HCYHongye Chris Yang
ResourcesAccounting Evidence Engine
Accounting Evidence EnginePrototypeprototype-0.1.0

Reliability and Cost Console

Separate measured prototype evidence from unavailable production metrics, with explicit populations, periods, denominators, release versions, and error categories.

  1. AuthoritySourceVerbatim bounded excerpt or public identifier
  2. ComparatorDeterministic baselineVersioned rule IDs, terms, and exclusions
  3. InterpretationStructured model outputInactive in this prototype
  4. ControlReview and releaseStatus, errors, metrics, version
Metric register

Measured or explicitly not measured

Every metric carries a basis, population, evaluation period, denominator, and release. The console does not convert absent telemetry into a decorative zero or synthetic production claim.

Open public JSON snapshot

Showing 20 of 20 metrics.

evidence itemsmeasured

Evidence schema validity

5 / 5

The build rejects malformed or impossible dates, noncanonical companies, mismatched SEC URL CIK/accession/document fields, stale release versions, duplicate excerpts, invalid rule IDs, inconsistent accounting codes, unsupported model metadata, and review digests that do not match the current excerpt.

Population
All evidence items imported by the prototype build
Evaluation period
Build validation for prototype-0.1.0
Denominator
5
Release
prototype-0.1.0
ground-truth casesmeasured

Known-target candidate coverage

3 / 3

The known public-company target appeared anywhere in the bounded generated candidate list; final acceptance is evaluated separately by the benchmark-agreement metric.

Population
Benchmark cases with one known public-company target
Evaluation period
Public-safe benchmark frozen 2026-08-21
Denominator
3
Release
prototype-0.1.0
public-safe benchmark casesmeasured

Linkage benchmark agreement

5 / 5

A case passes only when both the expected decision class and accepted entity agree. This small demonstration benchmark is not a production-quality estimate.

Population
Five frozen demonstration cases spanning identifiers, aliases, subsidiary risk, and common-name collision
Evaluation period
Public-safe benchmark frozen 2026-08-21
Denominator
5
Release
prototype-0.1.0
evidence itemsmeasured

Source-check coverage

5 / 5

Each checked item retains the review date, reviewer role, method, protocol version, source anchor, and a SHA-256 digest bound to the current verbatim excerpt; this is not an assurance opinion.

Population
All evidence items in the prototype release
Evaluation period
Source check completed 2026-08-21
Denominator
5
Release
prototype-0.1.0
public-safe benchmark casesmeasured

Unresolved benchmark share

1 / 5

Unresolved is an intended safe outcome when a common name lacks decisive public identifiers.

Population
All five frozen demonstration cases
Evaluation period
Public-safe benchmark frozen 2026-08-21
Denominator
5
Release
prototype-0.1.0
evidence itemsmeasured

Verbatim-excerpt provenance coverage

5 / 5

Every displayed item includes a verbatim bounded filing excerpt, canonical company identity, CIK, accession number, filing date, form, source anchor, SEC archive link, and passage digest.

Population
All evidence items in the prototype release
Evaluation period
SEC filings dated 2024-07-30 through 2025-07-30
Denominator
5
Release
prototype-0.1.0
dynamic API requestsnot-measured

Dynamic API error rate

Not measured

The current JSON interfaces are pre-rendered release snapshots; a runtime API error rate would be fabricated.

Population
No dynamic product API is deployed in this static prototype
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
labeled passagesnot-measured

Extraction F1

Not measured

F1 requires the same frozen benchmark and fixed baseline/model releases as precision and recall.

Population
No frozen extraction benchmark exists for the prototype
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
labeled passagesnot-measured

Extraction precision

Not measured

Precision will not be displayed until a pre-sampled labeled benchmark and fixed extraction release exist.

Population
No frozen extraction benchmark exists for the prototype
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
labeled passagesnot-measured

Extraction recall

Not measured

Absence of extracted evidence cannot be interpreted as absence of disclosure.

Population
No frozen extraction benchmark exists for the prototype
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
adjudicated filing passagesnot-measured

False negatives

Not measured

A false negative count requires a pre-sampled filing population with relevant and irrelevant passages labeled independently of system output.

Population
No exhaustively labeled filing benchmark exists for the prototype
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
adjudicated evidence itemsnot-measured

False positives

Not measured

A false positive count requires source-level adjudication of every emitted evidence item under a fixed extraction release.

Population
No frozen extraction benchmark exists for the prototype
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
independently reviewed itemsnot-measured

Human-review disagreement

Not measured

Source-check status is recorded, but reviewer disagreement requires independent duplicate coding and a frozen adjudication protocol.

Population
No double-coded adjudication sample exists for the prototype
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
USD per filing and evidence itemnot-measured

Model cost

Not measured

Cost will be reconciled to actual input/output usage and the frozen provider price schedule when a model release exists.

Population
No model run is active
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
milliseconds per filingnot-measured

Model latency

Not measured

Latency will be reported from actual scheduled run metadata, not estimated from a generic provider specification.

Population
No model run is active
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
evaluated passagesnot-measured

Model-versus-rule disagreement

Not measured

The prototype deliberately exposes an inactive model layer rather than inventing agreement statistics.

Population
No model output is active
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
changed evidence and metric recordsnot-measured

Release-to-release changes

Not measured

No prior public release is invented as a comparison baseline.

Population
Only one public prototype release exists
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
ingestion runsnot-measured

Scheduled ingestion health

Not active

No success, failure, stale-data, or API-error rate is shown before a real backend exists.

Population
No scheduled Worker is provisioned for this prototype
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
in-scope filingsnot-measured

Stale filing count

Not measured

The interface reports its snapshot date, but cannot count stale filings without a scheduled ingestion contract and current expected-filing universe.

Population
No scheduled ingestion universe exists for the prototype
Evaluation period
Snapshot last source-checked 2026-08-21
Denominator
Not available
Release
prototype-0.1.0
model responsesnot-measured

Structured-model schema failures

Not measured

The evidence data schema is build-validated, but model-response schema failures cannot exist until a fixed structured-model release runs.

Population
No structured-model run is active
Evaluation period
Not applicable
Denominator
Not available
Release
prototype-0.1.0
Operational state

What is active in this release

The release has a source-checked static evidence snapshot and client-side linkage benchmark. Scheduled ingestion, model execution, D1/R2 persistence, and production APIs are not presented as active.

SEC snapshot

Static, source-checked sample

Five evidence items; last checked 2026-08-21.

Scheduled ingestion

Not active

No Cron Worker, D1 database, or public-product R2 bucket is claimed in this prototype.

Structured model

Not active

No model run is presented in this prototype. The interface exposes the planned structured-output fields while keeping them explicitly unavailable until a fixed model, prompt, benchmark, and run record exist.

Public data interface

Static release snapshots

Read-only JSON snapshots are generated from the same typed data used by the interface.

Stale-data interpretation: the prototype snapshot was last source-checked on 2026-08-21. Because no scheduled ingestion exists, the interface does not claim current filing coverage after that date.

Error taxonomy

Failure modes remain inspectable before metrics exist

Error categories define what the public-beta benchmark must measure. A defined category is not reported as an observed error count unless an evaluation population exists.

Irrelevant keyword contextDefined; not yet benchmarked

A term appears, but the passage does not describe the company action or accounting object of interest.

Negation or hypothetical languageDefined; not yet benchmarked

The passage discusses a possible or negated event rather than an observed disclosure object.

Table or header extraction failureDefined; not yet benchmarked

HTML structure is mistaken for a substantive passage or loses the intended reading order.

Passage truncationDefined; not yet benchmarked

The retained passage omits qualifying language required to interpret the evidence.

Duplicate evidenceDefined; not yet benchmarked

The same filing statement is emitted more than once through overlapping sections or rules.

Wrong topicDefined; not yet benchmarked

A valid passage is assigned to an unsupported topic family.

Wording change without meaning changeDefined; comparison review required

Text differs across filings, but the substantive disclosure object may not have changed.

Missing filing sectionDefined; not yet benchmarked

A relevant section is absent from the normalized document or prefilter output.

HTML parsing failureDefined; not yet benchmarked

The filing cannot be normalized into stable passages with a usable source locator.

Schema violationBuild-time evidence validation active; model validation not active

Required provenance or structured-output fields are missing, malformed, or internally inconsistent.

Unsupported summaryPrevented in prototype by inactive model layer

A generated interpretation includes a claim that is not supported by the exact source passage.

Entity ambiguityDemonstrated in two public-safe benchmark cases

The available fields cannot distinguish a parent from a subsidiary or one common-name entity from another.

Release control

One prototype release, no invented trend

Release-to-release change is unavailable because this is the first public prototype snapshot. The page does not manufacture a prior baseline for visual comparison.

prototype-0.1.0

Prototype

Released
2026-08-21
Data checked
2026-08-21
Evidence schema
evidence-item@1
Topic rules
topic-rules@1
Company universe
prototype-us-public-2026-08-21
Linkage benchmark
5 / 5 cases agree with the frozen expected outcome

Accepted limitations

  • The prototype contains a small source-checked filing sample rather than the planned 30-company public-beta universe.
  • There is no scheduled SEC ingestion, D1/R2 persistence, or public model execution in this release.
  • Precision, recall, F1, latency, and model cost are not reported because no frozen labeled benchmark or model run exists.
  • Displayed evidence is illustrative of the inspection workflow and is not an exhaustive review of any filing or company.
Scope boundary

Evidence inspection, not automated judgment

This prototype does not provide investment recommendations, assurance opinions, causal conclusions, unrestricted prompting, arbitrary URL retrieval, confidential uploads, or user accounts. Missing extracted evidence does not establish that a disclosure is absent.