Remove evidence-certification bookkeeping; keep the benchmark harness

[deepfates]

What a reader of the public repo gains

Before this change, most of the tracked bytes were a record that certain benchmark
runs happened: sealed .tar.zst archives, content-addressed "admitted evidence"
files, registries binding claims to those files, and tests whose only assertion was
that a digest still matched. None of it let an outside developer run anything.

What is left is everything a stranger can run. The dataset fetchers and loaders, the
scorers, the campaign runners, the benchmark mix tasks, the differential harness
against pinned DSPy and GEPA, the upstream pins it reads, and every test that
exercises a task, a campaign runner, a scorer or a differential against a local
fixture or the pinned upstream. mix differential.check runs the whole
:dspy_parity lane unchanged.

Clone size: 1599 tracked files / 87.9 MB becomes 1249 files / 15.3 MB.

Deleted, by category

Category	Files	Size
evidence/ sealed archives	15	25.78 MB
benchmarks/ results, runs, admitted evidence, locks, claim and reproduction registries	179	28.82 MB
examples/ matched-experiment output dumps	112	17.47 MB
test/ certification tests	15	0.12 MB
docs/ maintainer registry docs and the generated conformance ledger	9	0.11 MB
lib/mix/tasks/ registry and coverage-ledger tasks	7	0.07 MB
bench/ registry and admission modules	6	0.12 MB
scripts/ with no runnable caller	5	0.05 MB
identity/	1	0.00 MB
.github/workflows/evidence.yml	1	0.00 MB

350 files, 72.5 MB.

What was kept, and why

benchmarks/data/, benchmarks/config/, benchmarks/authorities.json (whole, including
its families section — five differential Python scripts and four differential mix tasks
read it), benchmarks/authority_sources/, benchmarks/upstream/gepa-corrected/ (the
patch scripts/setup_corrected_gepa_comparator.sh applies in the CI differential job),
and every requirements lock the setup scripts install.

Every benchmark mix task that runs something is kept, including
imp.benchmark.parity.aggregate, imp.benchmark.gepa_replication,
imp.benchmark.gepa_merge, imp.benchmark.optimize_anything,
imp.benchmark.support_ticket_lift, imp.benchmark.provider_training,
imp.benchmark.instruction_optimizer_campaign, the two OpenRouter free-route canaries,
and imp_acp.demo_mcp_http_server. Missing documentation for a runnable task is a
documentation gap, not a reason to delete the task.

Mix tasks deleted

imp.evidence.admit, imp.evidence_authorities, imp.reproductions,
imp.research_portfolio, imp.upstream_fidelity — registry validators and
documentation generators for ledgers that no longer exist.

imp.benchmark.live_matrix — a coverage ledger over stored campaign artifacts, keyed on
the deleted benchmarks/model_availability.json; it calls no provider and measures
nothing. imp.benchmark.hotpotqa_analysis — a post-hoc classifier over stored parity
artifacts.

Bench modules deleted

Imp.EvidenceAuthorities, Imp.ReproductionRegistry, Imp.ResearchPortfolio,
Imp.UpstreamFidelity, Imp.BenchmarkTruth.EvidenceAdmission,
Imp.BenchmarkTruth.ReproductionArtifactValidator. Verified unreachable from every
surviving lib or mix task, bench module, script and test.

Imp.BenchmarkTruth.Paths loses admitted_root/0 and admitted/1.
Imp.BenchmarkTruth.OptimizeAnything.Campaign reads
benchmarks/authorities.json directly instead of through the deleted loader.
Imp.BenchmarkTruth.SupportTicketLiftCampaign no longer digests two deleted
predecessor result files while validating a preregistration manifest.
scripts/dspy_optimizer_public_workflow_gate.py defines its three-line
OperationalSafetyAbort locally instead of importing it from a deleted example dump —
that class is what the gate measures.

Tests: one line each

Deleted whole — bookkeeping (its only assertions are against a deleted validator,
registry or archive digest):

Test	Why
claims_inventory	validates benchmarks/claims.json against maintainer docs
evidence_admission	exercises the deleted admission module
evidence_authorities	exercises the deleted authority-ledger loader
evidence_provenance	asserts every admitted artifact's sha is an ancestor of origin/main and is cited by a ledger
reproduction_registry	validates benchmarks/reproductions.json
research_portfolio	validates benchmarks/research_portfolio.json
upstream_fidelity	exercises the deleted conformance-ledger generator
tutorial_ticket_routing_artifact	asserts a committed artifact is named by its own SHA-256 and exists

Deleted whole — behaviour lost because its subject is gone:

Test	Why
experiment_reference_graph	its subject was the six deleted matched directories, claims.json, and a sha256 migration ledger; scripts/experiment_reference_graph.exs goes with it
matched_gepa_mipro_ifbench	reads the deleted examples/matched_gepa_mipro_ifbench dumps
matched_ifbench_stop_accounting	requires stop_accounting.exs from two deleted example dirs
matched_instruction_family_ifbench_design	asserts a draft contract JSON in a deleted state
matched_instruction_optimizers_trec_example	reads the deleted TREC archive rows
case_study_trec_recomputation	same; the benchmark-surface branch moves those rows under examples/ and restores this test there
local_optimize_anything_retry_policy_three_seed_evidence	recomputes from a deleted archived manifest; its two sibling example tests still exercise the same runner

Restored after the first review — behaviour, kept with only the deleted-validator
assertions removed:

Test	What it exercises	What was dropped
instruction_optimizer_contract_artifact	runs imp.benchmark.instruction_optimizer_contract against pinned DSPy; 33 structural rows and the cross-runtime bootstrap fixture	the two ReproductionArtifactValidator calls and the tamper block
gepa_contract_artifact	GepaContract.compare against pinned GEPA v0.1.4; 15 required cases	the admitted envelope and its validator calls
support_ticket_lift_campaign (v2, v3, v3 full runner)	drives SupportTicketLiftCampaign.run through a local HTTP fixture for 48 requests and checks the response ledger; the four benchmarks/config preregistration files are restored	two digest-of-deleted-result-file assertions
optimize_anything_campaign	runs the campaign across seeds with fixture LMs, budget ceilings, cost guards	the mutation blocks whose only assertion was the deleted validator
optimize_anything_artifact	row validation plus the smoke and input CLI paths	the block reading a deleted archived artifact
optimize_anything_agent_config, optimize_anything_code_artifact, optimize_anything_scheduling_heuristic, optimize_anything_v1_4_readiness, circle_v1_4_cold_plan	evaluators, baseline-vs-comparator lift, held-out generalization, all provider-free	nothing
gepa_replication_artifact, gepa_merge_task, gepa_campaign	the replication and merge tasks against local fixtures	nothing
langprobe_heart_disease_product_fit	four-predictor program, pinned split, 147 grounded proposer calls, pinned DSPy product shape	nothing
musique_ans_product_fit, musique_ans_mipro_current	scorer agreement with the official Python scorer over 19,938 rows; real ReqLLM and DSPy wire rendering; MIPRO selection and fresh service	nothing
hover_papillon_calibration_pilot, hover_gepa_no_merge_plan	request serializers, route contracts, provider-disabled lifecycles with mutation, selection and fresh service, pinned DSPy loopback	nothing
gepa_study_plan	derives the study opportunity from the suite receipts	nothing
benchmark_paths	symlink containment and lane-escape rules on a surviving module	three Paths.admitted assertions
copro_isolation_artifact	runs the provider-free COPRO differential against pinned DSPy and validates the artifact the task emits	now captures a fresh artifact from a real task run instead of reading a deleted admitted file; the historical-ledger receipt block is dropped
authority_inventory	keeps benchmarks/authorities.json internally valid — the ledger the differential lane reads	two blocks reading claims.json and the deleted conformance doc

Trimmed, subject partly gone:

Test	Verdict
benchmark_truth	27 blocks removed, all of them live_matrix or hotpotqa_analysis, whose tasks are deleted. Nothing else was removed; the parity aggregator, fetcher, runner, integrity and scorer tests are untouched.
local_mlx_campaign	five blocks reading a committed admitted artifact removed — behaviour lost because its subject is gone. The acceptance, persistence and deployment tests remain and run fresh.
confidence_calibration_benchmark	one block reading a committed live result removed, same reason. The dataset preflight and provenance tests remain.
bfcl_adapted_artifact	one block removed whose every assertion was a deleted-validator call; the four scorer and mutation-corpus tests remain.
upstream_authority_registry	the UpstreamFidelity assertions removed; the registry fail-closed tests stay and a pin-resolution test replaces the deleted one.
documentation_contract	two blocks reading deleted maintainer docs.
operations_stress	one assertion that the run stays out of the deleted registries.
package_contract	exclusion-list entries naming deleted files.

Consistency

mix.exs loses only the aliases whose tasks are gone (upstream_fidelity.check,
reproduction.check, research.portfolio.check, benchmark.live_matrix,
benchmark.hotpotqa_analysis) and their preferred_envs entries. bench stays in
elixirc_paths for dev and test. .github/workflows/ci.yml needs no change: its jobs
run mix aliases, and every path in the paths-filter lists still exists.
.dialyzer_ignore.exs drops the entry for the deleted task and re-pins one line that
moved. .gitignore no longer ignores tracked files. test/test_helper.exs keeps the
:evidence_infrastructure exclusion — 40 tagged files remain. CONTRIBUTING.md,
decisions.md and docs/EVIDENCE.md lose their pointers to deleted paths.

Verification

• MIX_ENV=test mix compile --warnings-as-errors: clean.
• mix format --check-formatted: clean.
• mix quality.check (credo, mix deps.audit, mix hex.audit — the exact CI commands): exit 0.
• mix test on main: 54 doctests, 9 properties, 2998 tests, 0 failures, 11 skipped (171 excluded)
• mix test on this branch: 54 doctests, 9 properties, 2884 tests, 0 failures, 10 skipped (146 excluded)
• mix dialyzer: Total errors: 150, Skipped: 150, Unnecessary Skips: 0 — done (passed successfully).
• The :evidence_infrastructure suites outside the default run were exercised
  separately against a provisioned pinned environment: the COPRO isolation
  differential and artifact, the hover pilots, the GEPA contract, the BFCL scorer and
  the authority inventory all pass.
• CI on this PR, all ten jobs green: check, stranger.check, classify changes,
  package.check, integration.check, protocol.check, quality.check,
  dialyzer.check, differential.check (11m01s), docs.check.

Not in this change

docs/CASE_STUDY_TREC.md and docs/internal/ still reference deleted paths; the
benchmark-surface branch owns both.

Everything deleted here exists as a tarball on the owner's machine.