README · 616484d0caaa

EDCM

EDCM is the Energy–Dissonance Circuit Model measurement and evaluation repository.

The active package is 0.2.0, candidate edcm-measurement-v2. It preserves the frozen v1 canon data and metric equations while repairing transcript intake, result provenance and codec fidelity. Result schema 2.0.0 hashes every emitted compartment; codec 2 retains source text and token canon records. See the migration and repair record.

The repository separates current measurement, candidate experiments, and exact historical corpus/construction replay. Active language construction belongs to Stack and consumes UCNS constructors. No package, representation or passing test establishes empirical validity or selects canon.

UCNS–EDCM canon

The joint canon is reciprocal rather than inherited:

EDCM proposes observable distinctions and falsifiable readouts.
UCNS proposes structural policies and candidate instruments.
Experiments test both directions.
A separate evidence-bearing decision may later select canon.

Neither side transfers status automatically:

UCNS proof status -> EDCM empirical validity: false
EDCM empirical fit -> UCNS proof status: false
candidate registration -> canon: false
passing development fixtures -> canon: false
NA -> 0: false

See CANON.md and docs/UCNS_EDCM_EXPERIMENT_PROGRAM.md.

The complete compiled mathematics currently declared or implemented by EDCM is indexed in docs/EDCM_MATHEMATICS.md. The reference separates the frozen baseline candidate, signed-axis construction, implemented v0.3.1 architecture layer, joint experiment candidates, controlled goal-vector candidate, and unresolved formal boundaries. A textbook or website may carry a source-pinned copy but is not the mathematical authority.

Gonol language boundary

The current authority split is:

For new language construction, start in Stack English or Stack Python, consuming UCNS constructors and geometry. These source links pin the authority comparison used for this repair; resolve current owning commits before new research.

Historical edcm.gonol, edcm.language, lexical-floor and sealed profile artifacts remain explicit replay surfaces. They confer no active construction authority on EDCM. There is no universal mandatory adjacent-scale ladder. Construction reproducibility does not validate EDCM measurement; an unresolved UCNS operation remains hmmm.

See the language boundary and the 0.2 migration.

First joint experiment

The initial fixed corpus pressures four load-bearing distinctions:

• order;
• multiplicity;
• constraint pressure;
• resolution timing.

It compares:

• the maintained EDCM baseline;
• a transparent sequence-sensitive EDCM candidate;
• unit-, token-, and pressure-weighted UCNS cell encodings;
• UCNS ordered, multiset, and set structural views;
• noncanonical UCNS product-character and faithful-breadth candidate families.

Here token names a frozen support unit in that historical experiment. It must not be read forward as a tokenizer layer in the current gonol language path.

The experiment records supported, falsified, and errored hypotheses. A falsified hypothesis is valid research evidence and does not fail the build merely because a preferred candidate lost.

Run locally

The base package remains dependency-free. Joint experiments are opt-in and pin the exact reviewed UCNS commit.

python -m pip install -e .[dev,ucns-experiments]
python -m pytest -q tests/test_ucns_edcm_experiments.py
python -m edcm.ucns_edcm_experiments \
  --ucns-source-root /path/to/ucns-checkout \
  --output artifacts/ucns-edcm-report.json

The dedicated workflow checks out UCNS at:

The-Interdependency/ucns@5331ae9a4cf7eddfa1de72b8caed28e2358cc0ed

It runs the experiment twice and requires byte-identical reports before uploading the evidence artifact.

Evidence states

The following remain distinct:

1. represented evidence;
2. constructed representation;
3. candidate-measured evidence;
4. experiment-supported evidence;
5. canonically measured evidence.

The repository currently supports the first four. The experiment report always contains:

canon_selection = null

First controlled goal-vector measurement

The historical exact UCNS observation profile feeds one bounded EDCM candidate-measurement experiment. Two dialogues contain the same four exact utterance occurrences in different orders: one resolves an availability contradiction and one leaves it active. Every source word, turn, declared fixture relation, component state, contradiction, and scalar information-loss record remains inspectable.

The sealed 2026-08-02 report supports all eight declared findings. The resolved order ends at candidate projection 1/1, motion variance 1/8, and zero active contradictions; the reordered case ends at projection 1/2, motion variance 9/64, and one active contradiction. The cases share one occurrence-multiset identity but have different ordered and UCNS-observation identities.

python -m edcm.goal_vector_experiment \
  --ucns-source-root /path/to/ucns-at-a98c9e6c69804a8a08d0786b1d8b450bb2c49a97 \
  --output /tmp/goal-vector.json

This is controlled candidate-measured evidence, not empirical validation. METAPAT constraints, formal UCNS geometry, formal completion, closed floor-definition-gonol construction, proof transfer, and canon selection were not part of this sealed experiment and are not retroactively supplied by the later language-boundary correction.

First externally labelled booking-outcome holdout

The sealed MultiWOZ 2.1 booking-outcome experiment evaluates the maintained terminal progress proxy before source-annotated Booking-Book and Booking-NoBook responses. Development alone fits calibration, validation alone selects the threshold, and the 661-event test partition is evaluated after that freeze. The labelled response and all later turns remain outside candidate input.

The test produced TP=249, FP=56, FN=281, and TN=75: sensitivity 0.4698, specificity 0.5725, and balanced accuracy 0.5212. The 95% dialogue-cluster interval for balanced accuracy is 0.4656–0.5739, so the small above-chance point estimate is weak evidence. The declared sensitivity hypothesis is falsified and remains in the report.

• frozen design
• findings
• aggregate report
• completion receipt

python -m edcm.corpora.multiwoz21_booking_holdout \
  --archive /path/to/MULTIWOZ2.1.zip \
  --edcm-repository-root /path/to/edcm \
  --edcm-commit "$(git -C /path/to/edcm rev-parse HEAD)" \
  --output /tmp/multiwoz-booking-holdout.json \
  --receipt /tmp/multiwoz-booking-holdout-complete.json

The historical MultiWOZ profile seals remain represented evidence. The new derived report is bounded candidate-measured holdout evidence; it does not retroactively turn the profile observation into formal geometry or empirical validation. canon_selection remains null, and EDCM/METAPAT activation remains inactive.

Maintained measurement candidate

edcm/measurement/ consolidates the reviewed structural-measurement lineage from:

• The-Interdependency/edcmbone; and
• the earlier erinepshovel-code/EDCM application lineage.

It contains:

• frozen measurement canon data;
• deterministic transcript parsing;
• the eleven-component round vector;
• circuit-state recurrence;
• projection and risk surfaces;
• provenance and integrity gates.

The v1 implementation remains frozen at its historical commits; 0.2 uses candidate v2. Its maintained identity is useful as a baseline candidate. Integrity means the implementation did not drift; it does not mean the readouts are empirically validated.

Install and validate the baseline

EDCM requires Python 3.11 or newer.

python -m pip install -e .[dev]
python -m edcm.integrity
python -m pytest -q
python tools/check_metadata_contracts.py
python -m build
python -m twine check dist/*

METAPAT integration remains optional:

python -m pip install -e .[dev,metapat]

The exact historical/replay word-gonol profile is also optional:

python -m pip install -e .[dev,ucns-profile]

full-stack installs both the exact METAPAT producer and the historical EDCM UCNS profile producer. Use ucns-experiments only for the historical v0.1–v0.4 experiment epoch. Package availability alone attaches no evidence: UCNS activation also requires checkout package bytes that match the exact Git tree or a VCS installation whose package bytes match EDCM’s pinned producer manifest and raw wheel RECORD; the verified module graph is freshly loaded, and every runtime-loadable cache must derive from those verified sources. Exact ordered ucns_turns are required before a historical observation is attached.

First real-system corpus runner

MultiWOZ 2.1 is the first admitted historical real-system source. Its runner verifies the exact University of Cambridge archive and every logical member, streams all 10,438 dialogues in source order, observes every exact speaker turn through the pinned EDCM UCNS word-gonol profile, and independently repeats the exact turn stream through the UCNS v0.14.1 full-corpus completion gate from the merged v0.19 producer commit with final integrity repairs. Completion requires both passes and their source-native turn chains to reconcile. Only aggregate evidence and completion or incompletion receipts are emitted. Raw corpus bytes remain outside Git. Carrier SPACE metrics are derived from profile assignment at alphabet position zero; the exact source code point remains independently serialized.

python edcm/corpora/run_multiwoz21_seal.py \
  --archive /path/to/MULTIWOZ2.1.zip \
  --ucns-source-root /path/to/ucns-at-a98c9e6c69804a8a08d0786b1d8b450bb2c49a97 \
  --output /tmp/multiwoz-2.1-ucns-v019.json \
  --receipt /tmp/multiwoz-2.1-ucns-v019-receipt.json \
  --checkpoint /tmp/multiwoz-2.1-ucns-v019.checkpoint.json

See docs/corpora/multiwoz-2.1.md. This is represented evidence, not an EDCM candidate measurement, formal UCNS geometry, or a canon selection. The original profile-0.1 report remains immutable but is explicitly superseded because tabs, newlines, and non-breaking spaces were misclassified as out-of-alphabet instead of SPACE manifestations. The sealed profile-0.2 replacement assigns all 4,094 occurrences to carrier position zero, reports 1,783,679 SPACE boundaries and no carrier-unassigned source code points, and preserves the original source and turn digest chains. That historical receipt remains immutable; it is not a UCNS v0.14.1 completion receipt. The authenticated v0.14.1 rerun is now independently sealed in its own aggregate report and completion receipt. It consumed all 143,048 turns, reconciled the repeated source-native chain, and left EDCM and METAPAT activation inactive.

The exact reviewed UCNS v0.19 producer was then repinned and rerun without changing the consumed profile or gate schema. Its sealed report and completion receipt cover all 10,438 dialogues and 143,048 turns. A completed-checkpoint repeat was byte-identical. The exact stream hash and UCNS v0.14.1 receipt id remain unchanged; the report and receipt identities bind the EDCM package tree and the UCNS producer commit. Measurement remains not-run, canon selection is null, and EDCM/METAPAT stay inactive.

Merged UCNS v0.19 commit a98c9e6c69804a8a08d0786b1d8b450bb2c49a97 was then freshly resealed in new immutable aggregate report and completion receipt. The completed-checkpoint repeat was byte-identical, all execution counts and source-native chains reconciled, and the exact stream hash plus UCNS receipt id remained unchanged. Measurement remains not-run, canon selection is null, and EDCM/METAPAT stay inactive.

Integrity gate

python -m edcm.integrity checks the frozen baseline:

• complete and exact *_v1.json canon bytes;
• measurement authority and compatibility policy;
• orthogonality-class no-fork identity;
• source and wheel behavior.

A legitimate baseline-canon change requires a new versioned file and migration record. Do not update identities merely to silence continuous integration.

Provenance-bearing pipeline

The active shared-stack boundary separates:

• exact source evidence and the declared measurement input profile;
• METAPAT semantic/affixiation authority;
• externally constructed gonol identities and receipts where a declared evaluator consumes them;
• UCNS carrier/geometry identity and unresolved geometry;
• exact pinned historical UCNS word-gonol observations where replay profiles are used;
• typed UCNS geometry and factorization absence in historical adapters;
• EDCM measurement policy and implementation provenance;
• readouts;
• status evidence.

The retired ordered-occurrence bridge, UCNSObject, and factorization inputs fail closed on the current runtime path. Historical experiment reports keep their original UCNS commit and are not rewritten as current profile evidence.

The pre-reset edcm.ucns_metrics resolver and its top-level exports are removed. They depended on archived UCNSObject, recursive_encode, and stable_hash surfaces rather than the exact historical profile. Use ucns_profile_observation when reproducing represented historical evidence; any future scalar projection must remain linked to its complete trajectory and declare information loss. See docs/ucns-metric-objects.md for migration guidance.

Typed absence

NA != 0 remains non-negotiable.

Unavailable evidence is typed absence. A candidate may fail scope, return an explicit error, or remain unmeasured; it may not invent a neutral measurement.

Proof and measurement firewall

No UCNS or METAPAT status validates EDCM readouts, external truth, diagnosis, intention, morality, or consciousness.

A completed gonol construction likewise does not establish measurement validity.

Repository provenance

Historical implementation, source packets, repair handoffs, and prior adapter contracts remain preserved in Git history and under archive/ or codex-handoff/ where already present. They remain evidence, not automatic current canon.

Usage guidance

Use edcm.build_default_layers().run({"transcript": text}) with raw input fields. Output-only fields are rejected; submit the original input rather than a previous result. The parser retains continuation lines and mixed recognized label formats; parsed.source_text retains exact input bytes after UTF-8 encoding. A line matching a speaker-label grammar starts a new turn; disambiguate label-like prose before measurement.

For active construction use the Stack workspaces above. For historical replay use both the exact EDCM and UCNS producer commits named by the artifact. New executions with 0.2 are new evidence, not byte-identical historical reruns.

See upgrade avenues for the next candidate and usability decisions.

hmmm

Language admission is owned by its Stack workspace. The source-supported complete English morphology law, exact UCNS native Möbius-carrier affixiation/coupling geometry, and direct distant-scale coupling remain unresolved. EDCM has not selected a lawful projection, information-loss account, metric, benchmark, or falsifier for evaluating completed recursive text-gonol constructions. Historical MultiWOZ seals remain represented evidence, not measurement validity; external replication and broader carrier coverage remain separate unresolveds.

Open exact README source