Parallax · Experiment 1 · public research release

Evaluating strategic AI responses when there is no single correct policy answer.

Experiment 1 asked whether advisory responses selectively changed when decision-relevant evidence changed, while remaining stable under matched task-preserving nuisance changes.

NOT SUPPORTED

Under the prospectively frozen two-provider criterion. This is the public static release of the accepted research artifact.

00 — Abstract

Parallax asks whether observable advisory-response behavior can be evaluated when there is no correct-policy oracle. Experiment 1 used three synthetic strategic scenarios and two provider-diverse tested configurations, with 210 confirmatory calls split evenly between Anthropic and Google. A prospectively frozen, pre-specified decision rule required target movement under decision-relevant evidence while both nuisance vectors matched baseline. The overall result was NOT_SUPPORTED: 154 calls produced structured observations and 53 were behavioral refusals; the remaining calls were schema or technical failures. Provider heterogeneity was strong. Google showed a descriptive target-specific pattern among 23 complete matched units, while Anthropic had refusal-heavy, limited complete-unit observability and only five complete units. These provider-local descriptions are not a second endpoint and do not infer why refusals occurred. The result makes no policy-correctness or strategic-competence claim, and it does not establish performance beyond these exact synthetic cases.

Open the abstract as a standalone text artifact.

01 — The question

Can observable advisory-response behavior be evaluated without an oracle that scores the best policy?

Parallax tests one necessary property: whether observable response categories move selectively under a controlled evidentiary update relative to ordinary repetition and task-preserving nuisance variation. The protocol does not score recommendations against a policy answer.

02 — The design

Diagram showing baseline, evidentiary update, two nuisance conditions, and the separate hard-constraint control
Figure 1. Experimental logic. XC is a separate measurement-instrument control, not part of the primary selective-response contrast.

Conditions

X0 baseline · XE decision-relevant evidence · XN-P paraphrase nuisance · XN-I irrelevant-information nuisance · XC hard-constraint control.

Allocation

210 historical calls: 105 Anthropic — claude-fable-5 and 105 Google — gemini-3.6-flash. Three synthetic scenarios were used.

Instrument

Each structured response declared one category: LEANS_H, UNRESOLVED, or LEANS_NOT_H.

03 — Confirmatory result

EXPERIMENT 1: NOT SUPPORTED

Over the same 28 complete matched units, XE differed from X0, but XN-P also differed from X0. Because the exact rule required both nuisance vectors to equal baseline, the selected rule was OTHERWISE=>NOT_SUPPORTED.

Primary categorical vectors for X0, XE, XN-P, and XN-I over 28 complete matched units
Figure 3. Primary categorical vectors. Counts are shown in the order LEANS_H, UNRESOLVED, LEANS_NOT_H.
Primary vectors over 28 complete matched units
ConditionLEANS_HUNRESOLVEDLEANS_NOT_H
X00280
XE3250
XN-I0280
XN-P0271

Complete-case disclosure: 48 provider × scenario × repeat units were planned; 28 were complete and 20 were incomplete. Anthropic contributed 5 complete units and Google 23. No missing assessment was imputed.

04 — Provider heterogeneity

The combined result is not a claim that all tested response behavior looked the same.

Google contributed 23 complete matched units and showed a descriptive target-specific pattern. Anthropic contributed only 5 complete units, all from S3, and those units did not show target movement. These provider-local findings are descriptive and do not create a second confirmatory endpoint.

Separate provider-local categorical vectors for Google n=23 and Anthropic n=5 complete units
Figure 4. Provider-local descriptive vectors. The different denominators are part of the interpretation.
Provider-local descriptive vectors
ProviderComplete unitsX0XEXN-PXN-I
Google23[0, 23, 0][3, 20, 0][0, 23, 0][0, 23, 0]
Anthropic5[0, 5, 0][0, 5, 0][0, 4, 1][0, 5, 0]
DESCRIPTIVE PROVIDER-LOCAL SELECTIVE PATTERN. Google’s complete-unit pattern is reported as a bounded description, not as a pass, proof, or general conclusion.

05 — Refusal and observability surface

DESCRIPTIVE OBSERVABILITY / MISSINGNESS SURFACE. This is an observed missingness description, not a confirmatory endpoint.

Structured observability was provider-, scenario-, and condition-dependent. A behavioral refusal remains an observed outcome; this artifact does not infer its cause.

Outcome observability bars by provider showing structured responses, refusals, schema failures, and technical failures
Figure 2. Outcome observability by exact tested configuration. Every bar retains its n=105 denominator.
Heatmap of structured, refusal, schema, and technical outcomes by provider, scenario, and condition
Figure 6. Descriptive refusal/observability surface. Each cell exposes its planned N.
Provider × scenario × condition observability
ProviderScenarioConditionNStructuredRefusalSchemaTechnical
AnthropicS1X086200
AnthropicS1XC33000
AnthropicS1XE88000
AnthropicS1XN-I80800
AnthropicS1XN-P81700
AnthropicS2X080800
AnthropicS2XC30300
AnthropicS2XE80800
AnthropicS2XN-I80800
AnthropicS2XN-P80800
AnthropicS3X088000
AnthropicS3XC33000
AnthropicS3XE86110
AnthropicS3XN-I87010
AnthropicS3XN-P88000
GoogleS1X088000
GoogleS1XC33000
GoogleS1XE87001
GoogleS1XN-I88000
GoogleS1XN-P88000
GoogleS2X088000
GoogleS2XC33000
GoogleS2XE88000
GoogleS2XN-I88000
GoogleS2XN-P88000
GoogleS3X088000
GoogleS3XC33000
GoogleS3XE88000
GoogleS3XN-I88000
GoogleS3XN-P88000

06 — XC control

18 planned · 15 structured · 3 behavioral refusals · 15/15 structured compliant.

Interpretation: CONSTRAINT_COMPLIANCE / MEASUREMENT_INSTRUMENT_CONTROL. The control shows that the instrument registered the explicit hard constraint when a structured response was available. It does not establish broad capability.

Complete matched unit coverage by provider and scenario
Figure 5. Complete matched-unit coverage. Google contributed 7, 8, and 8 units across S1–S3; Anthropic contributed 0, 0, and 5.

07 — What the protocol did and did not demonstrate

Integrity of the negative result

  • The endpoint was sealed before confirmatory execution.
  • Missing responses were not imputed.
  • Refusals were not retried into compliance.
  • Pilot and diagnostic observations were not pooled.
  • No provider was dropped after the result.
  • The nuisance movement remained in the decision rule.

What remains bounded

Experiment 1 demonstrates one implemented protocol instance that can freeze interventions, preserve refusals and failures, and produce a deterministic negative result without a policy-answer oracle. It does not establish recommendation quality, hidden reasoning, broad capability, or performance beyond these exact synthetic cases.

Limitations

08 — Data and reproducibility

This public release contains deterministic summaries and figures, not the raw provider-response corpus. The governed evidence is preserved internally by identity.

Local rebuild

From the repository root, run:

python scripts/build_run56_public_artifact.py

The build is zero-network and requires no credentials or runtime provider access.

Provenance

Run 55: fbe75991a4e56305fecdfe99a87ad040bfe37d7f6603e889e84bd2b3bad4d58c
Evidence: 4ac8293e829dc50e7c56e5a48e4984b18f80c185f7325feeeead5460f97f17a9
Package: d5b51b6324b3f2e3b2e6bbadb1c0b49f4a8165f5849161dd91d3317497f19e5f
Freeze: fc603c5dbc8d03727c87e690dad745b4beb5c78d81968d3c6915883178f95b16

Read the methods note. The next human boundary is HUMAN_ADJUDICATION_OF_PARALLAX_PUBLIC_ARTIFACT.

09 — Follow-up questions

Future work is bounded rather than launched here: can refusal/observability become a prospectively specified endpoint; does the Google pattern replicate on fresh scenarios; does Anthropic’s scenario × condition refusal geometry replicate; can future designs reduce complete-case attrition while preserving refusal behavior; and do other providers show similar response geometry?