ACT
Anchor-Calibrated Trust
Agreement can tell you who sides with whom.
It cannot tell you which side is right.
ZH://RESEARCH Trust orientation for multi-source systems
ACT · anchor-calibrated trust · unpublished figures
Sources9
Regime6 / 9 corrupted
Cohorts—
Orientation—
Chaptersources
A The item under review one decision
ILLUSTRATIVE MULTI-SOURCE EXAMPLE
“Your account is locked. Verify your card details at this link to restore access.”
Decision A—
Phishing — block it
Decision B—
Legitimate — deliver it
Majority—
Nine automated sources each make one binary call. We are not told which of them can be trusted.
B Source decisions —
Nine sources look at the same item. Each one makes a call.
SOURCE DECISIONS
C Trust orientation which side to trust
—awaiting
agreement structure
Agreement ρ
—
ACT ρ
—
ρ is the correlation between estimated and true source reliability. Negative means the ranking is upside down.
D What ACT produces reliability map
Source decisions
Agreement structure
Anchor orientation
Oriented reliability map
↓ Downstream aggregation
Any consumer of source reliability. RoT / DRWA / AVGA are used as evaluation contexts in the paper.
Anchors orient the partition. They do not train or update the underlying models.
EVIDENCE illustration→reported experiments corrupted-majority configuration · 6 of 9 sources corrupted
Trust-orientation correlation ρ against true source reliability
With six of nine sources corrupted, agreement ranks the sources upside down. ACT turns the ranking the right way up.
Agreementpairwise structure only
−1 INVERTEDALIGNED +1
−0.82
ACTagreement + 16 labeled anchors
−1 INVERTEDALIGNED +1
+0.82
Consistency across seeds orientation reversals
SASentiment · SST-2 · BERT25 / 25
TRLPreference · HH-RLHF · GPT-2-XL + LoRA20 / 20
—False reversals in the main trusted-majority configurationsnone
NOISERandom anchor label noise ≤ 30%≥ 97.5%
Reversals are counted per seed. The noise figure is random anchor corruption — it does not speak to adversarially chosen anchors.
The ambiguity is not specific to pairwise agreement five aggregation families
DAWID–SKENE↓ GLAD↓ SPECTRAL↓ TRUTH DISCOVERY↓ SNORKEL-FAMILY↓
↑ one shared anchor orientation ↑
SUBSTRATESentiment: SST-2 · BERT. Preference: HH-RLHF · GPT-2-XL + LoRA.
ANCHORS16 labeled items. They orient the partition; they do not retrain the sources.
SCOPEACT estimates trust orientation, not precise continuous source reliability.
Agreement
Who agrees with whom?
→
ACT
Which group should we trust?
Agreement finds the groups.
ACT finds the direction.
ACT estimates trust orientation, not precise continuous source reliability.
ZH://RESEARCH ANCHOR-CALIBRATED TRUST ORIENT WHEN NEEDED · PRESERVE WHEN ALREADY CORRECT
Paper · TBA
0:00 / 0:45