Experiment

Read the research constitution and execute the
managed research workflow and
scientific contribution decisions. Begin or resume with
exactory-research status --summary and next --summary. Keep the complete objective, admitted cycle,
resource account, advisory preparation, and exact output contract in view.
This is a tool available at every stage.
The caller states the question, depth and budget. Record effective amounts,
actual work, outcomes, stopping reason and cost. The orchestration holds the
gates that accept a direction or result; a tool record grants no approval.

Probe before or during any milestone

Use exactory-lab probe for a quick calculation before choosing a theme, while
comparing approaches, or during a later stage. Configuration is required; a
target and admitted cycle may still be absent. State the exact question, program,
inputs and wall-time allowance:

exactory-lab probe --id example-001 --question "Does the small example satisfy the conjectured relation?" \
  --script experiment/check.py --input experiment/data.json --output-dir results --timeout 30

The two-phase probe contract pins program
and inputs and records probe-intent before launch, reserving the timeout against
the experiment purpose. The observed probe outcome pins outputs, stdout and
stderr, records exit status and actual elapsed time, charges that usage and
releases the reservation, including failures, timeouts and output-bound outcomes.
Launch ownership is persisted and exclusive. The frozen supervision.json
binds worker/program identities, process group and deadline. An independent
watchdog enforces the deadline after worker loss even without a retry. It writes
supervisor-timeout.json, terminates the program group, verifies termination
and writes supervisor-terminal.json. A later retry records the outcome and
reconciles actual usage before releasing the reservation.

Repeating an ID returns its outcome
or pending for a live owner. Reconcile an unfinished intent only after owner
loss or checked terminal evidence and confirmation that the original computation
has stopped; retain the reservation until reconciliation. Preserve measured
elapsed time and record unmeasured usage as unknown.

Cite probes in dossiers with their record references. Probe artifacts retain
exploratory origin through references and imports, with no prospective-test
credit. Existing supported imported evidence retains its standing. Independent
support assessment determines the additional validation a claim needs; repeat an
execution when that validation requires it and retain the prior chronology.
Computations relied on by recorded mathematical strategy decisions, proofs,
finite checks, counterexamples or acceptances require native admission.
Follow the native provenance contract
for imported exploratory output and exploratory_evidence.

Plan and execute each cycle

Use preliminary implementation, baseline tuning, hypothesis tests, and ablation
where they answer the question. Choose repetitions, seeds, comparison methods,
and stopping conditions from noise, scientific requirements, and available
resources; no fixed number of cycles or seeds establishes validity.

Each cycle belongs to a current investigation decision and named resource tranche.
Run the earliest affordable test that can change the commitment. A necessary
baseline has a bounded completion condition and an exit into that deciding test.
Check whether the method can distinguish the hypotheses and whether its proxy
supports the target inference. A bounded adequacy probe may establish that link;
extra precision on a shared fit cannot establish it by itself.

For every admitted prospective run, write the cycle, pin actual program/input
bytes with artifact, admit the current plan and reservation, and bind-run the
exact backend, seed, timeout, paths, outputs, and usage. These are prerequisites
for debugging, tuning and plotting when they form part of that planned execution,
including a return from manuscript revision. Quick exploratory work uses the
probe contract above. Planning and admission require configuration and a current
cycle decision; the complete synthesis report remains advisory. Launch with the actual admission ID:

exactory-lab run code/program.py --admission run-001 --backend local --timeout 30 \
  --expected-revision 42 --request-id launch-cycle-001

The admitted argv[0] must resolve to the Python that runs bind-run and
exactory-lab run. For a venv interpreter, start both with it, where VENV is the
venv directory:

VENV/bin/python3 "$(command -v exactory-research)" bind-run --file run-binding.json \
  --expected-revision 41 --request-id bind-cycle-001
VENV/bin/python3 "$(command -v exactory-lab)" run code/program.py --admission run-001 \
  --backend local --timeout 30 --expected-revision 42 --request-id launch-cycle-001

The worker then starts that venv interpreter, so the program has the venv's packages.
The launch of a binding that exactory-client 0.47.0 or earlier recorded starts the
resolved file instead. To give the program the venv's packages, admit and bind a
new run.

Use the current revision and exact bound values, as described in the workflow.
The worker executes a private copy of pinned inputs and seals the observed output
inventory. Metrics may be printed as JSON or written to the bound fallback path,
but a process exit or metric alone is not a scientific assessment. Read the saved
outputs and original observation before recording conclusions. Keep the human
experiment/journal.jsonl consistent with the authoritative execution history.

Keep scripts and data within experiment/. Choose data, scale and methods adequate
for the question within authorized resources. Small public or synthetic data are
useful when they answer the deciding question or a named prerequisite. Set and record meaningful seeds,
versions, and null-seed reasons. A guard finding requires a code or design fix.
The local backend suits work that fits CPU or MPS. Bind Colab only for a live
runner and a justified test; if unavailable, prospectively admit an adequate local
test or preserve the unresolved resource requirement. Retrying the same admission
reconciles the existing execution and never launches it again.

Assess and deepen

Separate a valid negative result from invalid methodology, a code defect, timeout,
missing output, or unknown execution. Retain every outcome and its cost. Use
reconcile-run for interrupted observations; unknown released work retains its
reservation and cannot provide completed evidence. Assess the real result and
validity checks separately, with scope, novelty, contribution, assumptions,
failure signals, and remaining obligations for the full objective.

Save a durable checkpoint with its stable ID, exact claim, evidence, verification
status, unresolved obligations, and next hypothesis. Deepen promising branches
using prospective successor cycles and explicit inheritance. Preserve failed
branches and reopen them only when new evidence addresses the recorded obstruction.
A branch result does not complete the full objective by itself.

Link developments and reviewer requests to stable work items, retaining the
original request and evidence. Action disposition differs from execution status:
an adopted next_round step may remain unattempted. A validity gap blocks a
retained claim; a consequence-critical gap blocks the promised inference. Optional
development is compared by scientific value and opportunity cost.

When work exceeds current scope or resources, next_round records its schedule,
reason, evidence, resource implications, reopening condition and effect on claims.
It does not discharge a critical gap. At the next commitment, test that gap,
establish a concrete obstruction or failed bounded probe, or justify a different
route or branch closure. Feasible but unchosen work is not infeasible. Preserve
legacy carried dispositions and link them to the same work items.

At tranche end, question resolution or a failure signal, update the result dossier
with actual knowledge gained, failed approaches, realized or unmet consequences
and cumulative resource use. A new cycle ID cannot renew unchanged work without
a current decision. If the supported consequence is smaller, reconsider whether
it would justify starting the study now, comparing current alternatives without
credit for sunk effort. Preserve the full objective and every unresolved gap.

For scalar parameter optimization, use a noise-aware baseline and prospective
comparisons. Select the best supported source variant while retaining every
attempt, observation, failure, review, and resource charge. Restoring source code
never restores an earlier budget or erases the search history. Stop an optimization
phase according to its recorded scientific and resource conditions; an exhausted
budget is not evidence of success or impossibility.

Combined result assessment before writing

After selecting an assessed checkpoint, extend the same strategy dossier with
the candidate, plans, sources, sealed observations, exact supported claims,
comparisons, failures, work-item execution and consequence status. No manuscript
is needed. Deliver the canonical result packet to two independent assessors on
verified reviewer routes. Each checks validity, scope, novelty, contribution
support and branch obligations before value in one response. Preserve unchanged
outputs and exact assignment/evidence bindings. The canonical support and value
findings derive compatible readiness; do not add a separate full readiness-review
loop. Blind manuscript measurement later evaluates the paper separately.

Inspect each result review-run output artifact for source_requests before
recording final support or value findings. Apply the same rule to bar or slate
reviews when a changed intent requires them. Follow the native continuation contract
in the same reviewer slot. Acquire unavailable exact sources through the ordinary
pipeline and retain pending or unaffordable requests as unresolved. A request
response is not a final value-review; both final bars must exist before the
slate is released. Reuse unaffected existing bars when the intent is unchanged.

Record the evidence-bound research-decision and run
research-decision-assess --boundary write with the current readiness gate.
develop_manuscript needs both supported claims and a worthwhile realized
consequence. Version-2 decisions require q1_venue passed in both reviews on
the realized supported advance and actual evidence. deliver_requested can use an existing unconditional delivery
instruction for a valid limited result while retaining an unmet bar and objective.
An explicit quality condition remains binding. Selected source-limited publication
contracts retain gate manuscript-readiness; whole-objective gate readiness
continues to report the original objective.

Address findings with source corrections, appropriate validity checks, or further
scoped cycles and affected independent reassessment. A material disagreement uses
focused adjudication with the objection retained. If value is still unmet, make a
justified investigate, pivot, close-branch or authorized-delivery decision; do not
manufacture success by narrowing the original objective. Enter write with
--status pending only when the current action and applicable gates permit it.

Prepare result summaries and plots with stated units, populations/denominators,
intervals, outcomes, baselines, uncertainty, and actual evidence references.
Record every reportable claim in evidence/claims.json, including unknown or
inapplicable dimensions with reasons. Check derived equations with
exactory-derive check where applicable and report the check's actual scope.

Report the supported result, comparisons, validity limits, failed approaches,
remaining objective, and resource usage. Resume from current status, retained
admissions, checkpoints, and the search tree without resetting history.