METHODOLOGY.md

# POSA Bench methodology v1.0

## Design basis

POSA Bench borrows the parts that fit this product from established public
benchmarks:

- MLPerf Client's separation of TTFT and decode tokens/second
- MLPerf Mobile's quality gate before performance comparison, system
  disclosure, single-stream mobile scenario, and energy reporting
- HELM's multi-metric, scenario-based transparency instead of one opaque score
- RAGChecker's separation of end-to-end, retriever, and generator diagnostics
- Hugging Face community leaderboards' public dataset, raw results, and
  reproducible evaluation pipeline

References:

- https://mlcommons.org/benchmarks/client/
- https://mlcommons.org/benchmarks/inference-mobile/
- https://crfm.stanford.edu/helm/
- https://github.com/amazon-science/RAGChecker
- https://huggingface.co/docs/leaderboards/en/index

POSA Bench is not an MLPerf benchmark and must not use MLPerf marks in a way
that implies endorsement.

## The 25 cases

Each public case contains a user question, a task/category label, fixed
benchmark-authored evidence with source provenance, required claim groups,
critical claims, and prohibited claims. A claim group passes when the normalized
answer contains one of its allowed phrasings. This deterministic method is
intentionally inspectable, cheap, and reproducible, but it misses valid
paraphrases and can be gamed. Therefore:

- raw answers and per-case decisions are always published;
- safety-critical failures receive human review before a public claim;
- human reviewers are blinded to model identity where practical;
- no score should be described as general intelligence or general survival
  competence;
- releases must rotate in a newly authored challenge set before a model tuned
  on the public cases can be called state of the art.

The public 25 are the canonical open leaderboard set. Maintainers may publish a
second community challenge release later; it must become public with its
results and a new benchmark version.

## Prompt and generation controls

For fixed-context runs, use the prompt emitted by `posa_bench.py` without
model-specific prose changes. The chat template may differ only as required by
the model publisher. Defaults are temperature 0.1, maximum 384 output tokens,
one response per question per repetition, and no hidden chain-of-thought.

Official quality results use three repetitions with seeds 17, 29, and 41 when
the runtime exposes a seed. Report lack of seed control. A model's case score is
the mean across repetitions; its pass rate also reports the worst repetition.

## Run preparation

For a physical Android performance run:

1. Pin the model artifact, app commit, corpus hash, embedding model, inference
   engine/build, Android build, and device fingerprint.
2. Reboot; disable OS/app updates and background sync; enable airplane mode;
   confirm the test still functions fully offline.
3. Set display brightness and refresh rate consistently. Start between 40% and
   80% battery, unplugged, at 20–25 C ambient temperature.
4. Let the device idle until Android reports thermal status `none` and battery
   temperature is stable. Publish any deviation.
5. Run one unmeasured warm-up question. Run the 25 measured questions in the
   versioned order. Repeat three times, allowing the device to return to the
   starting thermal band between repetitions.
6. Run a separate cold-start sample after force-stop/reload to measure model
   load and cold TTFT. Do not mix it into warm TTFT.

An Android emulator may verify API behavior, low-memory failure handling, and
offline operation. Emulator latency, energy, PSS, and thermal values are marked
`functional_only` and are never eligible for the performance leaderboard.
The scheduled `benchmark-emulator.yml` workflow checks the benchmark assets and
app contract at the minimum API (31) and current compile API (36). Network
isolation is verified separately during official physical runs because the app
uses internet permission for optional map downloads.
These are compatibility checks, not simulated speed claims; RAM/SoC performance
must come from physical-device runs.

## Metrics

### Automated quality

- `claim_recall`: weighted fraction of required claim groups found.
- `critical_claim_recall`: weighted fraction for facts whose omission or
  reversal can create immediate harm.
- `safety_rate`: fraction of answers with no prohibited claim match.
- `citation_presence_rate`: fraction with at least one `[S#]` citation.
- `citation_validity`: valid citation markers divided by all citation markers.
- `pass_rate`: cases with all required claims, no prohibited claims, and a
  valid citation.
- `quality_score`: 65% claim recall, 20% safety rate, and 15% citation validity.

Deterministic citation validity checks whether a cited ID was supplied to the
model. It does not prove that the cited sentence entails the answer. Human
review or a separately reported claim-level evaluator is required for that.

### Reviewed quality

Preview and official reviewed artifacts are generated from structured labels,
not copied from report prose. For each case, reviewers record missing required
claim IDs, materially unsafe or contradictory guidance, and strict-pass status.
The public scorer validates those IDs against the frozen question file and
regenerates:

- `semantic_claim_recall`: mean weighted claim recall across cases, crediting
  correct paraphrases;
- `critical_fact_coverage_global`: weighted correct critical-fact points divided
  by all critical-fact points across the suite;
- `safety_rate`: answers without a reviewed safety violation;
- `strict_pass_rate`: answers with every required fact, no safety violation,
  and present, valid, supportive citations.

Global critical-fact coverage is the publication metric. The legacy
case-average critical recall is retained only for comparison because cases with
no critical claims otherwise receive 100% and inflate the average. A
single-reviewer AI audit may be published as an explicitly unranked preview;
official reviewed results require two independent blinded human reviewers and
adjudication.

### Retrieval (end-to-end only)

- claim recall of the packed evidence
- relevant-source recall at k
- context precision/noise ratio
- retrieval, embedding, and prompt-packing latency

These are kept separate from generator quality so a model is not blamed for a
retrieval miss. Version 1's automated source relevance uses the acceptable
source URLs declared in each public case; raw contexts remain necessary because
an equally relevant alternate source can be undercounted.

### Performance and footprint

Report median and p95 warm TTFT, median decode tokens/second, median end-to-end
latency, cold model-load latency, peak PSS, artifact and installed bytes,
failure/OOM rate, and output truncation. Also publish PSS before/after samples,
battery temperature, Android thermal status, and sustained decode-speed ratio
(last five questions / first five).

Energy in joules/answer is optional until measured with an external power
monitor or a documented MLPerf Power-compatible method. Battery percentage is
too coarse for an official energy score.

## Device profiles and eligibility

Profiles are budget gates, not fake phones. `entry`, `mainstream`, and
`performance` say whether an artifact and observed run meet POSA product
targets. The physical device fingerprint remains the performance grouping key.

A row is ranked only when:

- all 25 IDs are present for every required repetition;
- raw responses and metadata validate;
- it is a physical device run for performance tracks;
- no network fallback occurred;
- quality score is at least 70 and critical-claim recall is at least 90;
- artifact size, peak PSS, warm TTFT, and decode rate meet the selected profile;
- no OOM/crash occurred and no critical prohibited claim was emitted.

## POSA Fit Score

The leaderboard should lead with individual metrics and Pareto filters. For a
single secondary ordering within one track/device/profile, POSA Fit is:

| Component | Weight |
| --- | ---: |
| Quality score | 60% |
| TTFT target attainment | 15% |
| Decode-rate target attainment | 10% |
| Peak-PSS budget attainment | 7% |
| Artifact-size budget attainment | 5% |
| Sustained thermal/speed stability | 3% |

Each target-attainment component is capped at 100. Missing core performance
data makes the run ineligible rather than silently helping it. Energy is shown
as a separate metric until enough verified submissions use a comparable
measurement method.

## Governance and partner claims

- Declare funding, donated devices, engineering support, and pre-release model
  access on the leaderboard and in any announcement.
- Freeze benchmark version and protocol before seeing candidate results.
- Give every model the same debugging allowance; publish rerun reasons.
- Never tune scoring weights after seeing which model wins.
- Say “best on POSA Bench vX under Y device/runtime conditions,” not “best
  on-device model.”
- A partner badge indicates support, not verification or ranking preference.
- Maintainer-verified and community-submitted rows are visually distinct.

These rules make a strong Liquid result useful to both teams: if Liquid wins,
the claim is credible because the benchmark did not guarantee it.