Engram

Engram is an offline Tiny Expert board proof for extreme edge constraints.

This repository is not a cloud-LLM replacement and not a public release of the full training stack. It is a task-specialized, table-driven, auditable edge reasoning line that shows how benchmark capability can be crystallized into a very small, flash-resident runtime on a commodity ESP32-C3.

The repo now publishes two layers on purpose:

• Public board proof: a real ESP32-C3 fixed-batch compiled run for LogiQA 642
• Research capability line: the current audited host-side scientific surface with official, external, runtime, and overfitting status

Quick facts

Item	Value
Target board	ESP32-C3
Public board proof mode	logiqa_batch_compiled_probe_aggregated
Compiled probe mode	host_full_exact
Public board proof	249 / 642 = 0.3878504672897196
Board proof host alignment	host_full_match = 642 / 642
Runtime artifact	chip_suite_surface_round2_q4_tgz_expack_rgz.json
Artifact size	1,380,771 bytes
Default firmware size	1,784,352 bytes
Research line	official IFEval = 0.780037, official LogiQA = 0.392523
External guard	external_dev = 0.308908, external_blind = 0.425072
Trust boundary	clean holdout2 = 0.400000, hidden-family 0 / 85

Why this matters

Many edge deployments do not need a general cloud assistant. They need a tiny, deterministic, offline decision system that fits severe memory and power budgets and can still be audited when something goes wrong.

This repository is a public proof of that direction:

• Offline: no cloud inference is required for the published board proof
• Table-driven: the runtime is flash-resident and structured, not a dense open-input LLM stack
• Auditable: board reports, benchmark references, integrity checks, and overfitting evidence are all published
• Edge-constrained: the target is a commodity ESP32-C3, not a GPU server

What this repo proves

1. Public board proof

The board proof in this repository is a real ESP32-C3 execution line.

• Board summary: results/board_proof/esp32c3_logiqa642_board_proof_summary.json
• Board acceptance: results/board_proof/board_runtime_audit_acceptance.json
• Raw board batches: results/board_proof/raw

The published board result is:

• LogiQA 642 = 249 / 642 = 0.3878504672897196
• host_full_match = 642

That means the board-side compiled path exactly reproduces the host full decisions for the published fixed batches.

2. Research capability line

The current audited host-side scientific surface is published separately:

• results/research_line/current_scientific_surface_manifest.json
• results/research_line/current_scientific_surface_reference_status.json
• results/research_line/current_scientific_surface_guard_bundle_status.json

Current reference metrics:

• official IFEval = 0.780037
• official LogiQA = 0.392523
• external_dev = 0.308908
• external_blind = 0.425072

3. Trust and audit layer

The repository also ships the public evidence layer for:

• replay and integrity
• runtime and shadow gates
• overlap and external non-regression
• post-hoc clean holdout audit
• hidden-family forensic audit

See:

• docs/TRUST_AND_AUDIT.md
• results/audit/current_scientific_surface_overfit_audit_status.json
• results/audit/current_scientific_surface_hidden_family_forensic_audit_status.json

Hard boundaries

This repository does not claim:

• unrestricted open-input native LLM inference on ESP32-C3
• public release of the full training pipeline
• public release of the full materialization and candidate-generation workflow
• proof of general reasoning beyond the audited task surface

The current board line is:

• a fixed-batch compiled board proof
• aligned to host full decisions
• auditable through raw board reports and published summaries

That boundary matters. The point of this repository is to show a real, reproducible, edge-constrained Tiny Expert line without pretending it is a general-purpose on-device chatbot.

Quick start

See docs/REPRODUCE.md for the full walkthrough.

Shortest path:

1. Flash the published firmware:

py scripts/flash_firmware.py COM3

2. Read the board report back:

py scripts/read_board_report.py COM3 --expect-mode logiqa_batch_compiled_probe --expect-artifact-sha256 626a1bfcc0a86585db82130744094ee4512eaaead8b4d9f1dba07175c010719d

3. Compare the returned JSON against:

• results/board_proof/raw/esp32c3_logiqa_batch_current_surface_runtime_topk_official642_k24_hostfullexact_000_064.json
• results/board_proof/esp32c3_logiqa642_board_proof_summary.json

Important scope note:

• the default firmware reproduces a published fixed batch board report
• the full 642 score is the audited aggregate over the included 11 board-side raw reports

Repository layout

• firmware
  • the published ESP32-C3 board-proof binaries
• results/board_proof
  • the current board-proof summary, acceptance, and raw batch reports
• results/research_line
  • the current scientific surface manifest and replay status
• results/audit
  • the current overfitting and hidden-family forensic evidence
• docs
  • method boundary, audit notes, capability framing, and reproduction notes
• scripts
  • flashing and report readback helpers

Further reading

• docs/WHY_TINY_EXPERTS.md
• docs/CAPABILITY_MATRIX.md
• docs/BOARD_METRICS.md
• docs/TRUST_AND_AUDIT.md
• docs/RESEARCH_LINE.md
• docs/METHOD.md
• docs/FAQ.md