dsearch/documents
3
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs(devidence): preserve verification asymmetry research note

Browse Source
This commit is contained in:
Matthew Raymer
2026-10-08 08:37:40 +00:00
parent aa205d5c17
commit 4594791b0b
3 changed files with 286 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
research-notes/README.md
+7
View File
@@ -0,0 +1,7 @@
# DEVIDENCE Research Notes
Non-normative research notes preserve empirically important findings and open questions without changing Base Layer or conformance semantics.
<!-- BEGIN VERIFICATION-ASYMMETRY-R27A -->
- [Verification Asymmetry, Evidence Volume, and Independent Evidentiary Burden](verification-asymmetry-and-evidence-volume.md) — non-normative L4 research note grounded in AI-REG R27A.
<!-- END VERIFICATION-ASYMMETRY-R27A -->
research-notes/verification-asymmetry-and-evidence-volume.md
+235
View File
@@ -0,0 +1,235 @@
# DEVIDENCE Research Note: Verification Asymmetry, Evidence Volume, and Independent Evidentiary Burden
**Status:** Non-normative research note
**Promotion state:** L4 cross-project research promotion; **not** an L5 standards candidate
**DEVIDENCE effect:** No Base Layer change. No conformance-profile change. No new evidence type or authority rule.
**Evidence anchor:** AI-REG Verification Asymmetry R27A (`ebf55815a4a052b22db4aff0d4473a58f822bf0b78847e8c27ca32225746e45b`)
**Repository base when added:** `aa205d5c1799b41870e7affa9d0cd933b87ee232`
**Recorded UTC:** `2026-10-08T08:27:36+00:00`
## Why this note exists
AI can lower the marginal cost of producing plausible submissions, claims, reports, filings, requests, or other substantial text faster than institutions lower the cost of trustworthy verification, triage, and adjudication.
This creates a **verification asymmetry**:
> generation/submission capacity can scale faster than reliable verification capacity.
The problem does not require falsehood. A large volume of plausible or partially correct material can saturate scarce review capacity simply because each item imposes a non-trivial verification burden.
R27A found strong evidence of this pattern in:
- scholarly preprint and journal submission/review;
- FOI/FOIA and public-records processing;
- vulnerability and security reporting; and
- legal-filing accuracy and verification duties.
R27A deliberately did **not** claim that every candidate institutional surface is already experiencing AI-driven saturation. Regulatory comments and filings, grants, procurement, standards submissions, complaints, appeals, and compliance submissions remain monitoring targets unless and until evidence supports stronger claims.
## Core DEVIDENCE constraint
A formally conformant evidence packet is not necessarily strong evidence.
In particular:
- schema validity is not evidentiary sufficiency;
- acquisition success is not truth;
- provenance completeness is not truth;
- observation count is not independent-origin count;
- many citations are not necessarily many independent sources;
- many agents are not necessarily many independent evidentiary origins;
- repetition is not corroboration;
- rate limiting is not evidence quality;
- human review is not automatically reliable; and
- automated verification is not automatically adversarially robust.
DEVIDENCE therefore must remain capable of describing evidence and provenance without silently converting structural completeness into authority, truth, or priority.
## Verification asymmetry model
For a submission or claim set, let:
- `G` denote the marginal cost of generating or submitting material;
- `V` denote the cost of trustworthy verification, triage, or adjudication; and
- `N` denote the number of submitted claim units.
This is a conceptual model, not a universal numeric metric. In practice, verification cost is heterogeneous and depends on source availability, domain expertise, representation type, contradiction state, reproducibility, and institutional requirements.
The important condition is:
> production can scale much faster than trustworthy verification.
The architectural objective is therefore not merely to suppress volume. It is to shift as much verification work as possible into inspectable, reproducible, deterministic checks while leaving unresolved judgment explicit.
## Evidence-bearing submission candidate
A useful response is an **evidence-bearing submission** pipeline:
1. receive the submission;
2. preserve submission identity and provenance;
3. identify individual claim units;
4. bind claims to exact source identities and versions;
5. preserve claim-to-source relationships and quotation/observation coordinates;
6. validate source materialization, hashes, quotations, and coordinates where deterministically possible;
7. identify duplicate and derivative sources;
8. build source-dependency and independent-origin relationships;
9. surface contradictions and qualifications;
10. expose unresolved provenance and verification work;
11. apply an explicit, replaceable prioritization or queue policy; and
12. send unresolved judgment to human adjudication.
This pipeline must not become an opaque replacement gatekeeper. Queue policy, ranking, evidentiary weight, authority, and truth remain distinct.
## Independent evidentiary burden
R27A introduced the following as **derived analytical dimensions**, not as a DEVIDENCE standard or global evidence score:
- `CLAIM_COUNT` — claim units requiring support or adjudication;
- `OBSERVATION_COUNT` — raw observations or repetitions;
- `INDEPENDENT_ORIGIN_COUNT` — estimated distinct upstream evidentiary origins after dependency clustering;
- `DERIVATIVE_RATIO` — observations tracing to already represented origins;
- `SOURCE_DEPENDENCY_DEPTH` — depth of source-on-source reliance relevant to the claim;
- `CIRCULAR_DEPENDENCY_PRESENT` — whether support contains dependency cycles;
- `UNSUPPORTED_CLAIM_COUNT` — claims without an adequate supporting representation;
- `UNRESOLVED_CONTRADICTION_COUNT` — material contradictory or qualifying claims not adjudicated;
- `PROVENANCE_GAP_COUNT` — missing identity, version, route, or coordinate information needed for replay;
- `DETERMINISTIC_CHECK_FAILURES` — failed hash, quote, coordinate, schema, materialization, or related deterministic checks; and
- `HUMAN_ADJUDICATION_QUEUE` — explicit unresolved judgment after deterministic checks.
These dimensions may support replaceable workflow or queue policies. DEVIDENCE and DSEARCH should **not** canonize them into one hidden global quality, trust, or importance score.
## Evidence-volume attack model
DEVIDENCE must be red-teamed against evidence volume, not merely malformed packets.
R27A identified at least these attack/failure classes:
### Packet flooding
Large numbers of formally valid packets can exhaust moderation or verification capacity. Duplicate fingerprints, origin clustering, rate/account provenance, and explicit queue policy can reduce avoidable work, but high-value novel claims may still be hidden within the flood.
### Derivative-source laundering
Many publications can derive from one upstream observation while appearing to provide independent corroboration. Source-dependency graphs, canonical-source linking, syndication detection, and similarity analysis should expose common ancestry.
### Citation rings
Mutually citing sources can create apparent support without independent evidentiary origin. Dependency cycles and publication chronology should be inspectable. A cycle is not automatically invalid; a later source may still add genuinely independent evidence.
### Quote mining
A quotation can be exact yet materially misrepresent source scope or conclusion. Quote coordinates and surrounding context reduce, but do not eliminate, interpretive burden.
### Authority laundering
A secondary source must not silently inherit the authority of a primary source it discusses. Source authority class and source-vs-quotation origin must remain separate.
### Contradiction suppression
A packet may present only supporting evidence while excluding known contrary or qualifying material. Claim-neighborhood search and contradiction discovery are therefore part of verification, not optional decoration.
### Evidence-graph bombs
A dependency graph can be formally valid but intentionally expensive to traverse. Bounded traversal, origin collapse, lazy verification, and explicit verification budgets are legitimate operational tools as long as they do not masquerade as truth judgments.
### Synthetic-origin multiplication
Agents or accounts can repeatedly restate the same claim. Observation count must remain separate from independent evidentiary origin. Identity and Sybil resistance may remain incomplete and must be marked accordingly.
### Formally complete weak packets
Every required field can be present while the supplied sources are weak, nonresponsive, derivative, or inadequate for the proposition. Bounded representation adequacy and claim-to-source analysis remain necessary.
### Verification-cost attacks
A submission can select artifacts that are disproportionately difficult to acquire, render, reproduce, or inspect. Verification-route cost and materialization requirements may be recorded, but high verification cost alone does not make a claim unworthy.
## Relationship to current DEVIDENCE semantics
R27A did **not** demonstrate that the current DEVIDENCE Base Layer or acquisition/representation conformance profile loses required information.
The current research position is therefore:
> Treat source dependency, independent-origin analysis, contradiction discovery, and verification burden as derived analyses first. Promote new semantics only after repeated cases demonstrate material information loss that cannot be represented cleanly with existing provenance, derivation, relationship, adequacy, and uncertainty semantics.
This follows the empirical-accretion rule: real case → current vocabulary → material loss? → repeated cases → smallest useful distinction → cross-domain test → promote only if generalized.
## DSEARCH relationship
DSEARCH is a natural place to make evidence-volume structure searchable without converting search rank into evidentiary authority.
Candidate query/facet capabilities include:
- claims with many observations but few independent origins;
- circular source dependencies;
- duplicate/derivative source clusters;
- claims with missing source versions or provenance;
- citation/quotation validation failures;
- contradiction neighborhoods;
- claims whose deterministic checks pass but still require substantial human adjudication;
- source/version drift; and
- evidence-bearing submissions grouped by explicit verification-burden dimensions.
No mandatory global score should be introduced. Alternative ranking and prioritization policies must remain possible over the same preserved evidence.
## Cross-project implications
### Project Console
PC can govern the intake and adjudication process: deterministic validation receipts, queue-policy versions, adjudicator actions, promotion/rejection rationale, and the boundary between candidate evidence and accepted project state.
A DEVIDENCE-conformant packet must not automatically become accepted PC state.
### Devctl
Future Devctl validators can reduce verification cost by producing reproducible receipts for source materialization, hashes, quotations, citation coordinates, dependency fingerprints, duplicate/origin clusters, and executable reproducers.
A Devctl `PASS` is evidence that a bounded check succeeded; it is not semantic truth or legal merit.
### DSEARCH
DSEARCH can index source dependencies, independent-origin clusters, contradiction neighborhoods, duplicate/derivative relationships, and verification-burden facets.
DSEARCH ranking or priority is not evidentiary weight or truth.
### DEVIDENCE
DEVIDENCE should continue to preserve source identity, version identity, provenance, derivation, claim relationships, bounded adequacy, contradictions, qualifications, and explicit `UNKNOWN`.
The next DEVIDENCE question is empirical: whether repeated volume/independence cases expose a semantic distinction that current relationships cannot preserve without material loss.
## Boundary principles
The following boundaries are part of this research note:
- conformance does not self-authorize;
- provenance does not establish truth;
- quantity does not establish independence;
- ranking does not establish evidentiary weight;
- deterministic validation does not replace judgment;
- anti-spam policy must not silently become censorship or authority scoring;
- source or agent identity must not be treated as independent-person identity without evidence; and
- absence from public evidence must not be treated as absence from provider, participant, or regulator-compellable evidence.
## Promotion criterion
This topic is already important enough to remain permanently visible in DEVIDENCE research.
It should become an L5 standards candidate only if repeated, cross-domain cases establish that current DEVIDENCE semantics materially fail to represent one or more necessary distinctions, such as source dependence, independent evidentiary origin, or review/verification provenance.
Until then, the correct state is:
**preserve → dogfood → red-team → measure semantic loss → promote only if necessary.**
## Evidence provenance
This note is grounded in the R27A execution-evidence handoff:
- archive SHA-256: `ebf55815a4a052b22db4aff0d4473a58f822bf0b78847e8c27ca32225746e45b`
- R27A result: `VERIFICATION_ASYMMETRY_IS_A_FIRST_CLASS_AI_CAPABILITY_INSTITUTIONAL_CONSEQUENCE`
- R27A promotion state: `L4_CROSS_PROJECT_PROMOTED`
- R27A DEVIDENCE adjudication: current profile sufficient for source states; independent evidentiary burden remains a derived dogfood candidate
- R27A standards mutation: `NONE`
- R27A Base Layer change: `NONE`
The R27A evidence set reported strong support for scholarly publishing/review, FOI/public-records processing, vulnerability/security reporting, and legal-filing accuracy/verification. Other institutional surfaces remain monitored hypotheses unless later evidence promotes them.
research-notes/verification-asymmetry-and-evidence-volume.r27a.json
+44
View File
@@ -0,0 +1,44 @@
{
"anti_score_rule": "Do not canonize these dimensions into one global evidentiary quality, trust, importance, or priority score.",
"base_layer_change": "NONE",
"conformance_profile_change": "NONE",
"derived_dimensions": [
"CLAIM_COUNT",
"OBSERVATION_COUNT",
"INDEPENDENT_ORIGIN_COUNT",
"DERIVATIVE_RATIO",
"SOURCE_DEPENDENCY_DEPTH",
"CIRCULAR_DEPENDENCY_PRESENT",
"UNSUPPORTED_CLAIM_COUNT",
"UNRESOLVED_CONTRADICTION_COUNT",
"PROVENANCE_GAP_COUNT",
"DETERMINISTIC_CHECK_FAILURES",
"HUMAN_ADJUDICATION_QUEUE"
],
"monitoring_hypotheses": [
"REGULATORY_COMMENTS_AND_FILINGS",
"GRANTS_PROCUREMENT_STANDARDS_COMPLAINTS"
],
"promotion_state": "L4_CROSS_PROJECT_PROMOTED",
"recorded_utc": "2026-10-08T08:27:36+00:00",
"repository_base_commit": "aa205d5c1799b41870e7affa9d0cd933b87ee232",
"schema": "devidence.research-note.verification-asymmetry.r1",
"source_handoff": {
"adjudication_schema": "devidence.ai-reg.coupled-r27a-adjudication.v1",
"archive_sha256": "ebf55815a4a052b22db4aff0d4473a58f822bf0b78847e8c27ca32225746e45b",
"devidence_result": "CURRENT_PROFILE_SUFFICIENT_FOR_R27A_SOURCE_STATES; DEPENDENCY_BURDEN_REMAINS_DERIVED_DOGFOOD_CANDIDATE",
"result": "VERIFICATION_ASYMMETRY_IS_A_FIRST_CLASS_AI_CAPABILITY_INSTITUTIONAL_CONSEQUENCE",
"source_acquisition_counts": {
"FAIL": 2,
"PASS": 11
}
},
"standards_candidate": false,
"status": "NON_NORMATIVE_RESEARCH_NOTE",
"supported_surfaces": [
"SCHOLARLY_PREPRINTS_AND_JOURNALS",
"PUBLIC_RECORDS_FOI_FOIA",
"VULNERABILITY_AND_SECURITY_REPORTING",
"LEGAL_FILINGS_ACCURACY_AND_VERIFICATION"
]
}