Scientific record
Historical v0.1.0 checks — 2026-10-11
Runtime: Python 3.12.14 on Windows; standard library only. Commands ran from a recovered staging copy of the package. Raw staging results (not included in this public selection: verification/staging-results.json) retain output and exit codes.
| Check | Observed result |
|---|---|
python -B -m unittest discover -s tests -v | 10 test methods passed; includes 150 inventory/action-limit/cost-budget variants of one fixed token recipe system and all 5,370 bounded action sequences. |
python -B verify_showcases.py | Eight fixed showcase checks passed, including both token witnesses and both schedules. This is fixed-witness sanity checking, not validation of seven additional full graders. |
python -B harness.py inspect fixtures/token-workshop-001.json | Strict fixture validation succeeded. |
python -B harness.py solve fixtures/token-workshop-001.json | Independent transition/search implementation found a four-action solution of cost 4; six nodes expanded. |
python -B harness.py grade fixtures/token-workshop-001.json fixtures/private/token-workshop-001.witness.json | PASS; exact final inventory, four actions, cost 4. |
The tests cover strict parsing, duplicate keys, non-JSON constants, response/world schemas, resource legality, budgets and intermediate invariants, acceptance of alternative valid solutions, renaming/order invariance, search limits and zero-cost actions. Two regression tests check reported token costs under changed costs and refusal of unsupported showcase semantics.
Local installation and records
The same five commands were executed from the installed local Hermes package using Python 3.12.14. All completed with exit code 0: 10 test methods passed, all eight showcase witness checks passed, fixture inspection succeeded, the solver found a valid four-action/cost-4 solution, and the retained witness graded PASS. Raw installed results (not included in this public selection: verification/local-results.json) record the exact commands, working directory and observed outputs.
All 17 prior canonical record bytes matched the existing SHA-256 snapshot manifest before and after the save. The archived root README remains intact; the snapshot overview uses SNAPSHOT.md to avoid a name collision. All 246 planned local file links resolved against the prepared integration; final installed file/link/hash verification is recorded in the file manifest (not included in this public selection: verification/file-manifest.json).
The recovery retained all 40 drafts, saved their package/record integration through scoped filesystem grants and normal file tools, corrected the showcase cost report and unsupported-semantics handling, and added two regression tests. No unrestricted terminal-input retries were used during recovery.
Evidential limits
This is an offline engineering result. Frontier-model calls: 0; paid calls: 0. No model deficit or human/process/reward result is inferred. The oracle shares fixture validation with the grader, so its transition/search independence is narrower than a separately formalized independent specification. Exhaustive checking here concerns bounded variants of one recipe system, not arbitrary generated worlds.
At the v0.1.0 freeze, the whole-solution-space necessary-motif certificate and matched F/C demonstration/solver audit were unimplemented. The v0.1.1 extension below supersedes that status for its supported public token scope. Private generators, seven full family graders, powered confirmation and provider/budget freeze remain pending. The scheduling showcase replays valid schedules; it does not independently prove the recorded optimum.
2026-10-12 — Token audit extension v0.1.1
Raw audit result (not included in this public selection: verification/token-pair-audit-v0.1.1.json) and new test transcript (not included in this public selection: verification/token-pair-tests-v0.1.1.json) record Python 3.12.14, exact commands and exit code 0. Only the new audit/tests were executed; historical completed suites were not rerun.
- Twelve new test methods pass, including false pruning, incorrect non-goal state/cost, partial failure state, metadata/role corruption, preloaded targets, incomplete enumeration, renaming and a copying-solvable relaxed-buffer mutant.
- TW-001: all 121 grammatical sequences within four actions replayed; exactly two valid plans, optimum length/cost 4; both remain accepted.
- Imported public F/C targets: each grammar has 87,381 sequences. A disjoint illegal-prefix cover certifies all of it using 77 full-state comparisons (27 legal prefixes, 50 rejected next steps), retaining both valid eight-action/cost-8 solutions. This is not 87,381 direct grader calls.
- Every valid target plan requires the reviewed type-level make/coat/join dependency. No individual fungible-token ancestry or internal innovation event is certified.
- Each source demo is legal with a verified six-state trace. Source F/C have 2/13 valid plans, 10/52 legal prefixes and 23/69 rejected extensions; this is an observed matching limitation.
- Each copying audit replays 912 declared subsequence/repeated-block/bijective-rename candidates without an accepted adaptation. Arbitrary reordering/mixed fragments remain allowed.
- Primitive evidence matches under the explicit role maps: 14 examples each, 4 legal/10 invalid. Evidence is 6,485 serialized UTF-8 bytes per condition; complete standalone inputs are 7,447 bytes each. Provider tokens are unmeasured.
The earlier oracle/checker comparison checked only PASS/goal on legal prefixes and NOT-PASS on rejected extensions. NOT-PASS also describes legal non-goal prefixes. The new audit requires precise failure classification and committed-state/cost/count agreement before pruning. Existing candidate certificates remain consistent under this stronger check; the assurance gap is recorded without rewriting historical test reports.
Prior state is byte-recoverable from the 17-record history capsule (not included in this public selection: verification/prior-records-v0.1.0.json); compressed JSON contains original bytes and SHA-256 values. Creation of a new archive folder was denied, so this normal-file-tool artifact resides in the writable package verification directory. Current integrity record (not included in this public selection: verification/token-pair-file-manifest-v0.1.1.json) checks it, existing snapshots and input rendering.
Original checker/oracle/fixture and all EXP-0003 source/evidence files are retained. Model/human/paid runs remain zero. The current prototype covers one supported contract and an imported public related base. Source matching, order/counterbalance, provider/P/R/E/compute/criteria/spend freeze, private generators and confirmation remain unresolved.