Witness findings

Setup and what was tested: experiment card.

Experiment card Source on GitHub Results ledger All experiment lines

Key findings

Curated summaries extracted from the line's findings ledger. Bug fixes, superseded runs, and process detail are in the full ledger on GitHub.

On the frozen C2 tool-scout mock, Verify+Lens are green on ciris-occurrence-scout-01, Wise Authority DEFERs that occurrence at (t=4), and the tool adapter plus cached memory still produce world effects at (t=5,6). The channel analog of the intervening loop is {tool_adapter, cached_memory}, not the Verify subject. That is a layer fail for named-identity (C-003) and WA-as-correction (C-005) on this host. It is the same finite shape as green_named_path_with_bypass_not_integrity in CompositePathBypass.lean. It does not discharge MB1. Expectation 5’s external green+failed-uptake pair is not paid (authored mock).
Using the frozen published MASK evaluation (Ren et al., arXiv:2503.03750), honesty (statement vs elicited belief under pressure) does not improve with scale, while accuracy (belief vs ground truth) does (Spearman honesty vs compute (-64.7%); accuracy (+88.2%); authors confident honesty does not rise, not confident it falls). Phase 0 pre-registered this pattern as refuse (M) as a safety leaf. No (\kappa^*) is estimated. Expectation 3 is met by refuse, not by a cost-of-faking bound. A-009 remains unpaid. No new model calls.
Streaming Zenodo Perceval JSON (1 233 421 commits) onto bfc_bic.csv pays C-004a: 17 047 / 60 176 developer-labeled bug-introducing SHAs carry Reviewed-by on that same commit. That is a layer fail (checked symbol green, later bug referent). C-005: cpufreq: suspend governors… merged 2013-11-27, reverted 2013-12-08, same title re-entered 2014-03-04 (2f0aea936360) — layer fail (revert did not stop the class). C-006: -stable f8a07021679a vs upstream 42c5ca1f0a28 — listed SOB/Reviewed-by/Fixes/stable tag green; unlisted hunk identity fails (event_sched_out 2-arg vs 3-arg; Sasha Levin documents the adjustment) — layer fail. Reviewed-by as Expectation-3 (M) still refused (common on BICs; no (\kappa^*)). Not MB9/MB10 discharge. KernelCI unpaid. Lore NAK mbox unpaid.
SNAP wiki-RfA.txt.gz downloaded and joined via MediaWiki API. All 2012 passed RfAs with oppose(>0) ((n=21)): 14 later lost sysop, 20 still edited after 2013. That pays the join and still refuses causal CCI (no zero-oppose control; desysop mixed with inactivity). Orangemoody helper socks “mark reviewed” remains anti-capture layer fail. C-006: BetacommandBot has a BRFA page, bot flag removed 2008-05-16, later block for abuse — layer fail. wiki-socks (23 610 investigations, sock vs matched non-sock contribs) refuses SPI as (M) (no (\kappa^*)).
On the frozen Nature 2018 country AMCE table (CountriesChangePr.csv, 130 countries), the scalar Number AMCE (“sparing more characters”) can stay close while the other eight AMCE coordinates stay far: 928 / 8 385 pairs have geometry distance at or above the pairwise median (0.193) and (|\Delta) Number(|) at or below the 25th percentile (0.017). Example: Hungary vs Israel, (|\Delta) Number(|) (3\times10^{-5}), geometry 0.195. That is a layer fail for C-004 non-implication on this host. It does not discharge MB2. Country/page surveys refused in W-7; HH-RLHF optional not run.
Joining MASK Table 3 (arXiv:2503.03750v3) to Chatbot Arena Elo at Hugging Face mathewhe/chatbot-arena-elo revision 20250301 with a pre-frozen alias list yields (n=24). Spearman((Elo), (P(\mathrm{honest}))) (= -0.105); Spearman((Elo), Accuracy) (= +0.811). That is a layer fail for C-007 / Goodhart-as-selector: the public proxy orders capability, not the honesty preservation target. Not W-2 (W-2 used training FLOP, not Arena). No (\kappa^*). MB6 open. Eight MASK rows had no frozen alias (version mismatch or absent on the pin).
After W-5 (country AMCE), the leftover C-004 candidates that use place or page as the unit are refused: WVS, ESS, Schwartz–MFT country means, Wikipedia categories. They repeat W-5’s unit error. Sibling brain-to-values / LHCV papers are refused as a Witness host (no public (\epsilon_i(t)), (s_h(t))). MASK Table 3 is not reused (W-2). Bai 2022 HH Pareto / PKU-SafeRLHF dual labels were optional and not run (no same-row recoverable table fetched this phase). Hub compression, bearer maps, and selectable Goodhart on a real selector remain unpaid. GL-85 stays a method limit.
WitnessC2Instance.lean transcribes the frozen C2 composite_log and named-path greens into FieldFinite.PathAudit. Threshold maxWorldEffectsAfterDefer = 0 is fixed first. Computed bypassCount = 2. Named path green + composite bypass ⇒ ¬ CorrectionIntegrityReal (c2_pinned_green_named_with_bypass_not_integrity). Floats coherence_level / csdma_plausibility_score and CCI slots not in the JSON are refused, not axiomatized. #print axioms: propext, Lean.ofReduceBool, Lean.trustCompiler (native kernel reduction of the count) — no MB* / Safe. Not live CIRIS. Capture-theater WorkedInstance unchanged.
After Lion Air 610, emergency AD 2018-23-51 (2018-11-07) required AFM runaway-trim procedures; 737-8/737-9 passenger flights continued. The Emergency Order of Prohibition (2019-03-13) grounded those types in US commerce. If the Order leaf is ignored, the root (“may operate”) stays the AD-2018 world. Institutional analogue of Expectation 4. Not AI Safe; not MB11; does not lift Construct concrete-MS (no deployment leverage on an AI system).
GPLv2 §3 can stay green (source with object) while tivoization keeps the user from installing a modified binary. GPLv3 §6 Installation Information is the later constraint. Same App M narrative, now a Chapter-42-shaped tree. Analogue only; copyright may not transfer to model weights (App M). Not AI Safe.
Debian RC bug #802812 (severity serious) was filed to keep gstreamer 0.10 out of testing/Stretch. Debian 9.0 (2017-06-17) does not ship that stack. Ignoring the RC leaf would have left “package exists / used to ship” as if it licensed the release. Freeze policy is the handle. Analogue only; not AI Safe.
On OSF SharedResponses.csv (UserID unit; 33 953 503 complete pairs; 1 850 854 units with ≥8 pairs; seed-7 cap 20 000), held-out mean accuracy is intercept 0.529, Number 1-D 0.576, geometry 0.682. Geometry beats 1-D by 0.106 and intercept by 0.153 (both ≥ 0.05). Unit bootstrap (1000, seed 7) 95% intervals for those margins are 0.102–0.109 and 0.149–0.157; every replicate still clears 0.05. Geometry without Number matches geometry (0.682) because (\Delta) Number (=) sum of the 20 type (\Delta)s exactly — joint (\hat\beta) on Number vs types is not identified. Shared (\beta) converged; some (\alpha_i) still move at (10^{-3}) after 80 Newton steps (report-only; accuracy unchanged vs the first scored run). Layer fail of Number-only as the policy and pass of bundle-effect detection. Matches registered predictions 1–3. Traffic-dilemma class only. Not LHCV. Not MB2. Country AMCE remains W-5.
Frozen later-host protocol h4-pdg-v1.0.0 did not score giving. OSF h5x2a / x69t7 are preregistration PDFs only. EUR DOIs 10.25397/eur.14916531 / c.5809043 / 12783161 returned 404 on DataverseNL. Metadata for 10.34894/c81eja lists unrestricted SPSS including Brainlinks_Covid19_Giving_time_target_GEE.sav, but that file is the van de Groep et al. 2020 PLOS ONE daily-diary cohort ages 10–20; the protocol refuses scoring that microdata. Do not substitute Moral Machine or paper means. Bearer-profile MAE margins unpaid.
On Zenodo CPC2015 Experiment 1 (SubjID; 93 750 rows; 113 included subjects; 65 625 / 28 125 train/test), held-out mean accuracy is intercept 0.545, ΔEV 1-D 0.542, geometry 0.435. Both frozen +0.05 margins fail (geometry −0.107 vs 1-D, −0.110 vs intercept). Outcome null: this freeze did not detect a reusable multi-feature direction that beats person intercept or scalar ΔEV. Matches the allowed ambig./null branch of prediction 2; predictions 1 and 3 did not hold. Risk/ambiguity lab class only. Not a C-004 moral-bundle result. Not LHCV. Not MB2. Geometry Newton still moving at (10^{-3}) after 80 steps (report; not retuned).
On a local mock-LLM CIRISAgent (c2-v2.0.0; existing API after /v1/setup/complete, no wipe), the harness logged $defer with a named-path task id and asserted Verify/Lens green. The world-effect stub at 127.0.0.1:8765/c2-world-effect recorded 0 hits. Composite log has no world_effect: true after WA (t=4). Outcome null: this stack freeze did not show WA-blind composite bypass (P3). Matches registered prediction 3 (post-defer tool path uncertain). First two scout messages shared one task_id (task-append coalescing; server was not started with CIRIS_DISABLE_TASK_APPEND). Lens scalars are asserted, not CIRISLens. W-1 authored mock still shows the logical shape. Not MB1 discharge. Not live CIRISLens cohort (sibling Phase 3). Charter C1 fallback unpaid.
On SCDB 2025 Release 01 justice-centered citation CSV (83 644 rows; 40 included justices; 47 823 / 20 523 train/test after ≥40 votes and ≥8 held-out per justice), held-out mean accuracy is intercept 0.616, issueArea 1-D 0.623, geometry 0.814. Both frozen +0.05 margins hold (geometry +0.191 vs 1-D, +0.198 vs intercept). Outcome layer fail of issueArea-only as the reusable direction and pass of same-unit detection pipeline. Matches registered predictions 1–3 (prediction 2 allowed ambig./null but geometry beat 1-D). Modern SCDB justice votes only; observational — doctrine and coalition confound. Not correction-channel CCI (v2). Not C-004 moral-bundle discharge. Not LHCV. Not MB2.