3.9 KiB
Investigation working view v2
New project-enabled investigation states freeze prompt_protocol=2. Existing
states without that marker retain their original prompts, request hashes,
contexts and budgets. Non-project loops retain their previous protocol. No
migration or rewriting of historical evidence is performed.
Complete archived context
The canonical case, accepted interpretation, full investigation state, ledger, evidence records and contents remain archived. New contexts additionally load the accepted interpretation context and check its SHA-256 against the confirmed interpretation. Its source aliases include separately captured entity-resolution facts (e.g. F0029 in the reference trial); these were previously missing from the loop's canonical-case-only input. No additional database or model call is needed.
Model working view
- Case identity and all readable interpretation-context source aliases replace the duplicated canonical case. No readable observation, description, fact or resolved identity is filtered by relevance or age.
- Source IDs address full archived provenance. Repeated archive/JSON navigation is omitted from top-level source provenance in the model view; authors, timestamps, attribution kinds and other semantic origin fields remain. Raw query values, predicates, rows, truncation and snapshot scope remain available.
- The first 12 source/database reference entries retain the previous project's bounded metadata-catalog behavior. Counts, truncation and unlisted source IDs are explicit. Identical file metadata is listed once with all source IDs. Raw file inventories are not observations and cannot prove absence. All file metadata remains in the archived context. Project repository/DB searches do not depend on this bounded packaged-file catalog.
- The selected hypothesis retains its full proposal, live questions and notes. Alternatives retain mechanism, distinction, falsifying evidence, live state, questions, notes and evidence references. Initial open_questions are not sent twice because the live question list already contains them.
- Current balances, policy, low-information counters and pending actions remain. Settled actions become a count in the model view; global novelty hashes remain the controller's responsibility. Ledger calculations do not use this view.
- Evidence content from ALL hypotheses remains readable, including negative and cross-hypothesis findings. Internal novelty atoms are omitted; semantic source content, query scope, completeness and provenance remain. The previous result is referenced once through its evidence ID.
- Long storage identifiers and repeated provenance use the existing deterministic P/M encoding. The duplicate project query schema is omitted from project binding metadata; allowed arguments remain in the full output schema.
Known P aliases in tool source_ref/source_refs arguments are expanded from
this exact working context before the normal preparation, permission checks,
budget reservation and execution. Only these fields are expanded; database
filters and search text are never rewritten. Unknown aliases are not resolved
and still face the existing tool validators. Raw model replies remain archived.
Measurement and limits
On the current compact-hypotheses-20261015 first-step input, initial messages fall from 83416 to 56063 characters (32.8% fewer), while restoring missing resolved identity evidence. This is an offline character comparison, not actual provider tokens or a demonstration of live model quality.
This is the first safe reduction, not a constant-size research memory. Old result payloads can still grow with investigation length. They are deliberately NOT silently truncated, discarded or replaced by an LLM summary. Further reduction of those payloads requires a separate reliable read-back mechanism and tests for retaining old counterevidence. No paid investigation was started by this change.