INCLINATION.json

{
  "schema_version": "inclination.v0",
  "intents": [
    {
      "id": "INT00001",
      "origin": "Engineering with LLMs in production; safety dependent on agent self-policing.",
      "intent": "Agentic action can be bounded, observable, and verifiable, without trusting the agent or forfeiting expected utility.",
      "horizon": "now"
    },
    {
      "id": "INT00002",
      "origin": "Frontier-model dependence in independent labs; open-model performance gap.",
      "intent": "Open-model agent performance can be stabilized to approach prior-generation frontier performance, with closed vocabulary, deterministic tools, cross-model evaluations, and interpretability-informed training.",
      "horizon": "next"
    },
    {
      "id": "INT00003",
      "origin": "Vendor-platform lock-in observed across operator workflows.",
      "intent": "Rebuild-from-runtime sovereignty, achieved by deriving replacements for vendor components when a vendor dependency's friction exceeds its operational value.",
      "horizon": "later"
    }
  ],
  "hypotheses": [
    {
      "id": "HYP00001",
      "intent_ids": ["INT00001"],
      "hypothesis": "Closed-vocabulary AST-CI gates reduce off-vocabulary generation rate compared to no gate, measured on generated code across imperative languages.",
      "ablations": [
        "Vocabulary as advisory warnings instead of a gate produces equivalent off-vocabulary rate.",
        "Prompt instructions alone produce equivalent off-vocabulary rate.",
        "Disabling the AST walker produces no measurable change in off-vocabulary rate."
      ],
      "type": "empirical",
      "dependencies": [],
      "experiments": ["EXP00001"],
      "conclusion": null
    },
    {
      "id": "HYP00002",
      "intent_ids": ["INT00001"],
      "hypothesis": "Substrate records (Atlas, Git, Gateway) contain sufficient evidence to reconstruct any agent action without reliance on the agent's self-report.",
      "ablations": [
        "Reconstruction from substrate records alone fails to recover what occurred in audited sessions.",
        "Agent self-report alone produces reconstruction equivalent to substrate records.",
        "Atlas-disabled sessions remain equivalently reconstructable."
      ],
      "type": "operational",
      "dependencies": [],
      "experiments": ["EXP00002"],
      "conclusion": null
    },
    {
      "id": "HYP00003",
      "intent_ids": ["INT00001", "INT00003"],
      "hypothesis": "Propagate's State load at session start re-orients a new agent session to current work context without operator intervention.",
      "ablations": [
        "Propagate's State load drifts from current work context in ways the operator must correct.",
        "Operator must manually re-orient the new agent session despite Propagate's State load.",
        "Vendor-managed memory (Claude memory, ChatGPT memory) provides equivalent re-orientation; Propagate's State load adds nothing.",
        "Propagate's State load fails to support continuation when vendor memory is also disabled."
      ],
      "type": "operational",
      "dependencies": ["HYP00002"],
      "experiments": ["EXP00010"],
      "conclusion": null
    },
    {
      "id": "HYP00004",
      "intent_ids": ["INT00001"],
      "hypothesis": "Substrate enforcement layers (Seatbelt, Intermediary, Atlas, identity isolation) compose such that bypassing any single layer does not produce unauthorized capability.",
      "ablations": [
        "Bypassing Seatbelt alone yields Atlas write capability.",
        "Removing Intermediary mediation yields equivalent boundedness.",
        "Collapsing identity isolation to a shared POSIX user yields no measurable change in bypass surface."
      ],
      "type": "operational",
      "dependencies": [],
      "experiments": ["EXP00003"],
      "conclusion": null
    },
    {
      "id": "HYP00005",
      "intent_ids": ["INT00001"],
      "hypothesis": "Cross-model review composed with eval produces a higher defect detection rate than weaker baselines (eval-only, same-vendor review, review-without-eval) on a labeled task corpus.",
      "ablations": [
        "Removing the eval step produces equivalent detection rate.",
        "Same-vendor self-review produces equivalent detection rate.",
        "Random-verdict placebo produces equivalent detection rate."
      ],
      "type": "empirical",
      "dependencies": [],
      "experiments": ["EXP00004"],
      "conclusion": null
    },
    {
      "id": "HYP00006",
      "intent_ids": ["INT00001", "INT00003"],
      "hypothesis": "Substrate enforcement holds equivalently under remote operator access (SSH-over-Tailscale) compared to local operation.",
      "ablations": [
        "Remote sessions exhibit measurable enforcement relaxation.",
        "Vendor-mobile harness (Claude Code web, Codex cloud) covers operator mobile need without requiring remote substrate enforcement.",
        "Remote operation produces measurably degraded enforcement integrity."
      ],
      "type": "operational",
      "dependencies": [],
      "experiments": ["EXP00005"],
      "conclusion": null
    },
    {
      "id": "HYP00007",
      "intent_ids": ["INT00001"],
      "hypothesis": "Substrate-active agent work produces equivalent work-quality scores compared to substrate-ablated work, with model and task held constant.",
      "ablations": [
        "Substrate-active sessions produce measurably lower work-quality scores than ablated sessions.",
        "Substrate-induced friction reduces task completion rate below ablated baseline."
      ],
      "type": "empirical",
      "dependencies": [],
      "experiments": ["EXP00006"],
      "conclusion": null
    },
    {
      "id": "HYP00008",
      "intent_ids": ["INT00002"],
      "hypothesis": "Substrate-active agent runs produce lower across-run variance compared to substrate-ablated runs, with model and task held constant.",
      "ablations": [
        "Substrate-active runs produce equivalent or higher across-run variance.",
        "Observed variance reduction is attributable to randomization rather than substrate composition."
      ],
      "type": "empirical",
      "dependencies": [],
      "experiments": ["EXP00006"],
      "conclusion": null
    },
    {
      "id": "HYP00009",
      "intent_ids": ["INT00002"],
      "hypothesis": "Interpretability-informed training applied to an open model produces a tuned variant that performs measurably better in the substrate than the untuned base model.",
      "ablations": [
        "Training without an interpretability-derived signal produces equivalent improvement.",
        "The tuned variant performs equivalently or worse than the untuned base model in the substrate.",
        "The interpretability-derived preference signal correlates with no measurable model behavior."
      ],
      "type": "empirical",
      "dependencies": [],
      "experiments": ["EXP00007"],
      "conclusion": null
    },
    {
      "id": "HYP00010",
      "intent_ids": ["INT00002"],
      "hypothesis": "An open model tuned via interpretability-informed training (HYP00009), run within the full apparatus, produces performance comparable to prior-generation frontier on a fixed work-task corpus.",
      "ablations": [
        "The tuned open model in the full apparatus fails to approach prior-generation frontier performance.",
        "The untuned open model in the apparatus produces equivalent performance, showing tuning contributes nothing.",
        "Performance improvement is attributable to raw model capability rather than substrate or tuning."
      ],
      "type": "empirical",
      "dependencies": ["HYP00009"],
      "experiments": ["EXP00008"],
      "conclusion": null
    },
    {
      "id": "HYP00011",
      "intent_ids": ["INT00002"],
      "hypothesis": "Substrate stabilization holds across multiple open instrumentable model families (Qwen, Gemma, gpt-oss), not only the family used to develop the apparatus.",
      "ablations": [
        "Stabilization holds only for the family used during apparatus development.",
        "Performance varies significantly across families with apparatus held constant, indicating family-specific tuning rather than substrate stabilization."
      ],
      "type": "empirical",
      "dependencies": ["HYP00010"],
      "experiments": ["EXP00008"],
      "conclusion": null
    },
    {
      "id": "HYP00012",
      "intent_ids": ["INT00002", "INT00003"],
      "hypothesis": "Substrate-mediated runs produce reproducible per-task cost compared to ad-hoc vendor-API agent invocation, measured as low cost variance across reruns of fixed task sets.",
      "ablations": [
        "Substrate-mediated runs exhibit cost variance comparable to ad-hoc agent invocation.",
        "Reproducibility requires extensive operator manual tuning per session rather than emerging from substrate properties.",
        "Cost reproducibility requires holding non-cost variables constant in ways that diverge from normative use."
      ],
      "type": "empirical",
      "dependencies": ["HYP00010"],
      "experiments": ["EXP00008"],
      "conclusion": null
    },
    {
      "id": "HYP00013",
      "intent_ids": ["INT00003"],
      "hypothesis": "Substrate enforcement, observability, and evidence guarantees hold equivalently regardless of which underlying infrastructure provides the substrate's runtime (harness identity, network layer, vendor service).",
      "ablations": [
        "Substrate guarantees vary measurably by underlying infrastructure component.",
        "Apparent neutrality is achieved only through component-specific compensations in substrate code.",
        "Substituting a previously-validated infrastructure component reveals contract gaps the original setup masked."
      ],
      "type": "operational",
      "dependencies": [],
      "experiments": ["EXP00009"],
      "conclusion": null
    }
  ],
  "experiments": [
    {
      "id": "EXP00001",
      "intent_ids": ["INT00001"],
      "hypothesis_ids": ["HYP00001"],
      "label": "Measurement of off-vocabulary generation rate in agent-produced code under closed-vocabulary AST-CI gating compared to no gating, across imperative languages.",
      "push": null
    },
    {
      "id": "EXP00002",
      "intent_ids": ["INT00001"],
      "hypothesis_ids": ["HYP00002"],
      "label": "Audit of agent action reconstruction fidelity from substrate records (Atlas, Git, Gateway) compared to reconstruction from agent self-report alone.",
      "push": null
    },
    {
      "id": "EXP00003",
      "intent_ids": ["INT00001"],
      "hypothesis_ids": ["HYP00004"],
      "label": "Validation of substrate enforcement composition by attempting single-layer bypass against each enforcement layer (Seatbelt, Intermediary, Atlas, identity isolation) and observing whether unauthorized capability emerges.",
      "push": null
    },
    {
      "id": "EXP00004",
      "intent_ids": ["INT00001"],
      "hypothesis_ids": ["HYP00005"],
      "label": "Measurement of defect detection rate against a labeled task corpus, comparing cross-model review composed with eval to weaker baselines (eval alone, same-vendor review, review without eval).",
      "push": null
    },
    {
      "id": "EXP00005",
      "intent_ids": ["INT00001", "INT00003"],
      "hypothesis_ids": ["HYP00006"],
      "label": "Comparison of substrate enforcement integrity between local operator sessions and remote operator sessions over SSH-over-Tailscale, executing equivalent substrate operations.",
      "push": null
    },
    {
      "id": "EXP00006",
      "intent_ids": ["INT00001", "INT00002"],
      "hypothesis_ids": ["HYP00007", "HYP00008"],
      "label": "Measurement of agent work quality and across-run variance on a fixed task set, comparing substrate-enabled sessions to substrate-ablated sessions with model and task held constant.",
      "push": null
    },
    {
      "id": "EXP00007",
      "intent_ids": ["INT00002"],
      "hypothesis_ids": ["HYP00009"],
      "label": "Comparison of in-substrate performance between an open base model and its variant tuned via interpretability-informed training, on a fixed task set.",
      "push": null
    },
    {
      "id": "EXP00008",
      "intent_ids": ["INT00002", "INT00003"],
      "hypothesis_ids": ["HYP00010", "HYP00011", "HYP00012"],
      "label": "Comparison of in-substrate performance and per-task cost across tuned open-model families (Qwen, Gemma, gpt-oss) against prior-generation frontier on a fixed work-task corpus; reruns measure cost variance.",
      "push": null
    },
    {
      "id": "EXP00009",
      "intent_ids": ["INT00003"],
      "hypothesis_ids": ["HYP00013"],
      "label": "Comparison of substrate enforcement, observability, and evidence guarantees across substitutions of underlying infrastructure components (harness identity, network layer, vendor service).",
      "push": null
    },
    {
      "id": "EXP00010",
      "intent_ids": ["INT00001", "INT00003"],
      "hypothesis_ids": ["HYP00003"],
      "label": "Comparison of new-session orientation under Propagate's State load, no load, and vendor-managed memory alone, measured by required operator intervention and drift from current work context.",
      "push": null
    }
  ]
}