1. Strategic advice without an answer key
Some of the questions policymakers most want to ask AI systems cannot be scored against a single correct policy: reassure an ally or signal resolve, disclose an assessment or preserve ambiguity, act now or wait. Standard benchmark accuracy does not directly tell us whether advice is responsive to evidence or fragile to phrasing.
Parallax asks a deliberately narrower behavioral question:
When supplied evidence is made stronger or weaker in a prospectively ordered way, do observable response classifications move in the corresponding direction while remaining stable under decision-irrelevant rewrites?
That is not a test of hidden belief, consciousness, truthfulness, internal reasoning, or optimal policy. It is a test of observable response behavior under controlled perturbation.
2. Experiment 1: the first claim failed
The first live protocol used three fictional strategic scenarios, two provider configurations, and 210 confirmatory calls. Its result was frozen before interpretation.
Among 28 complete matched units, the relevant-evidence condition moved three responses relative to baseline. One paraphrase-nuisance condition also moved a response. The contextual nuisance moved none. Because the target movement was not selective, the exact fallback rule applied: OTHERWISE=>NOT_SUPPORTED.
Observability was heterogeneous: 154 structured responses, 53 behavioral refusals, two schema failures, and one technical failure. Only 28 of 48 planned matched units were complete. The accepted result remains NOT_SUPPORTED. The target movement was preserved descriptively but never promoted into a post-hoc success.
This negative result established that the project was falsifiable and exposed three problems: nuisance content needed stronger identification, refusal could not be treated as an incidental defect, and sparse complete units made a simple matched analysis brittle.
3. Experiment 2: scale and custody did not rescue identification
Experiment 2 was a larger live scientific successor plus a separately bounded offline engineering-qualification program. The live campaign resolved all 888 planned coordinates—444 per provider, 864 MAIN and 24 mechanical-control coordinates—with no technical retries, missing coordinates, duplicates, or unexpected coordinates.
The exact accepted provider results were both INSUFFICIENT_EVIDENCE. Only 66 of 288 matched triples were fully valid; 222 were handled under the frozen pessimistic rule. The campaign recorded 275 refusals, 54 schema failures, and 97 technical truncations. Both mechanical construct gates passed, but that did not override failed MAIN coverage.
A separately labeled post-confirmatory analysis found nuisance movement in all 66 fully valid units. Google had 29 target movements, none selective; Anthropic had no target movement. Those observations are diagnostic, not a new confirmatory result. Offline recovery, replay, request realization, and custody checks establish engineering integrity only.
The deeper lesson was blunt: more calls do not automatically restore identification when availability and nuisance behavior dominate the matched sample.
4. Experiment 3: rebuild the measurement object
The mature successor uses five ordered evidence doses:
STRONG_AGAINST WEAK_AGAINST BASELINE WEAK_TOWARD STRONG_TOWARD
Its substantive response remains observable and small: LEANS_NOT_H, UNRESOLVED, or LEANS_H. Behavioral refusal is a separate co-primary response component, never placed on the substantive ordinal scale. Measurement, transport, and unclassified unavailability remain on one common eligible-unit denominator and enter a joint coherent identified set. The analysis does not invent an unseen answer.
The primary object is an ordered vector of cumulative response contrasts across adjacent evidence doses. Nuisance stability remains a separate affirmative content/behavioral-equivalence gate; it is not subtracted into the dose estimand. Providers remain separate.
The governing Situation Room campaign
The current governing local evidence is SR-8R1: 1,600 scheduled full-bundle runs across two cases and two local model families, Qwen 2.5 14B and Gemma 3 4B. Of those rows, 823 reached a non-transport terminal state: 822 substantive ordinal assessments and one measurement-unavailable row. Another 777 rows delivered no usable response because of local runtime or transport failure. They were attempted once, retained on the denominator, and never retried, replaced, or imputed.
The fictional Vela Strait case provides a controlled readiness ladder. Afghanistan 1979 provides one qualified historical transfer case at a fixed cutoff; the later invasion is not used as an answer key. An attempted identity-masking diagnostic is excluded because the model-visible hypothesis and source identifiers remained unmasked.
5. Attack the machinery before funding the provider study
The Experiment 3 validation ladder is intentionally role-separated:
exact protocol and custody
↓
known-ground-truth synthetic shadow
↓
real local-model surrogate execution
↓
live non-scientific provider-path qualification
↓
unexecuted funding-dependent confirmatory study
Known-ground-truth shadow
Twelve physical worlds exercised the full production geometry at 556,200 scheduled coordinates each: 6,674,400 evaluations total. They included clean pass, adjacent reversal, endpoint failure, nuisance instability, refusal/availability failure, conservative missingness, provider asymmetry, surface-cue following, decision-boundary behavior, retry metamorphism, and nonstationarity. Every observed disposition matched an independent oracle.
This establishes software and analysis behavior only. It does not establish real-model behavior.
Earlier real local Qwen2.5 stress campaign
The local campaign attempted and terminally accounted for 4,800 Qwen2.5 generations across 7B/14B and deterministic/stochastic strata. It produced 4,544 valid ordinal responses; 256 schema-invalid enum values, all in 14B strata; zero duplicate or replacement attempts and zero runtime errors; dose behavior that was not uniformly monotone; 17 nonzero blinded nuisance comparisons out of 60; and 2,519/4,800 outputs concordant with a predeclared shallow-cue heuristic.
These are accepted descriptive predecessor observations. They do not govern the current Situation Room, are never pooled with SR-8R1, and do not demonstrate a causal capacity effect, internal cue use, general Qwen behavior, Chinese-model behavior, or Anthropic/Google performance.
Current provider paths
Six live qualification calls—three Anthropic and three Google—tested current request, response, refusal/error-envelope, and custody behavior. No Parallax scientific prompt was sent. They establish dated transport-path qualification, not a scientific provider result or permanent compatibility guarantee. A full run would require fresh sacrificial qualification.
6. The full reference study that has not run
The sealed reference plan has two providers, 30 cells per provider, N=9270 per cell, 278,100 calls per provider, 556,200 total scheduled calls, M=210, and global alpha=.05. It preserves provider separation, fail-closed retry and availability rules, and no full-joint power claim.
Full confirmatory execution funding is not secured. Account capacity is unverified. Those facts block execution, not the existence or publication of the protocol. No Experiment 3 scientific provider result exists.
7. What is public now
Parallax now pairs this research narrative with a public Situation Room. A reader can inspect the governing five-dose local response profiles, matched-wording contrasts, failure classes, historical transfer, shortcut diagnostics, and exact claim boundaries without reading raw JSON first.
The story’s value is not a triumphant benchmark number. It is the practical discipline required to make an evaluation capable of saying “not supported” and “insufficient evidence” without quietly moving the goalposts.
8. Reproduction and the current public release
Run from the public-release root:
python reproducibility/verify_policy_reader_presentation.py
The check requires no GPU, provider credentials, network access, or model calls.
PUBLIC RESEARCH RELEASE = LIVE EXPERIMENT_1 = NOT_SUPPORTED EXPERIMENT_2 = INSUFFICIENT_EVIDENCE EXPERIMENT_3_CONFIRMATORY_EXECUTION = NOT_PERFORMED SCIENTIFIC_PROVIDER_RESULT = NONE FULL_CONFIRMATORY_EXECUTION_FUNDING = NOT_SECURED GOVERNING_LOCAL_CAMPAIGN = SR-8R1 PUBLIC_SUBMISSION_ACTIONS = OUTSIDE_THIS_RELEASE