flaukowski/kannaka-scientist-v1

No trained model yet. Every Kannaka Scientist model so far has failed its pre-registered go/no-go gate; see gates/gate_history.json.

What Kannaka Scientist is

A swappable Brain interface driving an ECDSA.fail knob-search harness: the loop repeats brain.propose(), scorer.score(), ledger.append() and brain.learn() against a fixed knob Space (space/ecdsa_knobs.json) until its budget is spent. random and tpe ship in this snapshot's code/scientist/brains/. ensemble -- the from-scratch model under test -- is not shipped here: it failed every gate it was run against (see Gate history below) and its code is parked, unpublished, on wip/* branches. See spec/design.md for the full architecture.

Gate history

Every go/no-go gate run for Kannaka Scientist v1, including failures; attempts are pre-registered unless marked diagnostic. Bar: beat random in >=8/10 seeds at a 40-config budget (V2 also requires >= TPE's count). Lower final score is better; 1.0 = the record.

ID	Date	Model	Scorer	Seeds	Result	Win counts	Cause / finding	Code commit
t5-mlp	2026-09-28	7 numpy MLPs (1x32 tanh, score+failure heads)	FakeScorer v1	100-109	FAIL	ensemble wins vs random=4; guard history blind wins=6	160 features vs 20-30 samples; members overfit (ensemble-mean corr ~0.035)	876a585 (branch wip/t5-gate-fail)
t5b-bayes-linear	2026-09-28	7 Bayesian linear regressors (empirical-Bayes alpha/beta) + L2 logistic failure head	FakeScorer v1	200-209	FAIL	ensemble wins vs random=4; guard history blind wins=1; old seeds 100 109 wins=4	beta hits its cap (noise-free fake, n < p); predictions track the true score only weakly	d48e932 (branch wip/t5b-gate-fail)
tpe-reference-v1	2026-09-28	Optuna TPE (reference, not our model)	FakeScorer v1	200-209	diagnostic	tpe wins vs random=5	FakeScorer v1 is unsearchable at 40 configs: every arm ends 0.97-0.99 while planted optima are 0.91-0.93	-
t5c-bayes-linear-v2	2026-09-28	7 Bayesian linear regressors + L2 logistic failure head (unchanged)	FakeScorerV2 (8 relevant knobs, smooth int bowls, noise 0.002)	300-309	FAIL	ensemble wins vs random=5; tpe wins vs random=7; guard history blind wins=6	does not learn from history within 40 configs (blind copy does as well); linear int features cannot represent bowls	d8314ed (branch wip/t5c-gate-fail; FakeScorerV2 lives there)

Every attempt above is FAIL except the Optuna TPE reference run, which is diagnostic (TPE is not our model). The full per-attempt detail, including per-seed final scores where recorded, is in gates/gate_history.json. The code_commit values above (the wip/* branches) are not public -- they are not part of this snapshot or this repo's public history; only the code under code/ here, at the commit named in MANIFEST.json, is published.

What is in this repo

• README.md -- this card.
• LICENSE and NOTICE -- the Space Child License, byte-identical to the source repo's.
• space/ecdsa_knobs.json -- the knob Space these attempts searched over.
• priors/ecdsa_history.jsonl and priors/ecdsa_history.counts.json -- the mined public-submission history prior (facts about public submissions only).
• gates/gate_history.json -- every gate attempt, including one non-pre-registered TPE diagnostic, compiled from the recorded counts; the failures are included.
• threat_ledger/ -- the BQR threat ledger's first artifact: the ECDSA.fail record curve (CSV, chart, methods note); see Threat ledger below.
• spec/design.md -- the design spec these attempts were run against.
• results/e001/ -- the e001 run as it happened: the control rows, the warm-up, every screened config per arm and seed, every confirmation row, and the corrected report.
• spec/v2-design.md -- the Kannaka Scientist v2 design spec (amendments 1-4), and spec/c005-prereg.md -- campaign c005's pre-registration.
• campaigns/c001/ .. campaigns/c005/ -- each v2 campaign's registration (experiment.json, experiment.md), seed hypothesis ledger (seed.jsonl) and results as they happened; see Kannaka Scientist v2 below.
• research/ -- the research notes the v2 campaigns were built on and closed with.
• code/ -- a snapshot of scientist/, tools/, pod/, tests/ and pyproject.toml at the published commit (git archive HEAD; only committed code, never the working tree); this is the v2 machine at the commit named in MANIFEST.json.
• MANIFEST.json -- the git commit, build time, sha256 of every file above, and "model": null.

History-prior caveat

The history miner examined 1,307 submissions; 21 rows mined; 6 usable within k=3. Public submissions also changed code, not just knobs, so a history row's score reflects code and knobs -- the prior is a head start, not ground truth (history rows are down-weighted 0.3 against 1.0 for our own scored runs). Mapping each mined row onto this Space and dropping any that change more than k=3 knobs leaves the 6 usable rows.

Threat ledger

The BQR threat ledger tracks progress on the ECDSA.fail benchmark, under its rules (approximate circuits allowed; average executed Toffoli x peak qubits), measured from the benchmark's first record. The benchmark's task is one secp256k1 elliptic-curve point addition as a reversible quantum circuit. As of 2026-09-28T16:15Z, it covers 533 promoted records:

• First record (2026-05-30): score 10,753,444,395 (3,960,753 Toffoli x 2,715 qubits)
• Latest record (2026-09-28): score 1,111,307,500 (889,046 Toffoli x 1,250 qubits)
• Factor: the latest record's score is 9.7x below the first record's.

Progress on this benchmark, under its rules (approximate circuits allowed; average executed Toffoli x peak qubits, over 9024 Fiat-Shamir shots), measured from the benchmark's first record. Each record is the best publicly verified cost of this one sub-circuit (one secp256k1 point addition) under the benchmark's rules; a full attack costs more, and better unpublished circuits may exist. This is not a full attack, and it gives no break date.

See threat_ledger/ for the CSV, the chart and the methods note.

What's next

• e001 (Track A, run 2026-09-28 on the qBraid Lab): arms tpe, random x seeds 1, 2 at a budget of 30 configs per arm and seed (120 configs total). Verdict, from results/e001/report.md: Track A: TPE beat random on best score in 1 of 2 seeds and on AUC in 1 of 2 seeds -- no evidence either way at n=2. Three configs passed the 32-nonce confirmation about 30 to 50 Toffoli below the record on the same nonce labels; they were judged with the +2.0 error margin against the record on those labels (the evaluator derives its inputs from a hash of each circuit, so this is a difference of means between two input sets, not a paired test), are not contract-checked (upstream PR 419), and are not submission candidates. The report's first published version carried a wrong FAIL headline (an ensemble pass rule that Track A never used) and listed the tpe seed-1 arm as incomplete when every one of its five confirmed candidates had failed the gate; both defects were fixed in the tool (commit 6efa10d) and this is the corrected report.
• Track B (laptop only, never on the Lab): a further from-scratch model attempt, gated on fresh FakeScorerV2 seeds 400-409 against the same bar (>=8/10 vs random, and >= TPE's count), openly labelled as shaped by knowing V2's structure.
• Track B's results, if it passes its gate, are added to this repo after Nick's go. A model, if one ever passes its gate, goes to a Hugging Face model repo of the same name.

Kannaka Scientist v2 (structural search)

v2 moves Kannaka Scientist from choosing knob settings to proposing, testing and remembering structural changes to the ECDSA.fail record circuit (base 454ff5a, 889,621 Toffoli x 1,249 qubits). The design is spec/v2-design.md, with amendments 1-4 binding. The loop is:

hypothesis ledger -> proposer -> policy check -> local gate (laptop build and evaluate) -> Lab screen -> Lab confirm -> ledgers (every attempt, failures included, appended to results.jsonl).

• The proposer is a restricted headless Claude session (claude -p in a throwaway git worktree of the base, with a dollar cap per call). Its tools are Read, Grep, Glob and Edit only: no shell, no web. From amendment 3 on it also gets one sandboxed tool, measure, served by our own local MCP server: it takes no arguments, runs the policy check and the local gate on the worktree's current diff, and returns numbers, never a shell (at most 4 measurements per session). The diff is taken from git diff, not from the model's text, and the driver re-runs policy and the local gate on it before any Lab spend.
• Policy and local gate. The diff may touch only src/point_add/**, at most 1,500 changed lines, with no process, network, unsafe or extern constructs; the laptop gate builds it, evaluates one nonce and runs the structural checks before any Lab time is spent.
• The PR 419 limiting contract. Every ECDSA.fail record since upstream PR 419 carries four terms: no refinement may raise the circuit's error Λ (failing 64-shot batches per 9,024 shots) above 18; no simple error-for-score trades, such as round reductions; inefficient error buys may be swapped for cheaper ones only while Λ stays at or below 18; and the note and contract are copied into later commits. The base reports Λ = 18.0 ± 0.89, so there is no headroom. v2's gates therefore measure an error e (classical mismatches + phase-garbage batches + ancilla-garbage batches, per nonce) against the base, and the proposer refuses bare error trades (amendment 1).
• The paired-vs-unpaired correction. Amendments 1 and 4 called the Lab comparisons "paired by nonce". The evaluator draws its 9,024 test inputs from a SHAKE256 hash of the whole op stream, so a changed circuit never sees the base's inputs, even at the same nonce label: the pairing matched labels, not inputs. Every number below is an unpaired difference of means, because the evaluator hashes the op stream. On the one file with both sides' per-nonce counts (a002, 128 nonces) the SE is 0.784 by the difference formula against 0.808 unpaired, correlation +0.06, so the figures stand and only the label changed (campaigns/c004/survey2-corrections.jsonl; spec/c005-prereg.md, section 0).

Campaign	What ran	Outcome	Claude spend
c001	8 seed hypotheses, proposer without the measure tool	8/8 refusals, each a code-cited analysis (research/2026-09-29-c001-refusals.md); no patch	1.71 USD
c002	hypothesize step (amendment 2: measured experiments allowed)	hypothesize refused: nothing clears the bar without being a bare error trade	1.17 USD
c003	proposer with the measure tool (amendment 3)	a002 found: Q 1,249 -> 1,248, T 889,621.9 -> 889,321.6 (-0.114% on T x Q); at 128 nonces on the Lab, e rose +1.83 ± 0.78 per nonce, all phase (9.80 -> 12.04), bought by narrowing one compare from 21 to 20 bits: rejected as an error trade	1.89 USD
c004	two Lab surveys stacked on a002 (5 widening variants, then 4 restorations of the narrowed compare), 128 nonces each	no variant passed; restoring the compare costs about 800 T per unit of error, while a qubit is worth about 713 T, so the widening branch is closed (research/2026-09-29-c004-summary.md)	none (no proposer call)
c005	pre-registered exact search (spec/c005-prereg.md), 4 proposer sessions, no Lab time	4 refusals, 3 candidates measured on shared inputs, all at zero effect; S1 failed as pre-registered, and ECDSA.fail is parked for this machine at this record (research/2026-09-29-c005-summary.md)	2.63 USD
c006	pre-registered exact search on the record circuit (spec/c006-prereg.md), 4 proposer sessions; the rebased analysis patch gained a fifth anchor	4 refusals; S1 failed as pre-registered; the record's post-pass chain measured at a fixed point and its history tape empty at the peak	2.34 USD
c007	pre-registered pricing of two error buy-backs against the record (spec/c007-prereg.md, revision 2), 3 arms x 2,048 paired items on a free qBraid CPU plan, SAT rewrite rows off in every arm	C1 (restore 195 room cuts) removes 0.265 ± 0.079 e for +551 T, 2,082 T/e; C2 (buy back compare band 0-99) removes 0.157 ± 0.074 e for +140 T, 890 T/e, interval contains the author's 715; no cheap buy-back (campaigns/c007/report.md, correction in public-correction.md)	0 USD (47.3 plan-h)

Each campaign's registration, seed ledger and results are under campaigns/<id>/ (experiment.json, experiment.md, seed.jsonl, results.jsonl, hypotheses.jsonl, report.md, diffs and Lab files). The campaign reports' header line still says "paired by nonce"; read it with the correction above. No campaign landed a patch. The code under code/ is the v2 machine (scientist/campaign.py, the driver; scientist/patch/, the proposer, policy and measure server; scientist/gates/) at the commit named in MANIFEST.json.

c006 and c007 were added on 2026-10-03 from source commit a71d647. Their pod logs and drivers had the operator's home directory replaced with /home/<user>; nothing else in them was edited. c007's items pair item by item (fixed test inputs per item), unlike c003-c005.

campaigns/heesch-f1/ is not an ECDSA.fail campaign. It holds the Heesch challenge leader's tile with only the tile fixed and pockets counted: proof-checked UNSAT certificates for defect at most 2, at most 3 and 4, and the ring-size search, which finds 359 as the largest ring (the upper side is an unchecked solver UNSAT). The LRAT proofs were streamed into the checker and never stored, so the folder holds their hashes, not the proofs. Added 2026-10-04; see its README.

Licence

Space Child License v1.0 -- free for peaceful use, war pays. See LICENSE and NOTICE.

The Space Child License covers our own code, docs and results. The ECDSA.fail-derived data (submission knob tables, official metrics, upstream context in the patch) belongs to the ECDSA.fail benchmark (Layr-Labs/ecdsafail) and its submitters, and is reproduced for research with credit.