AGENT_GUIDE.md

Implementation guide for AI agents and researchers

Eve’s Embrace of Extreme Quantization v1.0 is a research compendium: reference implementations, finite formats, public synthetic examples, frozen numerical receipts, audits, and a manuscript. Begin with the repository card, runtime guide, and reproduction protocol. The research/ directory preserves the original release files; work in a separate directory before running any command that writes results.

The research question is whether a paid shared-fork quantizer can retain useful neural quality at extreme storage rates when a small, finite correction repairs selected terminal probabilities. The promising practical components are exact byte accounting, strict finite decoding, teacher-selected probability geometry, matched repair controls, and protected-incumbent selection. The release does not contain a complete inference checkpoint or demonstrate whole-model SOTA.

1. Know the experimental boundary

The native experiment uses facebook/opt-125m and changes only six projection weight matrices in decoder block zero: self_attn.q_proj, k_proj, v_proj, out_proj, fc1, and fc2. They contain 7,077,888 weights. Repair additionally changes 768 fc2.bias entries through a paid overlay. All other 66 decoder projection matrices and remaining parameters stay FP32.

The released bit rate divides complete first-block bundle bytes by those 7,077,888 projection weights. It is not a whole-model bit rate. The unchanged resident complement costs 472,645,632 logical FP32 bytes, including the original resident fc2 bias. The bias overlay is additional paid state. Whole-model logical accounting is this complement plus the complete bundle, without claiming that logical bytes equal file-container size or process memory.

Real-model quantized frames, fitted correction factors, calibration/test token arrays, teacher/student activation caches, upstream model files, and optimizer states are intentionally absent. Hash receipts document them; independent reproduction recreates them from separately acquired inputs. The 4,590-byte public EEB1 example has constructed zero base weights. Its rate demonstrates framing and decoding, not neural compression performance.

2. Run useful offline checks first

From the repository root, create a disposable sibling copy. The destination must not already exist:

python -c "from shutil import copytree; copytree('research', '../eve-v1-offline-demo')"
cd ../eve-v1-offline-demo
python -m pip install -r requirements.txt
python -m unittest tests.test_v1_reservoir_codec tests.test_v1_bundle_codec tests.test_v1_bundle_operator
python -m unittest tests.test_v1_terminal_gauge tests.test_v1_select_release
python examples/generate_v1_codec_demo.py
python v1_runtime.py inspect examples/v1_bundle_demo.eeb1 --resident-bias examples/v1_bundle_demo_bias.npy
python v1_runtime.py apply examples/v1_bundle_demo.eeb1 --input examples/v1_bundle_demo_input.npy --output demo_output.npy --resident-bias examples/v1_bundle_demo_bias.npy --include-bias
python -c "import numpy as np; np.testing.assert_array_equal(np.load('demo_output.npy', allow_pickle=False), np.load('examples/v1_bundle_demo_output.npy', allow_pickle=False)); print('PASS')"

These commands download no model or dataset and run no neural forward pass. The codec/operator subset has 33 tests; the geometry/selection subset has 24 tests. The generator returns bundle_bytes: 4590; the applied output has shape (1, 768) and every value is 0.125. These commands were checked on an isolated extraction while preparing this Hub kit. That packaging check is separate from the original 386-test research receipt.

v1_runtime.py defaults to model.decoder.layers.0.fc2. To apply another projection, use its full name with --projection and supply an input whose final dimension matches that projection. --include-bias applies only to fc2 with the explicitly supplied resident bias. Bare EER/EEL correction application returns a correction action, not a complete installed model projection.

Attachment-dependent theory checks

examples/check_v1_source_theory.py as a complete CLI requires the original Piercing Light V4 ZIP. It verifies the exact attachment SHA256 d3a2e3d1e862e3b7c9dbe6eac5fdd34147418084473fdb4b6135215623b7a70d; that attachment is not redistributed here. Do not replace it with another ZIP or interpret a missing attachment as a codec failure.

The independent algebra routines can still run without the attachment. This command writes no receipt and executes no attached code:

python -c "import random; from examples.check_v1_source_theory import gradient_audit, conditioning_audit, geometry_audit, overlap_audit, finite_factor_audit; r=random.Random(1729); a=gradient_audit(r); b=conditioning_audit(r); c=geometry_audit(r); d=overlap_audit(r); e=finite_factor_audit(r); print('PASS', a['exact_rational_gradient_bound_checks'], b['exact_shear_formula_numeric_checks'], c['KL_partition_and_gradient_cases'], d['exact_rational_tree_cases'], e['finite_signed_dyadic_native_word_checks'])"

The checked output is PASS 1440 256 128 64 448. Floating numerical checks supplement the stated algebra; they are not interval certificates for an actual neural model. If you separately have the exact attachment, use --attachment PATH --output NEW_OUTPUT_PATH on the full CLI in your disposable workspace.

3. Read the implementation in this order

Component	Starting points in research/	Responsibility
Base quantizers	v1_base_build.py, adaptive_fork.py, fork_sum.py, scalar_quant_frame.py	Build complete binary, shared-fork, local one-bit, and local two-bit projection controls.
Finite projection frames	fork_frame_container.py, numeric_view_container.py, v6_projection_operator.py	Self-describing base payloads, reversible views, identities, bounded reference decoding.
Terminal geometry	v1_terminal_gauge.py, v1_teacher_collect.py	Fixed teacher-selected IDs, direct masked tail logsumexp, normalized selected-plus-tail KL.
Repair families	v1_gauge_reserve_fit.py, v1_lowrank_gauge_fit.py	Sparse shared supports and matched finite low-rank factors, paid bias repair, frozen fitting protocols.
Reserve bytes	v1_reservoir_codec.py	Strict finite FP16 coefficients; sparse IDs or both factors; raw incumbent plus ten reversible compressed choices.
Complete installation	v1_bundle_codec.py	All six base frames, correction, bias overlay, directory, headers, checksums, native cast rules.
Bounded CPU action	v1_bundle_operator.py, v1_runtime.py	Tiled FP64 reference action of installed FP32 values; no packed GPU inference claim.
Selection	v1_select_release.py	Train admission, all-epoch catalog, complete-byte caps, stable development selection, bare incumbents.
Native evidence	v1_model_probe_release.py, examples/check_v1_release_evidence.py	Frozen profiles, configuration canonicalization, exclusive journals, native losses, aliases, independent audits.

Keep the cast contract explicit: decode the finite base to FP64, cast the base to FP32, add the finite reserve in FP64, then install FP32. For factor repair the product is formed from decoded FP16 factors in FP64. Adding correction to the uncast FP64 base changes the endpoint. Keep the original resident bias hash check and pay for its overlay. Finite FP16 signed zeros remain part of the exact word representation.

The bounded reference operator materializes at most one 64 × 64 weight tile at a time. Its FP64 accumulation differs from native FP32 BLAS. The reported language quality comes from dense FP32 installation into the normal model, not from this operator. Declared numerical scratch bounds are not measured peak RSS. Tiled/full factor products may differ at rounding boundaries; check installed words and actual native outputs for a new implementation.

4. Reproduce the neural experiment in a new workspace

Use Python 3.12 and the recorded environment in requirements-recovery.txt. Main pins are NumPy 2.3.5, SciPy 1.17.0, Torch 2.14.1+cpu, Transformers 5.19.0, Datasets 5.1.0, Hugging Face Hub 2.2.0, Safetensors 0.8.0, Tokenizers 0.23.3, and threadpoolctl 3.6.0. These are historical experiment identities; changing them creates a new configuration. The CPU Torch index must be available:

python -m pip install -r requirements-recovery.txt --extra-index-url https://download.pytorch.org/whl/cpu
python v1_rebuild_inputs.py --project-root /absolute/path/eve-v1-replica/working/embrace

Run the rebuild command from an untouched extracted research/ directory. Substitute a genuinely empty absolute destination, outside the release and its cache root. On Windows, quote an absolute path such as "C:/experiments/eve-v1-replica/working/embrace". The helper refuses nonempty/nested/original destinations, verifies downloaded assets and original projection hashes, copies public source, and preserves published v1 receipts under reference_results/original_v1. New receipts go to results. It downloads inputs but performs no fitting or neural forward pass.

Input	Revision
facebook/opt-125m model and tokenizer	27dcfa74d334bc871f3234de431e71c6eeba5dd6
Salesforce/wikitext, wikitext-2-raw-v1	b08601e04326c79dfdd32d625aee71d232d685c3

Use the printed model_dir from the recovery receipt. Follow the complete ordered commands in V1_REPRODUCTION.md: collect teacher evidence; build/decode four bases; freeze selection; freeze/run sparse repair; freeze/run factor repair; score the bare one-bit control; audit calibration; propose/finalize roles; prepare tokens; freeze/evaluate the release probe exactly once; audit saved native results. Pass each printed protocol digest exactly as requested.

The terminal objective fixes teacher top-32 IDs plus one direct complement category. All vocabulary logits contribute to the normalizer; tail mass must not be computed as 1 - selected_mass. Bases use all 64 calibration contexts. Repair uses the first 48 contexts for fitting and the last 16 for development selection; that development split was already seen by the base fits. Four Adam epochs produce five finite checkpoints per configuration, including epoch zero. Compile, strictly decode, install, and score the finite endpoint rather than treating the shadow optimizer values as the released model.

The catalog contains 60 repaired checkpoints and four bare incumbents. Admit repair checkpoints only when native train KL is no worse than their own epoch-zero train KL. Minimize development KL under actual complete bare two-bit/one-bit bundle caps, then bytes, then stable name. Global primary selects bare local two-bit; global extreme selects hybrid factor rank 3 epoch 4. Source-specific roles remain separately disclosed. Incumbent protection guarantees this finite calibration selection property, not superior fresh perplexity or optimization over all possible quantizers.

Existing journals are evidence. Do not delete a start marker to repeat evaluation, overwrite frozen source, or copy historical failure receipts into a clean replica. Preserve a failed replication and begin another explicitly identified directory. The original test window test[81920:90112) is now public replication data; it cannot become untouched evidence again. A new scientific study needs newly reserved evaluation data before fitting or method selection. Final packaging also needs a newly authored, rendered, visually checked report; copying the original PDF QA receipt does not validate a new manuscript.

5. Extend the method with measurable milestones

First establish exact finite reconstruction and byte accounting for your target architecture. Then evaluate all relevant layers with the same matched sparse/factor repair controls and protected bare candidates. Report complete deployed bytes, including embeddings, auxiliary state, scales, supports, factors, routing, exceptions, alignment, bias overlays, and container overhead. Recompute whole-model denominators explicitly.

Use official baseline implementations and their documented recipes. The supplied scalar controls are independent local GPTQ-style implementations; they are not official GPTQ software or an industry-wide reference. Match calibration data, access to optimization, and encoder compute when attributing gains. Separate base-quantizer novelty from repair gains using repaired binary and scalar controls. The released fork-versus-matched-binary descriptive intervals include one; this window does not establish a reliable fork-specific advantage.

For quality, add standard full-corpus perplexity, downstream tasks, generation behavior, multiple seeds and model families, and independently reserved data. The current reset-128, 8,128-prediction diagnostic is not the standard full WikiText-2 benchmark. Selected-plus-tail KL omits conditional redistribution inside the tail and cannot certify full KL, next-token NLL, or autoregressive generation quality. Uniform neural reserve invertibility and protected branching-tree enclosures are unverified.

For deployment, implement packed GPU kernels only after fixing decoder and installation semantics. Compare latency, throughput, peak memory, and numerical quality across batch/context sizes on declared hardware. None of those GPU measurements exist in v1.0. Preserve failures, source hashes, protocols, and independent receipts as new evidence. A faster kernel, broader successful study, or smaller exact finite representation would be a useful contribution; claims of SOTA must follow the completed comparison, not precede it.