Skip to content

Repository files navigation

EDEN

EDEN measures what intelligence makes unnecessary. This repository is a single-node research prototype of a protocol that is intended to become a network. Today it is one Mac and a SQLite file — the claims below are scoped accordingly.

Bitcoin proves work by burning energy. EDEN goes the opposite direction:

Bitcoin:  Energy → Computation → Proof → Money
EDEN:     Task → Intelligence → Less Resource → Proof → Value

EDEN mints on observed dominance, not counterfactual savings. Receipts record observed facts. The Pareto frontier of verified results is the reference. New value is issued only when a record is actually broken — doing ordinary work moves existing CREDIT; pushing the efficiency frontier is the only thing that creates new CREDIT (1 CREDIT ≡ 1 certified joule is a v0 placeholder; cross-family valuation is an open problem in the spec).

Constitution

  1. Observation Before Prediction — predictions never justify issuance; minting is based on observations only.
  2. Result Before Efficiency — no efficiency has value until the result is proven correct.
  3. Net Efficiency Only — an improvement smaller than its own measurement, verification, and audit cost is not an improvement.
  4. Facts Outlive Rules — receipts are immutable observations; frontier and minting rules are versioned interpretations that may change, but must never break recomputability from past facts.

v0 pipeline

task → run → measure → verify → receipt → frontier (read-only SELECT) → mint (--commit, simulated)

Single-file CLI (eden.py), Python stdlib only, SQLite ledger. Self-tests:

python3 -m unittest discover tests

Requirements

  • macOS or Linux (uses the Unix-only resource module; Windows is unsupported)
  • Python ≥ 3.9, stdlib only
  • Level V meter: macOS only (powermetrics; sudoers grant below)
  • LLM runner: ollama; default model qwen2.5:7b (~4.7 GB, pulled automatically on first use)

Quickstart

python3 eden.py demo                     # top-k word-frequency family
python3 eden.py demo --scenario codefix  # code-fix family

Measured on one machine (MacBook Pro M5 Pro, 2026-08-16 → 08-18 JST)

Sample sizes and meters are stated everywhere, because they change the story. Real receipts from this ledger are committed under examples/.

Level S — estimated-cpu-v1 (child cpu-seconds × an assumed 6.0 W; highly repeatable, systematically biased, blind to GPU/ANE):

naive_count    n=10   5.680 J   σ 0.037 J (cv 0.7%)   # cpu-time repeatability, NOT energy accuracy
dict_loop      n=5    0.148 J
counter_fast   n=6    0.141 J
codefix_rules  n=5    0.343 J   ← code-fix frontier holder
codefix_brute  n=5    1.538 J
codefix_llm    n=3    2.439 J   # labeled lower bound in receipts (GPU invisible)

Level V — powermetrics-package-v1 (CPU+GPU+ANE package power minus a measured idle baseline):

naive_count    n=10   mean 8.96 J    σ 3.52 J (cv 39%)
codefix_llm    n=5    mean 93.7 J    σ 53.0 J (cv 57%)   # + LLM output-length nondeterminism

What we can honestly claim from this ledger:

  • Same meter, same verified result: the targeted rule-based fix (0.343 J) beats brute-force mutation search (1.538 J) by ~4.5× — tight intervals, certified dominance, one simulated mint of +0.998 CREDIT net of verification energy. This is not "rules beat LLMs": the rule runner carries the bug class as prior knowledge. What the ledger prices is the joule value of knowing something — here, knowing the fix pattern is worth ~1.2 J per task against blind search.
  • The LLM's real cost lives on the package meter: ~51–159 J per verified fix (n=5, cv 57%), against a cpu-time-only apparent cost of ~2.4 J. Cpu-time metering misses roughly 20–60× of the energy (GPU/ANE inference) — a range, not a constant. All 8 LLM attempts passed the 12-test suite.
  • Cross-meter ratios are not publishable. Earlier revisions of this README claimed "200–450×" (LLM vs rules) and a "~9.3 W" calibration constant; an adversarial audit showed the former divided numbers from different meters and the latter was the minimum of n=3. Extending calibration to n=10 on a busy machine exploded cv to 39% — the "exclusive use" boundary condition is not decoration; without it, package-level measurement is noise. σ really is EDEN's first research result.
  • A certified frontier update once minted 0 CREDIT because the gain was smaller than verification energy (Constitution III), and a cheap-but-wrong runner got no receipt (Constitution II). Both under Level S.

First challenge-sampled results (v0.2)

Frontiers must be "expected J over a distribution the protocol issues", not "minimum J over instances you chose". eden challenge open pins enrolled runner code hashes before generating instances (seeded bug injection into a correct module; the epoch seed is recomputable from the ledger). First epoch — 6 protocol-issued single-token bugs × 3 strategies:

codefix_brute  6/6 (100%)   8.4 J total   →  1.41 J/success   [estimated]
codefix_rules  0/6 (  0%)   0.9 J total   →  ∞                [estimated]
codefix_llm    1/6 ( 17%)  893.9 J total  →  894 J/success    [Level V]

Three things this measured:

  • Cherry-picking was real. On the hand-picked bug earlier in this README the LLM fixed 8/8; on the issued distribution it fixed 1/6. Self-chosen results do not predict distribution results — visible only after challenge sampling existed.
  • Failures are billed (the §8 metric works): five failed LLM attempts burned 130–165 J each, so the true cost per verified success is ~894 J.
  • Specialist knowledge dies off-distribution: the rule-based runner went 0/6 (giving up honestly at 0.15 J per attempt).

Declared bias: the injected-bug vocabulary is a subset of the brute-force searcher's mutation vocabulary, so this distribution structurally favors search. The honest claim is "search dominates single-token bug distributions", not "search beats LLMs". Harness validated: one failed instance passed on manual replay (stochastic), two failed again (real).

v0.3 — the frontier's input unit is now the distribution. An epoch folds into a Distribution Certificate (success rate with a Wilson 95% interval, total joules with verification inside the cost, J/success), and the record is a Pareto frontier over (success rate ↑, J/success ↓) per meter — a 100%/20J runner, a 95%/10J runner and a 60%/2J runner can all hold the record at once; a 96%/8J newcomer dominates only the middle one. Minting happens at certificate registration as a pure function of ledger order — a runner that solves one instance miraculously cheaply cannot escape the cost of its other five failures:

★ brute  rate 100% [61,100]   1.59 J/success   ← frontier (estimated meter)
  rules  rate   0% [ 0, 39]   ∞                (dominated)
★ llm    rate  17% [ 3, 56]   894 J/success    ← frontier (Level V meter, genesis)

The §9 experiment: intelligent frugality exists (v0.3, epoch cc846674)

Seven strategies, one protocol-issued distribution (12 instances, 8 distinct single-token bugs), all failures billed, verification energy inside the cost, every LLM measured on the package meter (Level V):

strategy                     success   J/success (verify incl.)
cascade (rule→search→1.5b→7b) 12/12      1.37   ← sole Level V frontier holder
brute search                  12/12      1.53   (estimated meter)
phi4 14b                      11/12    287
qwen2.5:1.5b                   6/12     81
qwen2.5:7b                     5/12    391
qwen2.5:7b × 3 retries         5/12    404
targeted rule                  0/12      ∞

What this measured: escalating-cheapest-first wins outright on this distribution (100% success at 1/210th of phi4's cost per success — the original spec's "intelligent frugality" hypothesis, answered); the small model is 3.5–4.8× cheaper per verified success than its bigger siblings and Pareto-dominates the 7b outright; and retries buy nothing against systematic failures — the 7b fails the same way every time, so three attempts cost 3× for the same 5/12.

Revised with bootstrap intervals (2026-08-19). Resampling the runs instead of trusting point estimates sharpened which of these claims the evidence actually carries:

cascade  1.37 J/success  ci95 [0.78, 2.07]  ⎫ overlapping — the evidence
brute    1.53 J/success  ci95 [1.20, 1.87]  ⎭ does not separate these two
1.5b       81 J/success  ci95 [54.7, 162.5] ⎫ disjoint — this domination
7b        392 J/success  ci95 [213, 1060]   ⎭ is real

So "the cascade beat brute-force search" was over-read from twelve runs; "the small model beats its larger sibling per verified success" survives the resampling. Minting now requires the intervals to separate, which is why the first claim no longer earns anything.

Paired against the issued instances (2026-08-19). An epoch issues the same instances to every strategy, and instances differ in difficulty. Two separately-widened intervals charge that shared difficulty to each strategy twice and then ask whether they happen to miss each other. Drawing the instance set once and letting both strategies face the same draw removes the part of the variation that belongs to the epoch:

                                unpaired intervals      paired difference
1.5b vs 7b  (token bugs)     [55,162] vs [213,1060]    Δ [-979, -118]  both
1.5b vs 7b  (semantic bugs)  [62,435] vs [300, inf]    Δ [-inf,  -52]  paired only

On the semantic-bug epoch the unpaired intervals overlap and the paired difference does not: the small model's advantage was there, and the unpaired test could not see it. The success-rate axis gets the same treatment, by exact McNemar over the instances the two disagree on — phi4 solved five instances the cascade missed and missed none the cascade solved (p = 0.031), which no Wilson interval on 11/12 against 6/12 will show.

This direction issues more, so it is worth saying plainly why it is not a loosened standard: the unpaired test was not conservative by design, it was the wrong instrument for a common-instance comparison, and being accidentally strict is not a safety property anyone should bank. Pairing is refused unless the two certificates come from the same epoch and every observation names the instance it was measured on — position in a list is not an identity.

The same pass removed a mint. A record standing on one success in six had an interval running to infinity; a challenger overlapping it was certified anyway, because the record's interval was recomputed at write time and never survived the trip back out of the ledger, leaving only a flat 20% margin on point estimates to decide. Replayed under the current rule, that certification does not happen.

Declared bias, as always: the bug vocabulary is a subset of the searcher's mutation vocabulary, so the cascade resolved everything in its cheap stages (its LLM stages never fired). On a distribution that defeats search, these numbers will look different — so we built that distribution next.

The frontier inverted (epoch f7f0863a, semantic bugs)

Six semantic bug classes — population-vs-sample variance, unsorted median, inverted mean, and friends — each proven by test to be irreversible by any single-token substitution. Same seven strategies, same rules:

strategy        success   J/success     (previous, token bugs)
phi4 14b         11/12    221   ← took the crown        (11/12, 287)
cascade           6/12    299   ← dominated by phi4     (12/12, 1.37 — was champion)
qwen2.5:1.5b      4/12    111   ← still cheapest/success (6/12, 81)
qwen2.5:7b        3/12    617                            (5/12, 391)
7b × 3 retries    3/12    742                            (5/12, 404)
brute search      0/12    ∞     ← extinct, as proven    (12/12, 1.53)

The frontier is a property of the distribution, not of the strategy. Yesterday's champion is today's dominated entry; search went extinct on schedule; the 1.5b kept the price-per-success crown on both distributions and Pareto-dominated its own 7b sibling twice in a row; retries bought nothing either time (systematic failures stay systematic). "Which intelligence is efficient" has no distribution-free answer — and EDEN can now measure that inversion in joules, with receipts, both ways.

First independent replication (two nodes)

On 2026-08-18 a second machine (M1 MacBook Air, Python 3.9) ran the same pipeline inside a signed, sandboxed relay job and its receipts were imported into this ledger as explicitly unsigned claims. The family id derived independently on both machines and matched exactly; so did every runner code hash. Hardware-fingerprint stratification kept the two machines in separate replication groups:

counter_fast @M5 Pro  n=2  0.094 J        counter_fast @M1  n=2  0.128 J
dict_loop    @M5 Pro  n=2  0.097 J        dict_loop    @M1  n=2  0.139 J
naive_count  @M5 Pro  n=2  0.408 J        naive_count  @M1  n=2  0.595 J

Same code, ~1.4× the cpu-time on the older chip — the first measured data point for the hardware-wave open problem in the spec.

Measurement honesty

Every condition that decides whether a receipt group may hold a record or trigger a mint lives in one auditable module, eligibility.py:

  • Replication groups are (runner, runner_code_hash, meter) — a renamed or rewritten runner never inherits another implementation's σ, and receipts on different meter boundaries never share a scale.
  • Protocol-assigned uncertainty (systematic) does not shrink with √n; only measured σ (n ≥ 3) does. Interval lower bounds clamp at 0.
  • Dominance across meter boundaries is undefined; records need n ≥ 3.
  • ρ gate: a group whose verification costs more than its run (ρ = E_verify/E_run > 1) can still take the record, but mints nothing — certification is history, minting is economics.
  • Verification energy is measured with the same meter class as the run.
  • frontier is read-only; issuance requires an explicit --commit, repeated transitions never re-mint.
  • Checks the spec requires but v0.1 does not yet enforce (challenge sampling, independent replication, challenge audit) are reported as pending on every assessment — the gate cannot pretend they were checked.
  • Raw observables (cpu-seconds, mW samples, model constants) are stored in every receipt so joules can be re-derived later (Constitution IV).
  • Tamper-evidence: eden chain build maintains an append-only journal where every entry commits to the previous one; eden chain verify re-hashes every receipt body and every link. Edits are not impossible on a single node — they are detectable, and the chain head is anchored in this repository's public commit history.
  • Scope of the claim (audit M8): EDEN measures marginal in-window execution cost. Challenge generation, model loads, warm-ups, downloads, idle power and the meter itself live outside the window — declared, not hidden. "Total physical resources" would be an overclaim.
  • The cultural layer: eden ore scan — receipts are sealed by the seed of the first epoch opened after them (unchoosable randomness); rare hashes are OREs. No economic linkage, by constitution and by test. If OREs ever acquire market value, the protocol does not guarantee their fairness.
  • eden html renders the ledger — frontiers, mints, ores — as one page.

Level V setup (macOS, optional)

printf '%s\n' "$USER ALL=(ALL) NOPASSWD: /usr/bin/powermetrics -i 100 --samplers cpu_power, /usr/bin/powermetrics -i 100 -n 1 --samplers cpu_power" | sudo tee /etc/sudoers.d/powermetrics
sudo visudo -c -f /etc/sudoers.d/powermetrics   # must print "parsed OK"

This grants passwordless root for exactly those two telemetry command lines and nothing else. Do not grant unrestricted NOPASSWD: /usr/bin/powermetrics — its -o/--output-file flag would then allow arbitrary root-owned file writes without a password. Remove the grant with sudo rm /etc/sudoers.d/powermetrics.

Related work (an honest map)

Measuring and rating AI efficiency is an occupied space: ML.ENERGY, the AI Energy Score (Salesforce/Hugging Face), MLPerf Power. Paying for verified efficiency improvements also exists: white-certificate energy markets, efficiency-challenge prizes (NeurIPS LLM Efficiency Challenge). EDEN's specific bet is narrower than "the position is empty": a baseline-free issuance rule — immutable receipts of observed results, minting only on certified Pareto-frontier updates, net of verification cost. Whether that is novel enough is an open question tracked in the spec, not settled in this tagline.

Adversarial audit

On 2026-08-18 this repository was audited by four adversarial reviewers (implementation, measurement statistics, protocol economics, publication). Surviving findings, the fixes, and the retracted claims are recorded in EDEN設計書.md (追補3), including open problems the protocol has not solved: sandbagged genesis records, measurement-window gaming, verifier-coevolution re-issuance, σ-assignment governance, and the honest limitation that EDEN measures marginal in-window resources, not "intelligence" as such.

Specification

Status

Research prototype, single node, unsigned receipts, simulated mint. Licensed under Apache-2.0 — the explicit patent grant matters for a measurement protocol with a hardware roadmap; the trademark exclusion (§6) keeps the EDEN name and future certification marks separate from the code.

About

EDEN — a network that measures what intelligence makes unnecessary. Efficiency receipts + frontier-mint (v0 research prototype)

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages