Skip to content

Atlas acceptance design

Status: planning draft, 2026-09-06. Companion to the target state and delivery plan. These files specify proposed behavior and contracts. They are not supported runtime configuration, qualified judges, or results from an upgraded Atlas.

1. Reviewable inputs

Artifact Contents Current standing
Behavior specification Nine commitments, rubric anchors, precedence, applicability Proposed product contract
Scenario families 18 cases in six families Draft expectations; diagnostic selection
Observed failures Original user/assistant records for Atlas q31 and q40 Real saved outputs, copied verbatim
Provenance Source path, selected lines, hashes, limitations Verified copy identity; historical environment incompletely known
Executor decision Executed acceptance fixtures and ownership decision Native path selected; local prototype limits documented

The behavior spec is independent of the implementation prompt. It can judge a prompt, retrieval, model, or application change against the same product commitments. It derives from Atlas's current working tree and existing tasks, with proposed clarifications called out in the YAML. In particular, correctly converted units and user-supplied quantities need different treatment from invented figures. Product review should settle that policy.

The first corpus deliberately avoids new clinical answer keys. Technical correctness and hazard expectations need suitable review; relevance and obvious resource conflicts can be inspected before those references exist. This is a qualification seed, not a representative estimate of all Atlas usage. Keep families together when splitting data.

2. What the real failures teach us

q31: no rope. The user asks what can substitute for rope when lashing a shelter frame. The saved answer tells the user to wrap their arms around the poles, then ends with timber hitches without identifying binding material. This gives a concrete proposed failure label for respect_constraints and useful_next_step. A relevance-only rubric might reward the lashing vocabulary; a good rubric asks whether the described procedure solves the missing-material problem. These are planning judgments awaiting review, not historical scorer outputs.

q40: leaving a homestead for a week. The saved answer emphasizes car winterization, insulation, snow shoveling, and frostbite. It does not organize the answer around what must be dealt with during the absence. The proposed task_match failure does not require a single canonical homestead checklist: acceptable answers may prioritize different systems after clarifying consequential facts. The family's livestock and empty-property variants test whether those facts actually change priorities.

The source excerpts are useful evidence about what was captured. The legacy adapter does not establish a complete historical prompt or a verified citation-number mapping merely by saving a list of excerpts. Import preserves that uncertainty. It must not manufacture a fully reproducible baseline from the current Atlas checkout.

3. Three execution surfaces

Surface Cases Purpose Requirement
Actual application 12 cases: materials, wet fire, winter absence, conversation Evaluate Atlas retrieval, answer, and session behavior together Request identity, reset, observed context, configuration capture
Controlled application 3 evidence-authority cases Hold evidence constant to test role and uncertainty behavior Explicit excerpt-injection test seam; record the intervention
Evaluator only 3 synthetic quantity cases Qualify conversion and contradiction detection Fixed answer, citation map, and evidence; no Atlas generation

Report these populations separately. Passing the quantity fixtures says something about the evaluator. Passing a controlled-evidence case says something about Atlas under that intervention. Neither establishes performance of the whole retrieval application.

An absent evidence field is a fourth condition for the capture contract, distinct from observing an empty retrieval. Pair the same observation with capture: complete and empty excerpts versus capture: unknown and no excerpts during schema/evaluator tests. The former can support an uncertainty-behavior judgment; the latter leaves evidence-dependent judgments unavailable. No-evidence behavior is not a grounding score of 1.0. Missing telemetry alone is not an application failure verdict.

The session adapter must deliver all user turns in order using actual earlier application answers. It must verify a clean start between cases and repetitions. The existing adapter forwards the final user message and accepts a response based on matching question text; that is insufficient to claim these follow-up and repeated-question tests have run correctly. Unsupported capabilities should fail preflight rather than silently degrade to single-turn tests.

4. What one evaluated case must retain

For binding_no_rope, an implementation should be able to follow this chain without reconstructing state from a report:

case revision + behavior revision + target/configuration snapshot
  -> planned sample (case, repetition, session identity)
  -> dispatch attempt (unique request ID, lease, budget reservation)
  -> observation (answer, evidence capture, usage, error/termination)
  -> evaluation (criterion version, observation/input digest, evaluator version)
  -> review annotation (reviewer, rationale, evidence spans, supersedes if revised)
  -> comparison snapshot and gate result

Public names remain readable; immutable digests identify content. A new request nonce distinguishes repeated identical questions. A retry is a new attempt within the sample; an intentional repeat is a new sample. Neither changes the case's meaning or identity.

Each criterion result carries execution status separately from its verdict. A scored constraint judgment might be violated, with the answer span requiring rope and the user span excluding it. A source-support judgment can be missing_evidence, with no numeric value or verdict. A technically correct procedure requires its own reviewed domain rubric; it is not inferred from passing style or constraint checks.

Human review changes labels by appending an annotation and selecting a new evaluation snapshot. It does not overwrite the observation or silently alter a previous CI decision. Model-proposed labels remain draft until reviewed. Quote spans are pointers into preserved artifacts, and every claimed source ID must resolve within the declared evidence selection.

5. Worked comparison and gate outcome

The table below is entirely synthetic. It illustrates result semantics, not measured Atlas improvement. Assume three independent cases with reviewed, blocking constraint criteria, and one planned sample per case for each target.

Case Baseline execution / criterion Candidate execution / criterion Saved comparison
A: missing material Succeeded / violated Succeeded / met Improvement on this criterion
B: material available Succeeded / met Succeeded / met Unchanged
C: follow-up removes substitute Succeeded / violated Timed out / unavailable No comparable judgment

Expected report:

Candidate decision: INCONCLUSIVE
Execution: 2 / 3 completed; 1 timed out
Required criterion coverage: 2 / 3 (66.7%)
Criterion met among scored cases: 2 / 2 (100%)
Demonstrated criterion successes among planned cases: 2 / 3 (66.7%)
Observed blocking violations: 0; unresolved required cases: 1
Comparable pairs: 2 / 3; improved: 1; unchanged: 1; regressed: 0
Missing pair: C
Gate: full required coverage not satisfied

The last success count is an operational denominator, not a claim that the timed-out answer was behaviorally wrong. Both conditional quality and completion remain visible. With full required coverage in policy, zero observed violations cannot turn this candidate green. Under the proposed CLI contract, this incomplete required evaluation returns exit code 3. A known blocking violation independently fails its gate even if other required results are missing; all failed and inconclusive gate reasons remain visible.

No confidence claim follows from one improvement in two comparable pairs. A release study needs a predeclared population, appropriate grouped sampling, sufficient evidence, and a policy for incomplete pairs. Changed evaluator versions require evaluating both saved output sets under the same version before making comparable claims. Legacy scores remain historical and are not silently relabeled.

Comparison itself reads saved results. A fresh pairwise judge is explicit evaluation work with a declared budget, producing saved results before comparison. Reopening the report, adding a slice, or exporting JUnit must not dispatch any model call.

6. First useful delivery to Atlas

  1. Import the existing 40-case run and these two failure records with honest provenance. Show missing fields, the original answer, proposed labels, and evaluators that cannot run.
  2. Use deterministic target/judge fixtures to prove recovery, status, budget, and coverage contracts before consuming physical-device time. Qualify quantity semantics independently.
  3. Upgrade the device bridge for identity and sessions, then run the existing suite through incremental persistence. A killed run must retain committed answers and resume safely.
  4. Review the behavioral families and qualify the first task/constraint judge against labels. Add new clinical or other domain expectations only with appropriate reference review.
  5. Produce an actual baseline/candidate comparison against a declared Atlas change, showing incomplete work and uncertainty as clearly as improvements. The pending stricter numeric prompt is a possible candidate, not an already demonstrated improvement.

Ordinary CI runs offline executor and evaluator contract fixtures. The physical-device suite is an explicitly selected job on a machine with the device bridge; generic hosted CI cannot be assumed to run it. Device availability and thermal conditions belong in the record, with scheduling policy agreed before comparing performance.

AverCare should add one real workflow to this same path once located. Its first planning input is the workflow and application boundary, not an assumed health-app rubric or a new platform abstraction. Atlas work can proceed independently.