68 lines
4.0 KiB
Markdown
68 lines
4.0 KiB
Markdown
# Early diagnosis review v1
|
|
|
|
An investigation may request independent final evaluation before all hypothesis
|
|
accounts become terminal. This is not acceptance of a cause or a coding task.
|
|
The service performs no patch, SQL write or client automation.
|
|
|
|
## Request and prerequisites
|
|
|
|
New investigation prompts (`investigation_v2`, `investigation_project_v3`) offer
|
|
`next.kind = request_diagnosis`, with `reason` and `countercheck_refs`.
|
|
Current prompts are `investigation_v3` / `investigation_project_v4`; older
|
|
persisted prompt versions remain reproducible.
|
|
The selected hypothesis must be `strongly_supported`, cite concrete nonempty
|
|
investigation evidence, and cite findings that meaningfully test an alternative.
|
|
Evidence is shared: collecting hypothesis IDs never restrict its use. Another
|
|
supported alternative does not by itself establish a material contradiction.
|
|
The controller validates IDs and structure; whether the evidence establishes
|
|
causality and whether a negative result has sufficient coverage remain explicit
|
|
responsibilities of the independent final evaluator.
|
|
|
|
A request first becomes `preparing`. Outstanding results must be assessed before
|
|
review: the controller schedules their hypotheses, and permits `defer` (assessment
|
|
without closing the hypothesis) or a justified `reject`. No fresh research action
|
|
is launched while preparing. `defer` is not available during ordinary research.
|
|
All action reservations must be settled before entering BUILDING_DIAGNOSIS.
|
|
No account is artificially exhausted and no alternative is implicitly refuted.
|
|
|
|
The requested state includes the leading hypothesis, evidence, countercheck,
|
|
reason and explicitly deferred alternatives. The existing confidence scheduler
|
|
and its initial turn for each hypothesis remain in use outside preparation.
|
|
|
|
## Independent decision and continuation
|
|
|
|
`diagnosis_early_v2` evaluates every hypothesis, all evidence and remaining
|
|
uncertainty independently. Acceptance requires the requested hypothesis,
|
|
`strongly_supported` confidence, direct evidence for it, an adequate evidence-grounded
|
|
alternative countercheck, and no explicitly unresolved material contradiction. Ordinary
|
|
final diagnosis validation remains mandatory. An accepted review creates the
|
|
normal diagnosis artifact; the existing separate Change Request and output
|
|
validators still decide whether a coding task can be issued. Missing coordinates
|
|
or other implementation prerequisites must remain explicit missing information.
|
|
|
|
An inconclusive review, insufficient confidence, unresolved material counterevidence or exhausted
|
|
invalid-response retries declines the request. The case returns to
|
|
INVESTIGATING_HYPOTHESES in the same analysis revision. The immutable review reason
|
|
is included in subsequent investigation context. No failure report or coding
|
|
task is issued by this rejection. Account balances, assessed results, questions
|
|
and evidence survive. The only reverse lifecycle transition is guarded by the
|
|
persisted pending request and its declined review. It is not a general rewind.
|
|
|
|
Each review has its own context, calls and error namespace under
|
|
`analysis/diagnosis/NNNN/early/NNNN/`. Historical requests and prompts stay intact;
|
|
completed acceptance still uses the canonical diagnosis completion checkpoint.
|
|
A process restart reuses the persisted model call rather than requesting it again.
|
|
Pending old investigation prompts retain their original protocol. Existing loops
|
|
adopt early-review support at their next decision boundary.
|
|
|
|
## Limits
|
|
|
|
`[investigation].max_early_reviews = 3` is a configurable operational default,
|
|
frozen with the revision policy. Older policies without it use three. Reaching
|
|
the limit disables further early requests, not ordinary investigation or normal
|
|
final evaluation. A declined request also needs new evidence content before
|
|
another request; a fresh ID for identical content does not suffice.
|
|
This prevents repeated unchanged final evaluations. It is independent of per
|
|
hypothesis research budgets, token limits and provider retry configuration.
|
|
Confidence levels are qualitative ratings, not calibrated probabilities.
|