

The RAPL and MSR meter tests drove their fake counters from a writer
thread on a 2 ms timer while the liveness probe slept between samples. On
a loaded host (load ~85 here) the writer was starved for whole sample
intervals, and the probe then judged a live fake counter frozen: the two
tests failed in 1-3 of every 5 runs on this branch and occasionally on
main. Each tick now writes the next rising step and then sleeps, so every
interval the probe samples holds exactly one advance whatever the
scheduler does. 16 of 16 runs passed, 6 of them with 40 extra busy
threads.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

93c6b2917c

• View at this point in history

2026-10-11 10:52:05 -04:00

Transaction Science

jouleclaw-fabric: seed the CPU and AMX matmul figures Estimated, not Benchmarked ...

local/cpu-neon (130 mJ/GFLOP) and local/amx (7 mJ/GFLOP) were seeded
Benchmarked, though each is a recorded throughput priced at an ASSUMED
~5 W core draw: the energy was never read. l3-jouleclaw's
projection_grade rightly treats Benchmarked as a calibrated input, so
with JOULECLAW_FABRIC=1 and no calibration row this host recorded, a
projection priced from these seeds was sealed ModelBased: an
uncalibrated constant claiming a model's grade (JCP SPEC §5.1 rung 6).

The seeds are now Estimated, as the GPU seeds already were, so a seeded
projection is an Estimator figure and a host calibration row still
replaces the seed at the grade its writer declared. EnergyProvenance's
docs say an assumed power makes a figure Estimated.

New tests: jouleclaw-fabric every_seeded_figure_is_an_estimate, and
l3-jouleclaw tests/fabric_grade_seeded.rs (fabric on, no calibration
file: the cpu-neon entry is Estimated and the receipt Estimator). The
existing calibrated/estimated fabric-grade tests are unchanged and pass.
No vector carries a fabric seed's grade.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

b22d3cd496

• View at this point in history

2026-10-11 10:11:54 -04:00

Transaction Science

wai-web: seal the OTT demo's page-supplied joules as Estimator, not ModelBased ...

OttSession::new seals whatever non-zero joules_micro the page hands it
as ModelBased. Nothing in the session measured, modelled or calibrated
that figure, so labelling it a model claims better provenance than how
it was obtained (the family's ladder, JCP SPEC §5.1: a model earns
ModelBased by its calibration). It is now sealed Estimator; zero stays
the unmetered marker with no class.

New test a_page_supplied_figure_is_sealed_as_an_estimate. No vector
changes: the OTT session is a demo surface, nothing under a conformance
or vectors directory is produced by it, and reach/record keeps the
package of record as it was.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

66ed977282

• View at this point in history

2026-10-11 10:11:54 -04:00

Transaction Science

sandbox-exec-microvm, sandbox-exec-wasm, sandbox-meter: grade nominal-cost figures Estimator ...

Three reference estimators priced work by a coefficient no host
calibrated and graded the figure ModelBased:
- sandbox-exec-microvm: with no live CPU counter, vCPU-seconds × the
  reference 12,000 µJ/vCPU-second (a nominal work-unit cost), and the
  accelerator's "coarse proxy" (that cost × 10);
- sandbox-exec-wasm: fuel × the reference 20,000 µJ per million fuel;
- sandbox-meter StorageActivity (from_samples, from_volume): ops and
  bytes × nominal NVMe-class pJ coefficients documented as "calibrated".
By the family's ladder (JCP SPEC §5.1) a model earns ModelBased by a
calibration on the host; a nominal coefficient is the whole of such a
model, so the figure is an estimate.

- MicroVmConfig.calibrated and WasmConfig.fuel_cost_calibrated (both
  false by default) keep a ModelBased path for a cost the operator
  calibrated on the host, as spec/receipts.md's "calibrated table"
  describes; the accelerator proxy is Estimator either way. Neither config
  is constructed outside its crate.
- Tests that pinned the over-claim now pin Estimator:
  cpu_only_run_meters_modelbased (renamed ..._an_estimate),
  a_cpu_counter_that_held_across_a_run_is_not_a_measured_zero, power.rs's
  two RAPL fallback tests, runs_a_unit_and_meters_cpu_energy,
  storage_globs_enforced_through_execution, contention_is_loaded_minus_
  baseline. New tests: a fallback is a model only at a host-calibrated
  cost (CPU) and never for the proxy (accelerator); the same for the fuel
  cost; both storage constructors are estimates.
- boot_microvm example's 15 W/vCPU fallback is Estimator too.
- SPEC §6.3, spec/receipts.md, SECURITY.md, README and sandbox-meter's
  Provenance docs say a nominal or assumed coefficient is Estimator.
No vector changes: sandbox-conformance passes every class, file-backed
vectors included, and conformance/v1 is untouched (its builders use
fixed constructors, none of these estimators).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

490a592223

• View at this point in history

2026-10-11 10:11:54 -04:00

Transaction Science

aac-core: grade the store's elapsed-time receipts Estimator, not ModelBased ...

timed_receipt prices every store mutation (register_user, create_board,
add_cell, record_utterance, predict, set_vocabulary_level) at elapsed
wall time × an assumed draw no device calibrated, and tagged it
ModelBased while its own doc called it "an estimate". By the family's
metering ladder (JCP SPEC §5.1) a model earns ModelBased by its
calibration; an assumed draw scaled by time is rung 6, Estimator.

The figure is unchanged. Its doc said "~2 mW", but 2.0 × elapsed µs is
2 µJ per µs, i.e. 2 W: the doc now says what the code computes.

access-core's Provenance / Receipt docs now say ModelBased is a draw the
device measured and an assumed draw is an estimate. New test
store_receipts_are_estimates. No vector changes (aac gen-vectors output
is byte-identical; the vectors carry no provenance).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

66137550bc

• View at this point in history

2026-10-11 10:11:54 -04:00

Transaction Science

proof-solvers: grade wall time × the default 15 W Estimator, not ModelBased ...

ns_to_uj prices an attempt at wall time × a fixed ~15 W that no host
calibrated, and Propositional, Z3Smt, LeanVerifier and every Z3Server
attempt that is not sealed as a tree reading graded that figure
ModelBased. By the JCP ladder (SPEC §5.1) and the proof SPEC's own
definition a model earns ModelBased by its calibration; an assumed power
scaled by time is an estimate, so the figure claimed better provenance
than how it was obtained.

- proof-schema: Energy::estimator(uj), and Provenance docs that say a
  default or assumed parameter makes a figure Estimator.
- proof-solvers: DEFAULT_ACTIVE_MILLIWATTS (15 W, same arithmetic) and
  wall_time_estimate; all four sites now Energy::estimator. The public
  Modelled / modelled_uj names are kept; their docs say estimate.
- Tests that pinned the over-claim now pin Estimator: grade_attempt's
  unit tests, z3_tree.rs (server(), refuted goal, settle bound, cascade
  total [Estimator, HwShunt] -> Estimator, undeclared z3), and proof-cli
  tests/measure.rs (receipt "provenance Estimator", "stay Estimator").
  New test: every wall-time rung (T0 proved/refused, one-shot z3, an
  absent Lean) is Estimator with no measurement.
- proof-cli, proof-cascade, proof SPEC §5 / README, and joule-code-remediate
  (FormalOutcome docs, joule_fix's printed label) say Estimator.
  proof/conformance/v1 is unchanged: gen-vectors output is byte-identical
  (it seals the stand-in ladder's estimate_tier figures, never ns_to_uj).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

e42c3d1c07

• View at this point in history

2026-10-11 10:11:54 -04:00

Transaction Science

jcp-energy: grade time × the default power Estimator, not ModelBased ...

SPEC §5.1 grades rung 5 as wall time × a HOST-CALIBRATED active power
(ModelBased) and rung 6 as an estimate. jcp-energy's ladder fell to
Meter::TimePower at DEFAULT_ACTIVE_MILLIWATTS (15 W, an assumed figure no
host measured) and every counter rung's span degraded to the same 15 W
figure, all graded ModelBased: better provenance than how the figure was
obtained. The power is the whole of that model, so a default power is a
constant scaled by time, an estimate.

- Meter::UncalibratedTimePower { milliwatts } and Rung::UncalibratedTimePower
  (Estimator), built by Meter::uncalibrated_time_power; detect() falls to
  it, so Detection.selected, rung() and provenance() all say estimate.
- Meter::TimePower / Meter::time_power stay ModelBased, documented as a
  power calibrated on this host.
- counter_span / ioreport_span / process_span degrade to the new
  uncalibrated_span (Estimator) instead of Energy::modeled.
- Tests that pinned the over-claim now pin Estimator: the RAPL/NVML/
  per-process/IOReport fallback-span tests, the declined-IOReport meter
  test, and the detect_with selection tests. New tests: a host-calibrated
  power stays ModelBased while the default is Estimator, detect() falls to
  the estimate, and the fallback span helper.
- jcp-conformance C9 now checks both grades.
- SPEC §5 / §5.1 rungs 5-6, jcp-schema's Provenance docs and the README
  say a default or assumed power is rung 6. No vector changes (jcp
  gen-vectors output is byte-identical; no jcp vector carries ModelBased).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

48db538a47

• View at this point in history

2026-10-11 10:11:53 -04:00