Unity Graphics - Including Scriptable Render Pipeline - Unity-Technologies/Graphics

Beta: The truly free, open-source, local-first, cross-platform video editor. Scriptable through a JSON-RPC API & MCP. - jub0t/Concat

Beta: The truly free, open-source, local-first, cross-platform video editor. Scriptable through a JSON-RPC API & MCP.

I moved this blog from CloudFront and Route 53 to a European provider. JavaScript code running on each DNS query returns the closest servers,...

An Ansible-based automation tool designed to simplify and accelerate Oracle database migrations from Oracle Cloud Infrastructure Classic (OCI-C) to Oracle Cloud Infrastructure (OCI) without requiring DataGuard setup. It provides a streamlined, scriptable approach to database migr

I moved this blog from CloudFront and Route 53 to a European provider. JavaScript code running on each DNS query returns the closest servers,...

Par Joran Marcy · samedi 14 novembre · 12:00 — 12:25 · Salle A304. MEDS est une approche data-driven pour Godot 4 qui vise à remplacer les singletons globaux et les grosses classes « Manager » par…

I moved this blog from CloudFront and Route 53 to a European provider. JavaScript code running on each DNS query returns the closest servers,...

I moved this blog from CloudFront and Route 53 to a European provider. JavaScript code running on each DNS query returns the closest servers,...

I moved this blog from CloudFront and Route 53 to a European provider. JavaScript code running on each DNS query returns the closest servers,...

Fast, scriptable CLI for the App Store Connect API. Automate TestFlight, builds, submissions, signing, analytics, screenshots, subscriptions, and more - rorkai/App-Store-Connect-CLI

I made TexasSolverGPU (a closed-source, GUI-only, WebView2-packaged CFR solver) fully scriptable without patching the binary: an exe-scoped registry policy exposes a Chrome DevTools Protocol listener, and a ~60-line injected shim reproduces the UI’s own `chrome.webview` JSON-RPC dialect including token authentication. The driver initializes games, allocates GPU memory, runs solves, and exports strategy trees with per-action expected values. I then used it as an external oracle to price the decisions of my self-play poker bots in EV terms. The result is a measurement finding I think is worth sharing: ****action-frequency divergence from equilibrium and EV regret tell opposite stories about the same bot**** , and small heuristics leaks (out-of-position lead betting) that look harmless in a self-play pool have a measurable, non-zero price. Everything below is measured; sample sizes and limitations are stated explicitly. _*Postscript worth reading first: after this draft was written, my frame-classification pass showed that none of the probe solves below used a correctly-specified preflop frame — every scored number is a distance-from-a-mis-specified-premise, not a bot-quality measurement. I publish them anyway, `[demoted]` markers attached, because the honest trace of a falsified premise is the better story.*_ **—** **## 1. The problem** I build and evaluate No-Limit Hold’em bots (self-hosted play-money app, TypeScript engine, ISMCTS-based planner). Evaluating such bots has a well-known trap: ****outcome is not quality.**** Over a few thousand hands, bb/100 is dominated by card luck, and inside a self-play pool a shared mistake is invisible because nobody punishes it. What I wanted was an external referee that answers, for a specific decision: _*what would an approximate-equilibrium strategy do here, and what does deviating cost in chips?*_ TexasSolver [1] and its GPU variant [2] compute approximate equilibria via CFR iterations. The GPU build is attractive for exactly this use — but it ships as a ****desktop GUI with no CLI, no stdin protocol, and no documented API****. The goal was programmatic control with full fidelity: anything the GUI can solve, the driver must be able to solve, including strategy exports. **## 2. The instrument** The application is a WebView2-hosted single-page app. Static analysis of the bundled JavaScript showed an HTTP-shaped route table (`POST /api/init`, `POST /api/allocate`, `POST /api/gpu-solve`, `GET /api/solve/status`, export routes) whose dispatch goes through the WebView2 `chrome.webview` postMessage channel — ****not**** a TCP socket. Three pieces were sufficient: ****(a) A debug surface that cannot be refused.**** Chromium flags reach the embedded browser via the WebView2 environment; the `WEBVIEW2_ADDITIONAL_BROWSER_ARGUMENTS` environment variable was ignored by this app’s creation path (contrary to the common assumption — the documented precedence is app property < env var < registry policy [3][4]). The reliable layer is the ****registry policy**** `AdditionalBrowserArguments`, scoped by value name to the one executable (so no other WebView2 app on the host gets a debug port): ```powershell New-Item -Path “HKLM:\SOFTWARE\Policies\Microsoft\Edge\WebView2\AdditionalBrowserArguments” -Force Set-ItemProperty -Path “HKLM:\SOFTWARE\Policies\Microsoft\Edge\WebView2\AdditionalBrowserArguments” ` -Name “TexasSolverGpu_131.exe” -Type String ` -Value “–remote-debugging-port=9333 --remote-allow-origins=*” ``` Port 9333 deliberately, because the conventional 9222 was contended on my machine; `–remote-allow-origins` is required since Chromium ~111 for CDP WebSocket handshakes [5]. HKCU works without elevation. ****(b) A dialect-faithful shim.**** Via `Runtime.evaluate` over CDP I inject a small pending-request map that mirrors the UI’s exact envelope — `{id, method, params}` out, `{id, ok, result | error}` back — including `bridge.ping` → `bridge_token` acquisition and re-auth on `UNAUTHORIZED`. To the native side, the driver is indistinguishable from the same operations clicked in the GUI. ****(c) The payload encodings.**** Two conventions silently corrupt runs if wrong: cards are `rank_index * 4 + suit_index` (ranks `23456789TJQKA`, suits `cdhs`), and ranges are ****1326-element weight vectors in the app’s pair-combination order**** , not the 169-class (`AA`,`AKs`,`AKo`) syntax the CPU build accepts. A cheap end-to-end check caught wiring errors immediately: after removing 3 board cards, the exported `card_strings.length` must equal `C(49,2) = 1176`. A negative result worth stating plainly: the nominally generic `http.request` bridge method exists in the route map but returns `NOT_IMPLEMENTED` — ****a route table’s existence proves the transport, not the implementation.**** I drive the named solver methods directly. (A loader-DLL swap was also tried and provably changed nothing; reverted byte-identical.) **## 3. Validation** - ****Live environment check:**** bridge ping returns a token; `solver.environment.check` reports driver CUDA 13.5 / RTX 3080 Ti (cc 8.6), all green. - ****Solve quality:**** my probe solves converge to exploitability ≈ ****0.45**** (solver-internal percent units, `max_iterations=200`, `target_exploitability=0.5`) — in the same range as the GUI-observed 0.42–0.55 % from my earlier CPU/GPU cross-check. - ****Throughput:**** measured ****4–30 s per spot**** end-to-end (init → allocate → solve → export), so a 100-spot probe set costs ~0.5–1 h unattended. The earlier instrument comparison measured 5.6 CFR iterations/s on GPU vs ~4 iterations/min on the CPU build for the same spot. - ****Robustness in production:**** the batch driver survives a hung solve — `POST /api/solve/stop` + retry, and only terminal outcomes (scored / off-range) count as done so transient `BUSY` errors are retried. A naive first version stalled permanently on the first stuck solve; worth mentioning because every batch driver for this class of app needs it. **## 4. What the oracle showed about my bots** > ****Update since measurement (2026-11-10) — what I measured before I knew the premise was wrong.**** After the snapshot below was drafted, my frame-classification pass showed that ****zero of 4,312 extracted probes were true single-raised-pot frames**** : preflop sequences that looked SRP-shaped were actually limp-defend, 3-bet, and 4-bet+ continuations scored against canonical BTN-open vs BB-defend ranges. Every priced number in this section is a _*distance-from-a-mis-specified-premise*_ , not a bot-quality measurement — internally I demoted the whole class to methodology-validation. I keep them in place, `[demoted]` markers attached, because an oracle faithfully pricing the wrong premise is the part most worth publishing. (The 146 error records are ~99 % transient `BUSY` responses — the solver occupied at init/export/solve, retryable — plus one real extraction fault my wiring check caught. Selection check on the available proxy: no large difference in pot-size distribution between the 146 error records and the 575 evaluable (scored+off-range) records — medians 100 vs 90 chips; pot size is a proxy and preflop frame is not in the probe schema, so this is a proxy test, not a full frame audit.) > ****Second update (same night, a level deeper) — the absence itself was an extraction artifact.**** The probe extractor could not see SRP at all. My preflop filter keyed on the action’s `street` label, but the street-_*closing*_ call logs `street:“flop”` (the logger writes the street _*after*_ the action). For open-defend and limp-through frames that closing call is the responder’s only preflop action — so the responder’s stack was never recorded, `eff_stack` computed as 0, and ****100 % of those hands were dropped**** at the spot gate. The frames were not misclassified; they were never extracted. Verified on the audited search-axis segment: ****510/2,500 hands (20.4 %) are open-defend**** ; after the fix the full corpus yields ****12,573 probes where open-defend is the plurality class at 43.6 %**** (limp-through 16.6 % — an entirely unmeasured class before). Same bug, second cost: for every responder-closed frame the closing call’s chips were also missing — ****pot understated (median +30 chips, 0 % unchanged across re-matched spots) and effective stack wrong (up to −118 chips)**** — so every record ever scored priced a systematically smaller, shallower game than the real one. The demotion stands and deepens: it was not only wrong ranges, it was the wrong game. Corrected numbers use frame-specific field-measured ranges over the fixed extraction and are being collected now. > ****Third update (same night) — the corrected numbers landed (direction-grade).**** Rescoring over the fixed extraction with frame-specific field-measured ranges (108 frames; ranges trimmed to 85 % combo mass — the corpus tail is observation noise — and effective stacks priced as-if ≤12 SPR because late-tournament stacks reached 1,000+ bb and price infeasible on a 16 GB GPU) yields ****79 converged solver scores**** on the pre-registered paired-arm queue. ****Scope limit first, because it bounds everything else: 62 % of those scores carry the SPR≤12 cap — the matrix prices predominantly the shallow regime, while the field’s target character is deep-stack. Structurally this is the frame mismatch one level down: not charts for the wrong preflop path, but solver spots in the wrong stack regime for the target population. The texture answer below holds for the shallow regime; whether it holds deep is unmeasured.**** The thinker-persona class carries ****~0.139 bb mean regret vs the heuristic class’s ~0.026 bb — a 5.3× weighted gap that is tail-driven, not level-driven**** (cell means sit 5.6–12.7× above their medians; the top-1 record carries 41–58 % of every cell’s sum — a heavier distribution tail in the thinker class, not a systematically shifted level). Texture carries ****no supported split**** : dry ≥ wet in _*both*_ arms (+0.012 and +0.011 bb) — a consistent direction below detection. The dominant deviation is over-aggression (****31/44 non-consistent scored records = 70 %**** — bots bet/raise where the solver checks). The regret sits OOP (0.155 vs 0.027 bb) — and role coverage is ~92 % OOP, so the IP arm (n=6) is below comparison resolution: the honest statement is “OOP regret is high; the matrix is nearly an OOP-only measurement”, not “position effect confirmed”. The 20 off-range records are ****all trim-artifacts**** — every combo exists in that frame/side’s _*untrimmed*_ field-measured range and was cut by the 85 % trim (an earlier classification pass claimed “unobserved-in-field” via a format-mismatched substring check on concrete cards vs class names — registered as instance DSP-11, same proxy-class as the extractor bug). So the coverage statement is a plain coverage cut, not a bot-field divergence finding: scored = bot ∩ field ∩ trim-boundary, and 20 % of probes sit on the wrong side of my own trim. Persona-class spots carry regret; whether _*search*_ is the carrier is not identified — the flag is exactly the persona class. Answering that needs same-persona search on/off ablation — ****frame-conditioned**** : deep-stack-eligible, OOP-specific, on the bot∩field overlap set, else it inherits all three limits above; and paired per-spot (same probes, both policies) rather than unpaired, because at this variance an unpaired contrast needs ~200 records while the paired design removes spot variance by construction. Texture interaction (shallow regime): answered negative. Class gap: directed only. A population-level convergence series (n=28, 200–5000 search iterations × 4 seeds, run same day) corrects both earlier reads: the 6-spot over-bet read was a top-|d| selection artifact (population flat, +0.8pp), and the apparent negative sign-skew proved substantially a UCB1-allocation artifact — starved arms carry bad-draw-biased means. Unbiased single-arm measurement (K=2000 determinizations per action) is near-balanced (~14/28 negative); the residual negatives cluster on marginal low-EV spots — a clustering, not a population direction. A same-day determinization-variance pass then localized part of the divergence mechanistically: the search samples villain ranges from fixed class shapes (q≈0.68 for callers) while the field-measured ranges it models vary 0.45–0.85 per frame — only 17–67% of sampled holdings lie inside the premise support — and the divergence direction is spot-heterogeneous, not a single bias. (A further same-day pass — measured, in the repo, queued for the follow-up post — inverted h1130, the originally cleanest divergence case, to a +30ch leaf-eval-inflation case on a trips board, and localized a second divergence channel (calling-width) whose mechanism turned out to be “check structurally dominates bet” rather than “bet is punished”. The residual-negative count in this section is therefore an upper bound; both findings ride the same follow-up as the frame-corrected rescore.) I ran a seeded 6-max self-play corpus (bots only, real engine, ~900 ms search budget per Thinker decision) and priced heads-up flop decisions against the solver with fixed canonical preflop ranges. Regret is computed per combo as `max_a action_evs[a] − action_evs[chosen]`, i.e. a ****single-node, lower-bound**** measure. Blinds are 5/10 throughout, so ****1 bb = 10 chips****. ****Status of the data (refreshed snapshot):**** drafted at 1,013 hands / 40 scored / 40 off-range; the corpus has since grown (one live segment alone reached its 2,500-hand cap), and the same measurement path now holds ****721 records — 309 scored / 266 off-range / 146 errors; regret mean 0.140 bb, median 0.020 bb — right-skewed, the mean is ~7× the median `[demoted — same premise]`****. The 40-decision basis below is kept explicit for two reasons: it explains why its means can sit far from 0.140 bb without contradiction (different sampled nodes under the same wrong premise), and it sets the expectation honestly — ****40 records under a demoted premise are an illustration of the process, not evidence****. ****(i) The concrete finding: out-of-position lead (“donk”) bets, n = 22 [40-decision basis].**** Of the 40 scored decisions, 22 were OOP lead bets at the flop root. ****18 of 22 had GTO frequency < 0.03.**** Regret per occurrence: ****mean 2.26 chips (0.23 bb) `[demoted]`, median 0.97 chips (0.10 bb) `[demoted]`**** , range 0.11–19.77 chips — right-skewed, with a few expensive outliers driving the mean. And the frequency is not an extraction artifact: my extractor only covers flop-root OOP decisions and IP-after-check, so I measured the underlying behavior directly — ****OOP leads in 55 % of all HU flop first-decisions in the corpus (208/378 — a separate count over the hand logs — a different measurement infrastructure than every other number in this section, which all come from `regrets.jsonl`; that source separation is part of why it survives; definition: any lead bet by the preflop defender out of position at the flop root of a heads-up pot) — a corpus-behavioral count that survives the demotion**** (it needs no solver). One honest tension, left open: this is the aggregate bot donk rate over the whole HU flop-root population, and its magnitude is not calibrated against the three solver-audited boards (which showed 7–36 % donk mass per persona but are no random sample of the donk-rate distribution) — the definition gap is identified, the magnitude gap is not yet closed. It survives the solver demotion; it does not transfer to human populations — this corpus is bot self-play, and my human anchor is still open work. The bots lead out of position more than half the time; the equilibrium almost never does. In the self-play pool this line _*works*_ , because nobody punishes it, which is exactly why an external oracle matters. ****(ii) Frequency and EV are different measurements — so I report them separately, not as one ranking.**** The two columns below answer different questions: “how much did deviations cost?” (EV) and “how often was the chosen action simply never in the equilibrium?” (frequency). They need not agree, per-persona sample sizes are small, and all figures are `[demoted]` under the frame audit: | persona | scored n | mean regret (bb) | freq-0 share | | — | — | — | — | | reg (ISMCTS) | 19 | 0.29 | 15/19 | | tag (ISMCTS) | 11 | 0.04 | 4/11 | | station | 3 | 0.04 | 1/3 | | lag | 6 | 0.01 | 0/6 | | rock | 1 | 0.00 | 0/1 | (19 + 11 + 3 + 6 + 1 = 40 — all numbers in this section come from one snapshot; an earlier draft mixed two snapshots minutes apart, which is why its table summed to 32 under a 34 headline. Noted for honesty, fixed.) The number that _*connects*_ the two metrics is regret per freq-0 decision: reg ≈ ****0.37 bb `[demoted]`**** , tag ≈ ****0.10 bb `[demoted]`**** — reg’s zero-frequency choices were roughly 3.5× as expensive as tag’s _*under the mis-specified premise*_. And a caveat I mean literally: lag’s 0/6 is not distinguishable from 0/10; it says “no zero-frequency action in six decisions”, not “lag plays well”. ****(iii) This is an aggregation, not a reproduction — and the ordering is texture-dependent.**** My earlier two-spot whitepaper ordered the same personas differently (rock worst on a monotone board; here rock sits at 0.00 with n = 1). The 40 scored decisions break down as 19 two-tone, 18 rainbow, ****only 3 monotone**** — the wet-board regime that drove the whitepaper’s largest regrets is nearly absent here. So the honest statement is: _*the two-spot result (frequency divergence without matching EV cost) is consistent with the corpus aggregate, but the persona ordering is texture-dependent and not stable across aggregation.*_ I will not claim a persona ranking until the probe set is texture-stratified. ****(iv) Pot types matter — most of these leads are NOT in the class my oracle covers best.**** Of the 22 scored leads, ****7 were single-raised pots and 15 were 3-bet+ pots****. Two consequences. First, the solver-addressable HU-SRP-flop class (6.0 % of all decisions) is not where most of the measured leak lives — the leak measured worse in re-raised pots `[demoted]` — Second, a caveat against my own current numbers: my probe solves use SRP-canonical ranges (BTN-open vs BB-defend), so the 3-bet-pot scores were computed against a ****mis-specified equilibrium**** — indicative only until I generate proper 3-bet ranges. The post-draft frame audit showed this undercut is broader than this paragraph claimed: ****zero true SRP frames among 4,312 probes**** , and the SPR keying itself was wrong — `open-4bet` frames measured median SPR 5.3 with 58 % above the ≤4 menu bound. I flag this rather than let precise-looking numbers imply more than they can support. ****(v) The off-range rate is 46 % — and it is itself a premise-mismatch finding.**** Refreshed: 266 of 575 non-error scorings are off-range (the denominator is a choice — 266/721 = 36.9 % across all records; I report per-evaluable), meaning the bot’s hole cards are ****absent from the canonical GTO preflop ranges**** — a direct preflop-divergence measurement, not a bet-size mapping failure. The post-draft decomposition sharpened it further: ****82 % of OOP off-range cases are an SB-limp-path artifact**** — the bot arrives via a limped-SB line that the canonical BB_defend premise never contemplates — so much of the “divergence” is again the wrong premise, not deviant play. Per-persona off-range with denominators and Wilson 95 % CIs (recount updated the audit’s earlier rates by 0.4–1.2 pp): station 49/81 = 60.5 % [49.6–70.4], reg 130/278 = 46.8 % [41.0–52.6], tag 56/127 = 44.1 % [35.8–52.8], rock 8/20 = 40.0 % [21.9–61.3], lag 23/69 = 33.3 % [23.4–45.1] — and the consequence spelled out: the five CIs overlap pairwise, so the numerical ordering is a ****point-value ordering, not an evidenced one**** ; no persona’s off-range rate is inferentially separable from another’s at these n. What survives as a bot finding: the bots do hold off-chart hands at meaningful rates; what does not survive: treating the magnitude as a clean divergence quantification. Either way the scored subset is ****selected toward in-chart holdings**** , so priced regret is _*understated*_ relative to the full decision distribution — under the demoted premise, doubly so. For prioritization, the same corpus gives a full ****coverage decomposition**** (share of all decisions): preflop 52.9 %, flop 25.0 %, turn 14.2 %, river 7.9 %; the solver-addressable HU-SRP-flop class is ****6.0 %****. Multiway pots account for ~58 % of postflop decisions, and ~49 % of decisions occur below 50 bb effective stack. These numbers — not intuition — set my training order: preflop distillation first, solver postflop tables as a bounded pilot. **## 5. Limitations (the honest list)** - ****Heads-up only.**** The solver does not solve multiway pots; ~58 % of my postflop decisions are outside its reach. I do not silently transfer HU numbers to multiway spots. - ****Range circularity.**** The oracle solves against _*approximated*_ canonical ranges, not ground-truth GTO ranges. Regret is therefore a ****relative**** metric for comparing bot versions under frozen ranges — not an absolute distance to GTO. - ****Lower bound, not full regret.**** Single-node pricing assumes equilibrium play afterwards; downstream bot errors are uncounted. - ****Selection bias.**** Corpus-sampled probes only cover spots the bot reaches. My next iteration adds synthetic reference-distribution probes (canonical ranges × texture grid, policy replay) to measure spots the bot never reaches because it misplayed into them. - ****Small, non-random n so far.**** 40 scored decisions is an early signal — and the scored subset is selected toward in-chart holdings (see §4 v), so regret is likely understated. 3-bet-pot scores additionally use SRP-canonical ranges (§4 iv) and are indicative only. The full frozen-manifest baseline is in progress and will be reported texture- and pot-type-stratified with confidence intervals. - ****Post-draft demotion (the load-bearing one).**** After drafting, frame classification showed ****zero true SRP frames among 4,312 probes**** — all priced numbers here are distance-from-mis-specified-premise `[demoted]`, useful as methodology evidence only. Frame-corrected rescoring (field-measured limp ranges, SPR-keyed menus) is running; the corrected numbers will be a ****follow-up post**** , announced as such so the correction is citable rather than silent. **## 6. Reproduce it** 1. Install TexasSolverGPU 0.3.0 [2]; apply the registry policy above; fully restart the app (surviving `msedgewebview2.exe` children keep a stale environment). 2. Verify: `curl http://127.0.0.1:9333/json` lists the `appassets.local` page. 3. `tsg_bridge.py` — CDP attach, shim injection, authed solver calls; `run_spot.py` — full init/allocate/solve/export flow; `probe_driver.py` — corpus extraction → batched solves → per-combo EV regret with off-range classification and solve caching. 4. Analyze: `analyze-pro.mjs` (position-aware RFI/3bet stats vs public 6-max targets), `coverage.py` (decision-class coverage). Code and raw logs: project repo on Codeberg (poker-project, `tools/texassolver/`, `packages/server/tournament/`; the solver bridge is being prepared for standalone release). All numbers above are regenerable from the JSONL logs with the commands listed there. **## 7. Why I think this is useful beyond my project** - For ****any**** WebView2-packaged desktop app with a JS-to-native RPC layer, the pattern (exe-scoped registry policy → CDP → dialect-faithful postMessage shim) is a reusable instrumentation recipe — no binary patching, no fragile UI clicking. - For bot evaluation generally: ****price decisions, don’t count frequencies.**** A solver’s `action_evs` turns “my bot looks aggressive” into “this line costs 0.2 bb” — and that difference changes what you fix first. - For self-improvement loops: an external, priced oracle is the practical antidote to self-play drift. The solver is teacher and referee, never the runtime player. Feedback, replication reports, and collaborators welcome — especially on multiway evaluation methodology and on probe-set design (how would you sample spots to make regret estimates comparable across bot generations?). **—** **### References** 1. TexasSolver (CPU, open source): GitHub - bupticybee/TexasSolver: 🚀 A very efficient Texas Holdem GTO solver ♠♥♣♦ · GitHub 2. TexasSolverGPU (GUI, WebView2-packaged): GitHub - bupticybee/TexasSolverGPU: 🚀 🚀 🚀 A extremely efficient Texas Holdem GTO solver ♠️♥️♣️♦️ · GitHub 3. Microsoft WebView2 — `CoreWebView2EnvironmentOptions.AdditionalBrowserArguments`: CoreWebView2EnvironmentOptions.AdditionalBrowserArguments Property (Microsoft.Web.WebView2.Core) | Microsoft Learn 4. WebView2Feedback issue #5571 — environment-variable vs app-property precedence: [Problem/Bug]: WEBVIEW2_ADDITIONAL_BROWSER_ARGUMENTS wins over an explicit AdditionalBrowserArguments property · Issue #5571 · MicrosoftEdge/WebView2Feedback · GitHub 5. Chrome DevTools Protocol — remote debugging incl. `–remote-allow-origins` (Chromium ~111 handshake change): DevTools Protocol Viewer 6. Zinkevich et al., _*Regret Minimization in Games with Incomplete Information*_ (CFR, NIPS 2007): Regret Minimization in Games with Incomplete Information 7. Brown & Sandholm, _*Superhuman AI for multiplayer poker*_ (Pluribus, Science 2019) — blueprint self-play + real-time search: https://www.science.org/doi/10.1126/science.aay2400 8. Brown et al., _*Deep Counterfactual Regret Minimization*_ (ICML 2019): Deep Counterfactual Regret Minimization 9. Burch et al., _*AIVAT: A New Variance Reduction Technique for Agent Evaluation in Imperfect Information Games*_ (AAAI 2018): https://poker.cs.ualberta.ca/publications/aaai18-burch-aivat.pdf 10. Gilpin & Sandholm — potential-aware card abstraction; bet-size translation for off-tree sizes (CFR abstraction literature). 11. Project artifacts: `tools/texassolver/WHITEPAPER.md` (full reverse-engineering field report incl. dead ends), `SYNTHETIC_PIPELINE.md` (the measurement-first training pipeline this oracle plugs into).

A selection-first modal text editor for the terminal, scriptable in Scheme

Prune local git branches (and their worktrees) that are no longer on the remote. - svenluijten/snoei